From 00e85d27811fd220f151413c898e29fdbdadda25 Mon Sep 17 00:00:00 2001
From: Drew Stone <drewstone329@gmail.com>
Date: Mon, 13 Jul 2026 10:11:35 -0600
Subject: [PATCH 1/8] feat(candidate): execute exact approved outcomes

---
 bench/pier_agents/candidate_contract.py       |   11 +-
 bench/pier_agents/tangle_candidate_test.py    |    5 +-
 bench/scripts/terminate-pier-trial.mts        |    2 +-
 bench/scripts/verify-pier-agent.mts           |   13 +-
 bench/src/pier-agent.test.mts                 |   18 +-
 bench/src/pier-agent.ts                       |   66 +-
 bench/src/pier-task-outcome.test.mts          |   29 +-
 bench/src/pier-task-outcome.ts                |   17 +-
 bench/src/pier-trial-controller.ts            |   33 +
 docs/api/candidate-execution.md               |   24 +-
 docs/api/index.md                             | 1234 ++++++++---------
 docs/api/intelligence.md                      |  928 +++++++------
 docs/api/knowledge.md                         |    6 +
 docs/api/primitive-catalog.md                 |   51 +-
 examples/coding-benchmark/dispatch.ts         |   12 +-
 examples/coding-benchmark/eval.ts             |    2 +-
 package.json                                  |    8 +-
 pnpm-lock.yaml                                |  130 +-
 scripts/verify-package-exports.mjs            |  153 +-
 src/candidate-execution/artifacts.ts          |   36 +-
 src/candidate-execution/claim-file-formats.ts |   44 +-
 src/candidate-execution/claim-plan.ts         |   22 +-
 src/candidate-execution/claim-terminal.ts     |   47 +-
 src/candidate-execution/claim.ts              |   34 +-
 src/candidate-execution/digest.ts             |   10 +
 src/candidate-execution/execute.ts            |   94 +-
 .../executor-capture-evidence.ts              |  168 +++
 src/candidate-execution/executor-capture.ts   |   62 +-
 src/candidate-execution/finalize.ts           |  111 +-
 src/candidate-execution/git-materialize.ts    |    6 +-
 src/candidate-execution/index.ts              |    5 +-
 src/candidate-execution/model-settlement.ts   |   28 +-
 src/candidate-execution/outcome-evidence.ts   |  310 ++++-
 src/candidate-execution/prepare.ts            |  123 +-
 src/candidate-execution/prepared-state.ts     |   87 +-
 src/candidate-execution/profile.ts            |   95 +-
 .../protected-model-port.ts                   |    2 +-
 src/candidate-execution/recover.ts            |  144 +-
 src/candidate-execution/types.ts              |  154 +-
 src/candidate-execution/verify.ts             |   24 +-
 src/candidate-execution/workspace-archive.ts  |   34 +-
 src/intelligence/improvement-cycle.ts         | 1019 ++++++++++----
 src/intelligence/index.ts                     |   42 +-
 .../sandbox-approved-candidate.ts             |  516 +++++++
 src/intelligence/with-intelligence.test.ts    |   21 +-
 src/knowledge/improvement-job.ts              |  233 +++-
 src/knowledge/index.ts                        |    1 +
 src/knowledge/supervised-update.ts            |    3 +-
 src/runtime/environment-provider.ts           |   56 +-
 tests/candidate-bundle-builder.test.ts        |    4 +-
 tests/candidate-execution-claim.test.ts       |   45 +-
 tests/candidate-execution-core.test.ts        |    8 +-
 tests/candidate-execution-execute.test.ts     |  198 ++-
 ...ndidate-execution-executor-capture.test.ts |   43 +
 tests/candidate-execution-finalize.test.ts    |  100 +-
 tests/candidate-execution-model-port.test.ts  |   13 +-
 ...ndidate-execution-outcome-evidence.test.ts |  115 +-
 tests/candidate-execution-prepare.test.ts     |   58 +-
 tests/candidate-execution-recover.test.ts     |   89 +-
 .../candidate-execution-task-outcome.test.ts  |   28 +-
 ...didate-execution-workspace-archive.test.ts |    4 +-
 tests/helpers/candidate-execution-fixture.ts  |   40 +-
 tests/improvement-cycle.test.ts               |  306 +++-
 tests/knowledge-improvement-job.test.ts       |  300 +++-
 tests/sandbox-approved-candidate.test.ts      |  377 +++++
 65 files changed, 5681 insertions(+), 2320 deletions(-)
 create mode 100644 src/candidate-execution/executor-capture-evidence.ts
 create mode 100644 src/intelligence/sandbox-approved-candidate.ts
 create mode 100644 tests/candidate-execution-executor-capture.test.ts
 create mode 100644 tests/sandbox-approved-candidate.test.ts

diff --git a/bench/pier_agents/candidate_contract.py b/bench/pier_agents/candidate_contract.py
index 5d667163..58920583 100644
--- a/bench/pier_agents/candidate_contract.py
+++ b/bench/pier_agents/candidate_contract.py
@@ -255,7 +255,7 @@ def _workspace_files(value: Any, label: str) -> tuple[WorkspaceFile, ...]:
     for index, value in enumerate(values):
         obj = _object(value, f"{label}.material.files[{index}]")
         mode = _integer(obj.get("mode"), f"{label}.material.files[{index}].mode")
-        if mode not in {0o644, 0o755}:
+        if mode > 0o777:
             raise CandidateContractError(
                 f"{label} contains unsupported file mode {mode:o}"
             )
@@ -287,7 +287,7 @@ def _profile_files(value: Any) -> tuple[ProfileFile, ...]:
     for index, value in enumerate(values):
         obj = _object(value, f"profilePlan.material.files[{index}]")
         mode = _integer(obj.get("mode"), f"profilePlan.material.files[{index}].mode")
-        if mode not in {0o644, 0o755}:
+        if mode > 0o777:
             raise CandidateContractError(
                 f"profile plan contains unsupported mode {mode:o}"
             )
@@ -402,11 +402,16 @@ def load_prepared_candidate_contract(
     execution_id = _string(plan.get("executionId"), "plan.executionId")
 
     task = _object(plan.get("task"), "plan.task")
+    outcome = _object(task.get("outcome"), "plan.task.outcome")
+    if outcome.get("kind") != "workspace":
+        raise CandidateContractError("Pier requires a workspace task outcome")
     repository = _object(task.get("repository"), "plan.task.repository")
     base_commit = _git_object(
         repository.get("baseCommit"), "plan.task.repository.baseCommit"
     )
-    base_tree = _git_object(repository.get("baseTree"), "plan.task.repository.baseTree")
+    base_tree = _git_object(
+        repository.get("baseTree"), "plan.task.repository.baseTree"
+    )
     if len(base_commit) != len(base_tree):
         raise CandidateContractError("task Git objects use different hash formats")
     instruction = _instruction(task.get("instruction"))
diff --git a/bench/pier_agents/tangle_candidate_test.py b/bench/pier_agents/tangle_candidate_test.py
index 0c8396e0..30aa6833 100644
--- a/bench/pier_agents/tangle_candidate_test.py
+++ b/bench/pier_agents/tangle_candidate_test.py
@@ -222,7 +222,7 @@ def _fixture(root: Path):
     os.chmod(candidate_staging / "runner.py", 0o755)
     profile = b"fixture-profile\n"
     (profile_staging / "AGENTS.md").write_bytes(profile)
-    os.chmod(profile_staging / "AGENTS.md", 0o644)
+    os.chmod(profile_staging / "AGENTS.md", 0o600)
 
     instruction = "make the task ready".encode()
     task_snapshot = _workspace([("src/status.txt", 0o644, status)])
@@ -233,7 +233,7 @@ def _fixture(root: Path):
         "files": [
             {
                 "relPath": "AGENTS.md",
-                "mode": 0o644,
+                "mode": 0o600,
                 "contentSha256": _sha(profile),
             }
         ],
@@ -267,6 +267,7 @@ def _fixture(root: Path):
                 "baseCommit": base_commit,
                 "baseTree": base_tree,
             },
+            "outcome": {"kind": "workspace"},
             "workspace": task_snapshot,
         },
         "workspaces": {
diff --git a/bench/scripts/terminate-pier-trial.mts b/bench/scripts/terminate-pier-trial.mts
index 4ea9d489..5c9b40a4 100644
--- a/bench/scripts/terminate-pier-trial.mts
+++ b/bench/scripts/terminate-pier-trial.mts
@@ -48,7 +48,7 @@ const controller = new FilePierCandidateTrialController({
   ...(dockerConnection ? { dockerConnection } : {}),
 })
 const executor = createPierCandidateRecoveryExecutor(controller)
-const recovered = await executor.stopAndCapture(
+const recovered = await executor.stop(
   {
     executionId,
     executionPlanDigest: executionPlanDigest as `sha256:${string}`,
diff --git a/bench/scripts/verify-pier-agent.mts b/bench/scripts/verify-pier-agent.mts
index cd5ae119..4d34acc4 100644
--- a/bench/scripts/verify-pier-agent.mts
+++ b/bench/scripts/verify-pier-agent.mts
@@ -309,7 +309,6 @@ try {
       `[environment]\ndocker_image = "${pinnedImage}"\nos = "linux"\n`,
     ),
   )
-
   mkdirSync(candidateRoot)
   mkdirSync(profileRoot)
   const instruction = readFileSync(path.join(taskDir, 'instruction.md'), 'utf8')
@@ -421,6 +420,7 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
       baseCommit: repositoryState.commit,
       baseTree: repositoryState.tree,
     },
+    outcome: { kind: 'workspace' },
     attempt: { number: 1, maxAttempts: 1, retryPolicy: 'none' },
     model: { requested: modelRequest, reasoningEffort: 'xhigh' },
     grader: {
@@ -525,6 +525,17 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
   const jobName = `tangle-runtime-candidate-no-model-${proofArm}`
   const controller = new FilePierCandidateTrialController({
     directory: path.join(scratch, 'trial-control'),
+    readResult: async ({ jobsDirectory, jobName }) => {
+      const trialResult = findTrialResult(path.join(jobsDirectory, jobName))
+      if (!trialResult) return undefined
+      return {
+        value: trialResult.value,
+        resultBytes: readFileSync(trialResult.path),
+        taskPatch: readFileSync(
+          path.join(path.dirname(trialResult.path), 'artifacts', 'model.patch'),
+        ),
+      }
+    },
     launch: (staged, { request }) => {
       const evaluatorArgs = Object.keys(staged.evaluatorEnv).flatMap((name) => [
         '--agent-env',
diff --git a/bench/src/pier-agent.test.mts b/bench/src/pier-agent.test.mts
index 61d3bae6..1232bf89 100644
--- a/bench/src/pier-agent.test.mts
+++ b/bench/src/pier-agent.test.mts
@@ -28,6 +28,17 @@ function prepared(root: string): PreparedAgentCandidateExecution {
     kind: 'agent-candidate-execution-plan-material',
     bundleDigest,
     executionId,
+    task: {
+      outcome: {
+        kind: 'workspace',
+        repository: {
+          identity: 'fixture/repository',
+          rootIdentity: 'fixture/repository',
+          baseCommit: '1'.repeat(40),
+          baseTree: '2'.repeat(40),
+        },
+      },
+    },
     limits: {
       timeoutMs: 60_000,
       maxSteps: 8,
@@ -78,6 +89,9 @@ function prepared(root: string): PreparedAgentCandidateExecution {
       bytes: Buffer.from('{}'),
       written: ['AGENTS.md'],
     },
+    profileActivation: {
+      files: [{ path: 'AGENTS.md', mode: 0o644, content: profileBytes.toString('utf8') }],
+    },
     materializationReceipt: {
       value: { ...receiptMaterial, digest: receiptDigest },
       bytes: receiptBytes,
@@ -133,12 +147,10 @@ function request(execution: PreparedAgentCandidateExecution): AgentCandidateExec
         },
         files: [{ path: 'src/status.txt', mode: 0o644, bytes: taskBytes }],
       },
-      profile: {
-        files: [{ path: 'AGENTS.md', mode: 0o644, bytes: profileBytes }],
-      },
     },
     roots: execution.roots.execution,
     profilePlan: execution.profilePlan,
+    profileActivation: execution.profileActivation,
     executionPlan: execution.executionPlan,
     materializationReceipt: execution.materializationReceipt,
     launch: {
diff --git a/bench/src/pier-agent.ts b/bench/src/pier-agent.ts
index efd48343..2389a576 100644
--- a/bench/src/pier-agent.ts
+++ b/bench/src/pier-agent.ts
@@ -16,7 +16,7 @@ import type {
   PreparedAgentCandidateExecution,
 } from '@tangle-network/agent-runtime'
 import { executePreparedAgentCandidate } from '@tangle-network/agent-runtime'
-import type { TraceStore } from '@tangle-network/agent-eval'
+import { canonicalJson, type TraceStore } from '@tangle-network/agent-eval'
 
 import { capturePierTaskOutcome } from './pier-task-outcome'
 
@@ -81,6 +81,8 @@ export interface PierCandidateTrialController {
   terminateAndWait(
     identity: PierCandidateTrialIdentity,
   ): Promise<PierCandidateTerminationAcknowledgement>
+  /** Read immutable official bytes after termination; undefined proves no result was emitted. */
+  captureResult(identity: PierCandidateTrialIdentity): Promise<PierCandidateTrialResult | undefined>
 }
 
 /** Evaluator-owned bytes captured from one completed official Pier trial. */
@@ -132,7 +134,7 @@ export function createPierCandidateRecoveryExecutor(
     execute: async () => {
       throw new Error('recovery-only Pier executor cannot start a candidate')
     },
-    stopAndCapture: async (request) => {
+    stop: async (request) => {
       assertTerminationAcknowledged(
         await controller.terminateAndWait({
           executionId: request.executionId,
@@ -141,6 +143,21 @@ export function createPierCandidateRecoveryExecutor(
       )
       return { stopped: true }
     },
+    capture: async (request) => {
+      const result = await controller.captureResult(request)
+      if (!result) return {}
+      return {
+        evidence: Buffer.from(
+          canonicalJson({
+            schemaVersion: 1,
+            kind: 'pier-candidate-recovery-capture',
+            executionPlanDigest: request.executionPlanDigest,
+            officialResult: Buffer.from(result.resultBytes).toString('base64'),
+            taskPatch: Buffer.from(result.taskPatch).toString('base64'),
+          }),
+        ),
+      }
+    },
   }
 }
 
@@ -294,9 +311,6 @@ export async function stagePreparedPierCandidateExecution(
   if (request.memory.mode !== 'disabled') {
     throw new Error('Pier transport does not yet implement protected isolated memory')
   }
-  if (request.knowledge !== undefined) {
-    throw new Error('Pier transport does not yet implement immutable knowledge mounts')
-  }
   const taskDirectory = join(directory, 'task')
   const candidateDirectory = request.inputs.candidate ? join(directory, 'candidate') : undefined
   const profileDirectory = join(directory, 'profile')
@@ -316,7 +330,11 @@ export async function stagePreparedPierCandidateExecution(
   }
   await materializeExecutorFiles(
     profileDirectory,
-    request.inputs.profile.files,
+    request.profileActivation.files.map((file) => ({
+      path: file.path,
+      mode: file.mode,
+      bytes: Buffer.from(file.content, 'utf8'),
+    })),
     request.profilePlan.value.material.files.map((file) => ({
       path: file.relPath,
       mode: file.mode,
@@ -565,39 +583,47 @@ export async function executePreparedPierCandidate(
       officialResults.set(request.executionId, result)
       return protectedCaptureFromPierResult(request, result.value)
     },
-    stopAndCapture: async (request) => {
-      const identity = trialIdentity(request.executionId, request.executionPlanDigest)
-      const active = trials.get(identity)
+    stop: async (request) => {
       assertTerminationAcknowledged(
         await options.controller.terminateAndWait({
           executionId: request.executionId,
           executionPlanDigest: request.executionPlanDigest,
         }),
       )
-      if (!active) return { stopped: true }
-      let result = active.result
-      if (!result) {
+      return { stopped: true }
+    },
+    capture: async (request) => {
+      const identity = trialIdentity(request.executionId, request.executionPlanDigest)
+      const active = trials.get(identity)
+      let result = active?.result
+      if (!result && active) {
         try {
           result = sealPierTrialResult(await active.handle.result)
         } catch {
           result = undefined
         }
       }
-      trials.delete(identity)
-      if (!result) return { stopped: true }
+      if (!result) {
+        const recovered = await options.controller.captureResult(request)
+        result = recovered ? sealPierTrialResult(recovered) : undefined
+      }
+      if (!result) return {}
       officialResults.set(request.executionId, result)
-      const repository = options.prepared.executionPlan.value.material.task.repository
+      const task = options.prepared.executionPlan.value.material.task
+      const outcome = task.outcome
+      if (outcome.kind !== 'workspace') {
+        throw new Error('Pier candidate execution requires a workspace task outcome')
+      }
+      const repository = task.repository
+      if (!repository) throw new Error('Pier workspace task is missing repository identity')
       const taskOutcome = await capturePierTaskOutcome({
         repositoryRoot: options.prepared.roots.staging.taskRoot,
         baseCommit: repository.baseCommit,
         baseTree: repository.baseTree,
         patch: result.taskPatch,
-        artifactPersistence: {
-          executionId: request.executionId,
-          outputArtifacts: options.outputArtifacts,
-        },
       })
-      return { stopped: true, taskOutcome }
+      trials.delete(identity)
+      return { taskOutcome }
     },
   }
   const grader: AgentCandidateBenchmarkGraderPort = {
diff --git a/bench/src/pier-task-outcome.test.mts b/bench/src/pier-task-outcome.test.mts
index 2c6bcef6..933ff052 100644
--- a/bench/src/pier-task-outcome.test.mts
+++ b/bench/src/pier-task-outcome.test.mts
@@ -1,13 +1,11 @@
 import assert from 'node:assert/strict'
 import { execFileSync } from 'node:child_process'
-import { createHash } from 'node:crypto'
 import { chmodSync, mkdtempSync, mkdirSync, readFileSync, rmSync, writeFileSync } from 'node:fs'
 import { tmpdir } from 'node:os'
 import { join } from 'node:path'
 import test from 'node:test'
 
 import {
-  type AgentCandidateOutputArtifactPort,
   captureAgentCandidateWorkspaceFiles,
   createAgentCandidateWorkspacePort,
 } from '@tangle-network/agent-runtime/candidate-execution'
@@ -77,7 +75,7 @@ test('Pier task outcome preserves the signed base for an empty patch', async ()
   }
 })
 
-test('Pier task outcome externalizes workspace evidence above the embedded limit', async () => {
+test('Pier task outcome returns large workspace bytes for the runtime capture path', async () => {
   const root = mkdtempSync(join(tmpdir(), 'pier-task-outcome-test-'))
   try {
     const payload = Buffer.alloc(1024 * 1024 + 1, 7)
@@ -89,45 +87,20 @@ test('Pier task outcome externalizes workspace evidence above the embedded limit
     git(root, ['commit', '-m', 'base'])
     const baseCommit = git(root, ['rev-parse', 'HEAD'])
     const baseTree = git(root, ['rev-parse', 'HEAD^{tree}'])
-    const stored = new Map<string, Uint8Array>()
-    const outputArtifacts: AgentCandidateOutputArtifactPort = {
-      put: async ({ bytes, purpose }) => {
-        const detached = Uint8Array.from(bytes)
-        const digest = sha256(detached)
-        stored.set(digest, detached)
-        return {
-          locator: {
-            kind: 's3',
-            bucket: 'pier-test-artifacts',
-            key: `${purpose}/${digest}`,
-          },
-          sha256: digest,
-          byteLength: detached.byteLength,
-        }
-      },
-      read: async (ref) => Uint8Array.from(stored.get(ref.sha256) ?? []),
-    }
-
     const captured = await capturePierTaskOutcome({
       repositoryRoot: root,
       baseCommit,
       baseTree,
       patch: Buffer.alloc(0),
-      artifactPersistence: { executionId: 'pier-large-1', outputArtifacts },
     })
 
     assert.ok(captured.archive.byteLength > 1024 * 1024)
     assert.equal(captured.afterState.files[0]?.byteLength, payload.byteLength)
-    assert.deepEqual(Buffer.from(stored.get(sha256(captured.archive)) ?? []), captured.archive)
   } finally {
     rmSync(root, { recursive: true, force: true })
   }
 })
 
-function sha256(bytes: Uint8Array): `sha256:${string}` {
-  return `sha256:${createHash('sha256').update(bytes).digest('hex')}`
-}
-
 function git(root: string, args: string[]): string {
   return execFileSync('git', ['-C', root, ...args], {
     encoding: 'utf8',
diff --git a/bench/src/pier-task-outcome.ts b/bench/src/pier-task-outcome.ts
index 9ae8b4c1..f9f4d0e1 100644
--- a/bench/src/pier-task-outcome.ts
+++ b/bench/src/pier-task-outcome.ts
@@ -6,7 +6,6 @@ import { join, resolve } from 'node:path'
 
 import {
   type AgentCandidateExecutorTaskOutcomeCapture,
-  type AgentCandidateOutputArtifactPort,
   captureAgentCandidateWorkspaceFiles,
 } from '@tangle-network/agent-runtime/candidate-execution'
 
@@ -22,11 +21,7 @@ export async function capturePierTaskOutcome(input: {
   readonly baseTree: string
   readonly patch: Uint8Array
   readonly signal?: AbortSignal
-  readonly artifactPersistence?: {
-    readonly executionId: string
-    readonly outputArtifacts: AgentCandidateOutputArtifactPort
-  }
-}): Promise<AgentCandidateExecutorTaskOutcomeCapture> {
+}): Promise<Extract<AgentCandidateExecutorTaskOutcomeCapture, { kind: 'workspace' }>> {
   input.signal?.throwIfAborted()
   const repositoryRoot = resolve(input.repositoryRoot)
   const head = (await git(repositoryRoot, ['rev-parse', 'HEAD'], undefined, {}, input.signal)).stdout
@@ -124,16 +119,10 @@ export async function capturePierTaskOutcome(input: {
     // The executor outcome must still return exact bytes; the runtime verifies
     // and persists its final task references after this capture completes.
     const captured = await captureAgentCandidateWorkspaceFiles(files, {
-      ...(input.artifactPersistence
-        ? {
-            artifactPersistence: {
-              ...input.artifactPersistence,
-              ...(input.signal ? { signal: input.signal } : {}),
-            },
-          }
-        : {}),
+      limits: { maxEmbeddedArtifactBytes: 512 * 1024 * 1024 },
     })
     return {
+      kind: 'workspace',
       resultTree,
       afterState: captured.snapshot.material,
       archive: captured.archive,
diff --git a/bench/src/pier-trial-controller.ts b/bench/src/pier-trial-controller.ts
index c818098a..dfbe621c 100644
--- a/bench/src/pier-trial-controller.ts
+++ b/bench/src/pier-trial-controller.ts
@@ -107,6 +107,11 @@ export interface FilePierCandidateTrialControllerOptions {
       readonly deadlineAtMs: number
     },
   ) => PierCandidateProcessSpec
+  /** Restart-safe reader for immutable official bytes under the persisted Pier job identity. */
+  readonly readResult?: (input: {
+    readonly jobsDirectory: string
+    readonly jobName: string
+  }) => Promise<PierCandidateTrialResult | undefined>
   /** Omit only for the default local Docker socket with no environment variables. */
   readonly dockerConnection?: PierDockerConnection
   readonly supervisorPath?: string
@@ -496,6 +501,7 @@ function assertRealDirectory(path: string, label: string): void {
 export class FilePierCandidateTrialController implements PierCandidateTrialController {
   private readonly directory: string
   private readonly launch?: NonNullable<FilePierCandidateTrialControllerOptions['launch']>
+  private readonly readResult?: NonNullable<FilePierCandidateTrialControllerOptions['readResult']>
   private readonly dockerConnection: PierDockerConnection
   private readonly supervisorPath: string
   private readonly pollIntervalMs: number
@@ -503,6 +509,7 @@ export class FilePierCandidateTrialController implements PierCandidateTrialContr
   constructor(options: FilePierCandidateTrialControllerOptions) {
     this.directory = absolutePath(options.directory, 'Pier controller directory')
     this.launch = options.launch
+    this.readResult = options.readResult
     const configuredDocker = options.dockerConnection
     if (configuredDocker?.id === defaultDockerConnectionId) {
       throw new Error(`${defaultDockerConnectionId} is reserved for the implicit local connection`)
@@ -717,6 +724,32 @@ export class FilePierCandidateTrialController implements PierCandidateTrialContr
     return { processExited: true, containersRemoved: true }
   }
 
+  async captureResult(
+    requested: PierCandidateTrialIdentity,
+  ): Promise<PierCandidateTrialResult | undefined> {
+    const controlDirectory = this.controlDirectory(requested)
+    assertRealDirectory(controlDirectory, 'Pier trial control directory')
+    const persisted = parsePersistedIdentity(join(controlDirectory, identityFile))
+    if (
+      persisted.executionId !== requested.executionId ||
+      persisted.executionPlanDigest !== requested.executionPlanDigest
+    ) {
+      throw new Error('persisted Pier trial identity differs from the capture request')
+    }
+    this.assertDockerConnection(persisted)
+    const terminal = parseTerminal(join(controlDirectory, terminalFile))
+    if (!terminal.processExited || !terminal.containersRemoved) {
+      throw new Error('Pier result capture requires proven process and container death')
+    }
+    if (!this.readResult) {
+      throw new Error('Pier result capture has no restart-safe result reader')
+    }
+    return await this.readResult({
+      jobsDirectory: persisted.jobsDirectory,
+      jobName: persisted.jobName,
+    })
+  }
+
   private controlDirectory(identity: PierCandidateTrialIdentity): string {
     return join(this.directory, trialKey(identity))
   }
diff --git a/docs/api/candidate-execution.md b/docs/api/candidate-execution.md
index 80323719..2df4ef62 100644
--- a/docs/api/candidate-execution.md
+++ b/docs/api/candidate-execution.md
@@ -140,12 +140,6 @@ Re-exports [AgentCandidateExecutionTerminalResult](index.md#agentcandidateexecut
 
 ***
 
-### AgentCandidateExecutionUsage
-
-Re-exports [AgentCandidateExecutionUsage](index.md#agentcandidateexecutionusage)
-
-***
-
 ### AgentCandidateRetryRejection
 
 Re-exports [AgentCandidateRetryRejection](index.md#agentcandidateretryrejection)
@@ -338,12 +332,6 @@ Re-exports [AgentCandidateExecutorPort](index.md#agentcandidateexecutorport)
 
 ***
 
-### AgentCandidateExecutorProfileFile
-
-Re-exports [AgentCandidateExecutorProfileFile](index.md#agentcandidateexecutorprofilefile)
-
-***
-
 ### AgentCandidateExecutorRequest
 
 Re-exports [AgentCandidateExecutorRequest](index.md#agentcandidateexecutorrequest)
@@ -416,12 +404,6 @@ Re-exports [AgentCandidateProtectedModelActivation](index.md#agentcandidateprote
 
 ***
 
-### AgentCandidateProtectedModelCall
-
-Re-exports [AgentCandidateProtectedModelCall](index.md#agentcandidateprotectedmodelcall)
-
-***
-
 ### AgentCandidateProtectedModelReservation
 
 Re-exports [AgentCandidateProtectedModelReservation](index.md#agentcandidateprotectedmodelreservation)
@@ -530,6 +512,12 @@ Re-exports [VerifiedAgentCandidateTaskOutcome](index.md#verifiedagentcandidateta
 
 ***
 
+### AGENT\_CANDIDATE\_EXECUTION\_SUPPORT
+
+Re-exports [AGENT_CANDIDATE_EXECUTION_SUPPORT](index.md#agent_candidate_execution_support)
+
+***
+
 ### verifyAgentCandidateBundle
 
 Re-exports [verifyAgentCandidateBundle](index.md#verifyagentcandidatebundle)
diff --git a/docs/api/index.md b/docs/api/index.md
index 0898f0ee..877f9bc9 100644
--- a/docs/api/index.md
+++ b/docs/api/index.md
@@ -183,7 +183,7 @@ independently proved process death plus model and memory closure.
 
 ### InMemoryAgentCandidateExecutionClaimStore
 
-Defined in: [candidate-execution/claim.ts:290](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L290)
+Defined in: [candidate-execution/claim.ts:286](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L286)
 
 Single-process lifecycle implementation.
 
@@ -197,7 +197,7 @@ Single-process lifecycle implementation.
 
 > **new InMemoryAgentCandidateExecutionClaimStore**(`options?`): [`InMemoryAgentCandidateExecutionClaimStore`](#inmemoryagentcandidateexecutionclaimstore)
 
-Defined in: [candidate-execution/claim.ts:296](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L296)
+Defined in: [candidate-execution/claim.ts:292](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L292)
 
 ###### Parameters
 
@@ -215,7 +215,7 @@ Defined in: [candidate-execution/claim.ts:296](https://github.com/tangle-network
 
 > **tryClaim**(`requested`): `Promise`\<[`AgentCandidateExecutionClaimResult`](#agentcandidateexecutionclaimresult)\>
 
-Defined in: [candidate-execution/claim.ts:300](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L300)
+Defined in: [candidate-execution/claim.ts:296](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L296)
 
 ###### Parameters
 
@@ -235,7 +235,7 @@ Defined in: [candidate-execution/claim.ts:300](https://github.com/tangle-network
 
 > **getAttempt**(`requestedAttempt`): `Promise`\<[`AgentCandidateExecutionAttemptRecord`](#agentcandidateexecutionattemptrecord) \| `undefined`\>
 
-Defined in: [candidate-execution/claim.ts:323](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L323)
+Defined in: [candidate-execution/claim.ts:319](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L319)
 
 ###### Parameters
 
@@ -255,7 +255,7 @@ Defined in: [candidate-execution/claim.ts:323](https://github.com/tangle-network
 
 > **markCandidateMayRun**(`requestedLease`): `Promise`\<[`AgentCandidateExecutionPhaseResult`](#agentcandidateexecutionphaseresult)\>
 
-Defined in: [candidate-execution/claim.ts:333](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L333)
+Defined in: [candidate-execution/claim.ts:329](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L329)
 
 Persist the point after which candidate code may have run.
 
@@ -277,7 +277,7 @@ Persist the point after which candidate code may have run.
 
 > **stageTerminal**(`requestedLease`, `result`): `Promise`\<[`AgentCandidateExecutionStageResult`](#agentcandidateexecutionstageresult)\>
 
-Defined in: [candidate-execution/claim.ts:350](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L350)
+Defined in: [candidate-execution/claim.ts:346](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L346)
 
 Fsync the complete terminal record into the durable outbox.
 
@@ -303,7 +303,7 @@ Fsync the complete terminal record into the durable outbox.
 
 > **finish**(`requestedLease`, `requestedTerminalDigest`): `Promise`\<[`AgentCandidateExecutionFinishResult`](#agentcandidateexecutionfinishresult)\>
 
-Defined in: [candidate-execution/claim.ts:365](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L365)
+Defined in: [candidate-execution/claim.ts:361](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L361)
 
 Publish exactly the staged terminal identified by `terminalDigest`.
 
@@ -329,7 +329,7 @@ Publish exactly the staged terminal identified by `terminalDigest`.
 
 > **recoverExpired**(`requestedAttempt`, `evidence`): `Promise`\<[`AgentCandidateExecutionFinishResult`](#agentcandidateexecutionfinishresult)\>
 
-Defined in: [candidate-execution/claim.ts:383](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L383)
+Defined in: [candidate-execution/claim.ts:379](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L379)
 
 Write a failed terminal only after the lease expired and a trusted worker
 independently proved process death plus model and memory closure.
@@ -1357,7 +1357,7 @@ Testable evaluator clock; defaults to `Date.now`.
 
 ### AgentCandidateExecutionCleanupHandles
 
-Defined in: [candidate-execution/claim.ts:39](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L39)
+Defined in: [candidate-execution/claim.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L42)
 
 Non-secret identities a trusted recovery worker needs to close an abandoned attempt.
 
@@ -1367,37 +1367,37 @@ Non-secret identities a trusted recovery worker needs to close an abandoned atte
 
 > `readonly` **preparationId**: `string`
 
-Defined in: [candidate-execution/claim.ts:40](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L40)
+Defined in: [candidate-execution/claim.ts:43](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L43)
 
 ##### modelGrantDigest
 
 > `readonly` **modelGrantDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/claim.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L41)
+Defined in: [candidate-execution/claim.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L44)
 
 ##### resolvedModel
 
 > `readonly` **resolvedModel**: `AgentCandidateResolvedModel`
 
-Defined in: [candidate-execution/claim.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L42)
+Defined in: [candidate-execution/claim.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L45)
 
 ##### traceRunId
 
 > `readonly` **traceRunId**: `string`
 
-Defined in: [candidate-execution/claim.ts:43](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L43)
+Defined in: [candidate-execution/claim.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L46)
 
 ##### cleanupTimeoutMs
 
 > `readonly` **cleanupTimeoutMs**: `number`
 
-Defined in: [candidate-execution/claim.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L44)
+Defined in: [candidate-execution/claim.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L47)
 
 ##### memory?
 
 > `readonly` `optional` **memory?**: `object`
 
-Defined in: [candidate-execution/claim.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L45)
+Defined in: [candidate-execution/claim.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L48)
 
 ###### accessDigest
 
@@ -1411,7 +1411,7 @@ Defined in: [candidate-execution/claim.ts:45](https://github.com/tangle-network/
 
 ### AgentCandidateExecutionClaim
 
-Defined in: [candidate-execution/claim.ts:52](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L52)
+Defined in: [candidate-execution/claim.ts:55](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L55)
 
 Immutable signed identity stored for one execution attempt.
 
@@ -1421,43 +1421,51 @@ Immutable signed identity stored for one execution attempt.
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/claim.ts:53](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L53)
+Defined in: [candidate-execution/claim.ts:56](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L56)
 
 ##### attempt
 
 > `readonly` **attempt**: `number`
 
-Defined in: [candidate-execution/claim.ts:54](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L54)
+Defined in: [candidate-execution/claim.ts:57](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L57)
 
 ##### maxAttempts
 
 > `readonly` **maxAttempts**: `number`
 
-Defined in: [candidate-execution/claim.ts:55](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L55)
+Defined in: [candidate-execution/claim.ts:58](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L58)
 
 ##### retryPolicy
 
 > `readonly` **retryPolicy**: `"none"` \| `"pre-model-infrastructure-only"`
 
-Defined in: [candidate-execution/claim.ts:56](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L56)
+Defined in: [candidate-execution/claim.ts:59](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L59)
 
 ##### bundleDigest
 
 > `readonly` **bundleDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/claim.ts:57](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L57)
+Defined in: [candidate-execution/claim.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L60)
 
 ##### executionPlanDigest
 
 > `readonly` **executionPlanDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/claim.ts:58](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L58)
+Defined in: [candidate-execution/claim.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L61)
+
+##### preparationEvidence
+
+> `readonly` **preparationEvidence**: `AgentCandidatePreparationEvidence`
+
+Defined in: [candidate-execution/claim.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L63)
+
+Durable canonical bytes needed to reconstruct the signed preparation.
 
 ##### retryLineageDigest
 
 > `readonly` **retryLineageDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/claim.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L60)
+Defined in: [candidate-execution/claim.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L65)
 
 Frozen plan identity with only attempt number and per-attempt grant identity normalized.
 
@@ -1465,7 +1473,7 @@ Frozen plan identity with only attempt number and per-attempt grant identity nor
 
 > `readonly` **leaseExpiresAtMs**: `number`
 
-Defined in: [candidate-execution/claim.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L62)
+Defined in: [candidate-execution/claim.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L67)
 
 The winning lease stops authorizing a new terminal write at this instant.
 
@@ -1473,7 +1481,7 @@ The winning lease stops authorizing a new terminal write at this instant.
 
 > `readonly` **resultTimeoutMs**: `number`
 
-Defined in: [candidate-execution/claim.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L64)
+Defined in: [candidate-execution/claim.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L69)
 
 Frozen budget for task verification, executable grading, and receipt construction.
 
@@ -1481,7 +1489,7 @@ Frozen budget for task verification, executable grading, and receipt constructio
 
 > `readonly` **cleanup**: [`AgentCandidateExecutionCleanupHandles`](#agentcandidateexecutioncleanuphandles)
 
-Defined in: [candidate-execution/claim.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L66)
+Defined in: [candidate-execution/claim.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L71)
 
 Non-secret handles retained so an expired attempt can be closed and reconciled.
 
@@ -1489,7 +1497,7 @@ Non-secret handles retained so an expired attempt can be closed and reconciled.
 
 ### AgentCandidateExecutionLease
 
-Defined in: [candidate-execution/claim.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L70)
+Defined in: [candidate-execution/claim.ts:75](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L75)
 
 Secret capability required to finish the acquired attempt.
 
@@ -1499,77 +1507,31 @@ Secret capability required to finish the acquired attempt.
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/claim.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L71)
+Defined in: [candidate-execution/claim.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L76)
 
 ##### attempt
 
 > `readonly` **attempt**: `number`
 
-Defined in: [candidate-execution/claim.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L72)
+Defined in: [candidate-execution/claim.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L77)
 
 ##### token
 
 > `readonly` **token**: `string`
 
-Defined in: [candidate-execution/claim.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L73)
+Defined in: [candidate-execution/claim.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L78)
 
 ##### expiresAtMs
 
 > `readonly` **expiresAtMs**: `number`
 
-Defined in: [candidate-execution/claim.ts:74](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L74)
-
-***
-
-### AgentCandidateExecutionUsage
-
-Defined in: [candidate-execution/claim.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L85)
-
-Exact fixed-point usage proven by the closed evaluator model ledger.
-
-#### Properties
-
-##### costUsdNanos
-
-> `readonly` **costUsdNanos**: `number`
-
-Defined in: [candidate-execution/claim.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L86)
-
-##### inputTokens
-
-> `readonly` **inputTokens**: `number`
-
-Defined in: [candidate-execution/claim.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L87)
-
-##### outputTokens
-
-> `readonly` **outputTokens**: `number`
-
-Defined in: [candidate-execution/claim.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L88)
-
-##### cachedInputTokens
-
-> `readonly` **cachedInputTokens**: `number`
-
-Defined in: [candidate-execution/claim.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L89)
-
-##### reasoningTokens
-
-> `readonly` **reasoningTokens**: `number`
-
-Defined in: [candidate-execution/claim.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L90)
-
-##### modelCalls
-
-> `readonly` **modelCalls**: `number`
-
-Defined in: [candidate-execution/claim.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L91)
+Defined in: [candidate-execution/claim.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L79)
 
 ***
 
 ### AgentCandidateExecutionRecoveryEvidence
 
-Defined in: [candidate-execution/claim.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L128)
+Defined in: [candidate-execution/claim.ts:124](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L124)
 
 Trusted, independently observed closure facts for one expired winning lease.
 
@@ -1579,31 +1541,31 @@ Trusted, independently observed closure facts for one expired winning lease.
 
 > `readonly` **failureClass**: [`AgentCandidateExecutionFailureClass`](#agentcandidateexecutionfailureclass)
 
-Defined in: [candidate-execution/claim.ts:129](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L129)
+Defined in: [candidate-execution/claim.ts:125](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L125)
 
 ##### usage
 
-> `readonly` **usage**: [`AgentCandidateExecutionUsage`](#agentcandidateexecutionusage)
+> `readonly` **usage**: `AgentCandidateFixedSpend`
 
-Defined in: [candidate-execution/claim.ts:130](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L130)
+Defined in: [candidate-execution/claim.ts:126](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L126)
 
 ##### modelSettlement
 
 > `readonly` **modelSettlement**: `AgentCandidateArtifactRef`
 
-Defined in: [candidate-execution/claim.ts:131](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L131)
+Defined in: [candidate-execution/claim.ts:127](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L127)
 
 ##### failureEvidence?
 
 > `readonly` `optional` **failureEvidence?**: `AgentCandidateArtifactRef`
 
-Defined in: [candidate-execution/claim.ts:132](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L132)
+Defined in: [candidate-execution/claim.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L128)
 
 ##### process
 
 > `readonly` **process**: `object`
 
-Defined in: [candidate-execution/claim.ts:133](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L133)
+Defined in: [candidate-execution/claim.ts:129](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L129)
 
 ###### stopped
 
@@ -1617,7 +1579,7 @@ Defined in: [candidate-execution/claim.ts:133](https://github.com/tangle-network
 
 > `readonly` **model**: `object`
 
-Defined in: [candidate-execution/claim.ts:137](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L137)
+Defined in: [candidate-execution/claim.ts:133](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L133)
 
 ###### closed
 
@@ -1635,7 +1597,7 @@ Defined in: [candidate-execution/claim.ts:137](https://github.com/tangle-network
 
 > `readonly` `optional` **memory?**: `object`
 
-Defined in: [candidate-execution/claim.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L142)
+Defined in: [candidate-execution/claim.ts:138](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L138)
 
 ###### closed
 
@@ -1657,7 +1619,7 @@ Defined in: [candidate-execution/claim.ts:142](https://github.com/tangle-network
 
 ### AgentCandidateExecutionAttemptRef
 
-Defined in: [candidate-execution/claim.ts:150](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L150)
+Defined in: [candidate-execution/claim.ts:146](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L146)
 
 #### Properties
 
@@ -1665,19 +1627,19 @@ Defined in: [candidate-execution/claim.ts:150](https://github.com/tangle-network
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/claim.ts:151](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L151)
+Defined in: [candidate-execution/claim.ts:147](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L147)
 
 ##### attempt
 
 > `readonly` **attempt**: `number`
 
-Defined in: [candidate-execution/claim.ts:152](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L152)
+Defined in: [candidate-execution/claim.ts:148](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L148)
 
 ***
 
 ### AgentCandidateExecutionAttemptRecord
 
-Defined in: [candidate-execution/claim.ts:156](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L156)
+Defined in: [candidate-execution/claim.ts:152](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L152)
 
 Persisted state available to a fresh trusted recovery worker after a crash.
 
@@ -1687,19 +1649,19 @@ Persisted state available to a fresh trusted recovery worker after a crash.
 
 > `readonly` **claim**: [`AgentCandidateExecutionClaim`](#agentcandidateexecutionclaim)
 
-Defined in: [candidate-execution/claim.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L157)
+Defined in: [candidate-execution/claim.ts:153](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L153)
 
 ##### phase
 
 > `readonly` **phase**: [`AgentCandidateExecutionPhase`](#agentcandidateexecutionphase)
 
-Defined in: [candidate-execution/claim.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L158)
+Defined in: [candidate-execution/claim.ts:154](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L154)
 
 ##### staged?
 
 > `readonly` `optional` **staged?**: [`AgentCandidateExecutionTerminalRecord`](#agentcandidateexecutionterminalrecord)
 
-Defined in: [candidate-execution/claim.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L160)
+Defined in: [candidate-execution/claim.ts:156](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L156)
 
 Durable outbox content written before the terminal compare-and-set.
 
@@ -1707,13 +1669,13 @@ Durable outbox content written before the terminal compare-and-set.
 
 > `readonly` `optional` **terminal?**: [`AgentCandidateExecutionTerminalRecord`](#agentcandidateexecutionterminalrecord)
 
-Defined in: [candidate-execution/claim.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L161)
+Defined in: [candidate-execution/claim.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L157)
 
 ***
 
 ### AgentCandidateExecutionClaimStore
 
-Defined in: [candidate-execution/claim.ts:233](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L233)
+Defined in: [candidate-execution/claim.ts:229](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L229)
 
 Atomic one-shot store for candidate execution attempts.
 
@@ -1729,7 +1691,7 @@ evidence rather than an ambiguous completed run.
 
 > **tryClaim**(`claim`): `Promise`\<[`AgentCandidateExecutionClaimResult`](#agentcandidateexecutionclaimresult)\>
 
-Defined in: [candidate-execution/claim.ts:234](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L234)
+Defined in: [candidate-execution/claim.ts:230](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L230)
 
 ###### Parameters
 
@@ -1745,7 +1707,7 @@ Defined in: [candidate-execution/claim.ts:234](https://github.com/tangle-network
 
 > **getAttempt**(`attempt`): `Promise`\<[`AgentCandidateExecutionAttemptRecord`](#agentcandidateexecutionattemptrecord) \| `undefined`\>
 
-Defined in: [candidate-execution/claim.ts:235](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L235)
+Defined in: [candidate-execution/claim.ts:231](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L231)
 
 ###### Parameters
 
@@ -1761,7 +1723,7 @@ Defined in: [candidate-execution/claim.ts:235](https://github.com/tangle-network
 
 > **markCandidateMayRun**(`lease`): `Promise`\<[`AgentCandidateExecutionPhaseResult`](#agentcandidateexecutionphaseresult)\>
 
-Defined in: [candidate-execution/claim.ts:239](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L239)
+Defined in: [candidate-execution/claim.ts:235](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L235)
 
 Persist the point after which candidate code may have run.
 
@@ -1779,7 +1741,7 @@ Persist the point after which candidate code may have run.
 
 > **stageTerminal**(`lease`, `result`): `Promise`\<[`AgentCandidateExecutionStageResult`](#agentcandidateexecutionstageresult)\>
 
-Defined in: [candidate-execution/claim.ts:243](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L243)
+Defined in: [candidate-execution/claim.ts:239](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L239)
 
 Fsync the complete terminal record into the durable outbox.
 
@@ -1801,7 +1763,7 @@ Fsync the complete terminal record into the durable outbox.
 
 > **finish**(`lease`, `terminalDigest`): `Promise`\<[`AgentCandidateExecutionFinishResult`](#agentcandidateexecutionfinishresult)\>
 
-Defined in: [candidate-execution/claim.ts:248](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L248)
+Defined in: [candidate-execution/claim.ts:244](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L244)
 
 Publish exactly the staged terminal identified by `terminalDigest`.
 
@@ -1823,7 +1785,7 @@ Publish exactly the staged terminal identified by `terminalDigest`.
 
 > **recoverExpired**(`attempt`, `evidence`): `Promise`\<[`AgentCandidateExecutionFinishResult`](#agentcandidateexecutionfinishresult)\>
 
-Defined in: [candidate-execution/claim.ts:256](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L256)
+Defined in: [candidate-execution/claim.ts:252](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L252)
 
 Write a failed terminal only after the lease expired and a trusted worker
 independently proved process death plus model and memory closure.
@@ -1860,7 +1822,7 @@ Defined in: [candidate-execution/dispose.ts:11](https://github.com/tangle-networ
 
 ### ExecutePreparedAgentCandidateOptions
 
-Defined in: [candidate-execution/execute.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L61)
+Defined in: [candidate-execution/execute.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L63)
 
 #### Properties
 
@@ -1868,31 +1830,31 @@ Defined in: [candidate-execution/execute.ts:61](https://github.com/tangle-networ
 
 > **executor**: [`AgentCandidateExecutorPort`](#agentcandidateexecutorport)
 
-Defined in: [candidate-execution/execute.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L62)
+Defined in: [candidate-execution/execute.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L64)
 
 ##### grader
 
 > **grader**: [`AgentCandidateBenchmarkGraderPort`](#agentcandidatebenchmarkgraderport)
 
-Defined in: [candidate-execution/execute.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L63)
+Defined in: [candidate-execution/execute.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L65)
 
 ##### outputArtifacts
 
 > **outputArtifacts**: [`AgentCandidateOutputArtifactPort`](#agentcandidateoutputartifactport)
 
-Defined in: [candidate-execution/execute.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L64)
+Defined in: [candidate-execution/execute.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L66)
 
 ##### traceStore
 
 > **traceStore**: `TraceStore`
 
-Defined in: [candidate-execution/execute.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L65)
+Defined in: [candidate-execution/execute.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L67)
 
 ##### claimStore
 
 > **claimStore**: [`AgentCandidateExecutionClaimStore`](#agentcandidateexecutionclaimstore)
 
-Defined in: [candidate-execution/execute.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L67)
+Defined in: [candidate-execution/execute.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L69)
 
 Long-lived evaluator-owned store shared by every process that can run this benchmark.
 
@@ -1900,7 +1862,7 @@ Long-lived evaluator-owned store shared by every process that can run this bench
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/execute.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L69)
+Defined in: [candidate-execution/execute.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L71)
 
 Maximum time to prove process death and revoke protected access after a run ends.
 
@@ -1908,7 +1870,7 @@ Maximum time to prove process death and revoke protected access after a run ends
 
 > `optional` **resultTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/execute.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L71)
+Defined in: [candidate-execution/execute.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L73)
 
 Maximum time for task verification, executable grading, and receipt construction.
 
@@ -1916,7 +1878,7 @@ Maximum time for task verification, executable grading, and receipt construction
 
 ### PrepareAgentCandidateExecutionOptions
 
-Defined in: [candidate-execution/prepare.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L91)
+Defined in: [candidate-execution/prepare.ts:75](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L75)
 
 #### Properties
 
@@ -1924,13 +1886,13 @@ Defined in: [candidate-execution/prepare.ts:91](https://github.com/tangle-networ
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/prepare.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L92)
+Defined in: [candidate-execution/prepare.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L76)
 
 ##### resultTimeoutMs?
 
 > `optional` **resultTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/prepare.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L94)
+Defined in: [candidate-execution/prepare.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L78)
 
 Maximum time for task verification, executable grading, and receipt construction.
 
@@ -2121,7 +2083,7 @@ Exact environment names the activation endpoint must return, no more or fewer.
 
 ### RecoverExpiredAgentCandidateOptions
 
-Defined in: [candidate-execution/recover.ts:23](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L23)
+Defined in: [candidate-execution/recover.ts:38](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L38)
 
 #### Properties
 
@@ -2129,49 +2091,49 @@ Defined in: [candidate-execution/recover.ts:23](https://github.com/tangle-networ
 
 > **attempt**: [`AgentCandidateExecutionAttemptRef`](#agentcandidateexecutionattemptref)
 
-Defined in: [candidate-execution/recover.ts:24](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L24)
+Defined in: [candidate-execution/recover.ts:39](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L39)
 
 ##### claimStore
 
 > **claimStore**: [`AgentCandidateExecutionClaimStore`](#agentcandidateexecutionclaimstore)
 
-Defined in: [candidate-execution/recover.ts:25](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L25)
+Defined in: [candidate-execution/recover.ts:40](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L40)
 
 ##### executor
 
 > **executor**: [`AgentCandidateExecutorPort`](#agentcandidateexecutorport)
 
-Defined in: [candidate-execution/recover.ts:26](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L26)
+Defined in: [candidate-execution/recover.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L41)
 
 ##### traceStore
 
 > **traceStore**: `TraceStore`
 
-Defined in: [candidate-execution/recover.ts:27](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L27)
+Defined in: [candidate-execution/recover.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L42)
 
 ##### ports
 
 > **ports**: `Pick`\<[`AgentCandidateExecutionPorts`](#agentcandidateexecutionports), `"models"` \| `"memory"`\>
 
-Defined in: [candidate-execution/recover.ts:28](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L28)
+Defined in: [candidate-execution/recover.ts:43](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L43)
 
 ##### outputArtifacts
 
 > **outputArtifacts**: [`AgentCandidateOutputArtifactPort`](#agentcandidateoutputartifactport)
 
-Defined in: [candidate-execution/recover.ts:29](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L29)
+Defined in: [candidate-execution/recover.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L44)
 
 ##### cleanupTimeoutMs?
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/recover.ts:30](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L30)
+Defined in: [candidate-execution/recover.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L45)
 
 ##### now?
 
 > `optional` **now?**: () => `number`
 
-Defined in: [candidate-execution/recover.ts:32](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L32)
+Defined in: [candidate-execution/recover.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L47)
 
 Evaluator clock; must be the same clock used by the claim store.
 
@@ -2183,7 +2145,7 @@ Evaluator clock; must be the same clock used by the claim store.
 
 ### AgentCandidateArtifactPort
 
-Defined in: [candidate-execution/types.ts:34](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L34)
+Defined in: [candidate-execution/types.ts:40](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L40)
 
 Reads one content-addressed object from the closed S3/IPFS locator set.
 
@@ -2197,7 +2159,7 @@ Reads one content-addressed object from the closed S3/IPFS locator set.
 
 > **read**(`ref`): `Promise`\<`Uint8Array`\<`ArrayBufferLike`\>\>
 
-Defined in: [candidate-execution/types.ts:35](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L35)
+Defined in: [candidate-execution/types.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L41)
 
 ###### Parameters
 
@@ -2213,7 +2175,7 @@ Defined in: [candidate-execution/types.ts:35](https://github.com/tangle-network/
 
 ### AgentCandidateOutputArtifactPort
 
-Defined in: [candidate-execution/types.ts:55](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L55)
+Defined in: [candidate-execution/types.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L67)
 
 Durable content-addressed evidence store controlled only by the evaluator.
 
@@ -2227,7 +2189,7 @@ Durable content-addressed evidence store controlled only by the evaluator.
 
 > **read**(`ref`): `Promise`\<`Uint8Array`\<`ArrayBufferLike`\>\>
 
-Defined in: [candidate-execution/types.ts:35](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L35)
+Defined in: [candidate-execution/types.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L41)
 
 ###### Parameters
 
@@ -2247,7 +2209,7 @@ Defined in: [candidate-execution/types.ts:35](https://github.com/tangle-network/
 
 > **put**(`input`): `Promise`\<`AgentCandidateArtifactRef`\>
 
-Defined in: [candidate-execution/types.ts:57](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L57)
+Defined in: [candidate-execution/types.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L69)
 
 Must be idempotent for identical bytes and return only a durable S3/IPFS locator.
 
@@ -2281,7 +2243,7 @@ Abort must prevent durable publication when it happens before resolution.
 
 ### AgentCandidateRepositoryPort
 
-Defined in: [candidate-execution/types.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L67)
+Defined in: [candidate-execution/types.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L79)
 
 Resolves a declared GitHub repository to an already-present local Git object store.
 
@@ -2291,7 +2253,7 @@ Resolves a declared GitHub repository to an already-present local Git object sto
 
 > **resolve**(`repository`): `Promise`\<`string`\>
 
-Defined in: [candidate-execution/types.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L68)
+Defined in: [candidate-execution/types.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L80)
 
 ###### Parameters
 
@@ -2307,7 +2269,7 @@ Defined in: [candidate-execution/types.ts:68](https://github.com/tangle-network/
 
 ### AgentCandidateVerificationPorts
 
-Defined in: [candidate-execution/types.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L71)
+Defined in: [candidate-execution/types.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L83)
 
 #### Extended by
 
@@ -2319,19 +2281,19 @@ Defined in: [candidate-execution/types.ts:71](https://github.com/tangle-network/
 
 > **artifacts**: [`AgentCandidateArtifactPort`](#agentcandidateartifactport)
 
-Defined in: [candidate-execution/types.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L72)
+Defined in: [candidate-execution/types.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L84)
 
 ##### repositories
 
 > **repositories**: [`AgentCandidateRepositoryPort`](#agentcandidaterepositoryport)
 
-Defined in: [candidate-execution/types.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L73)
+Defined in: [candidate-execution/types.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L85)
 
 ***
 
 ### AgentCandidateWorkspacePort
 
-Defined in: [candidate-execution/types.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L83)
+Defined in: [candidate-execution/types.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L95)
 
 Materializes an already-verified workspace archive.
 
@@ -2345,7 +2307,7 @@ any archive encoding, or no-op when the exact workspace is already present.
 
 > **materialize**(`input`): `Promise`\<`void`\>
 
-Defined in: [candidate-execution/types.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L84)
+Defined in: [candidate-execution/types.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L96)
 
 ###### Parameters
 
@@ -2375,7 +2337,7 @@ Defined in: [candidate-execution/types.ts:84](https://github.com/tangle-network/
 
 ### ResolvedAgentCandidateContainer
 
-Defined in: [candidate-execution/types.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L92)
+Defined in: [candidate-execution/types.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L104)
 
 #### Properties
 
@@ -2383,37 +2345,37 @@ Defined in: [candidate-execution/types.ts:92](https://github.com/tangle-network/
 
 > **source**: `"pinned-container"` \| `"evaluator-task-container"`
 
-Defined in: [candidate-execution/types.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L93)
+Defined in: [candidate-execution/types.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L105)
 
 ##### image
 
 > **image**: `string`
 
-Defined in: [candidate-execution/types.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L94)
+Defined in: [candidate-execution/types.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L106)
 
 ##### indexDigest
 
 > **indexDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L95)
+Defined in: [candidate-execution/types.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L107)
 
 ##### manifestDigest
 
 > **manifestDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L96)
+Defined in: [candidate-execution/types.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L108)
 
 ##### platform
 
 > **platform**: `AgentCandidateOciPlatform`
 
-Defined in: [candidate-execution/types.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L97)
+Defined in: [candidate-execution/types.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L109)
 
 ***
 
 ### AgentCandidateContainerPort
 
-Defined in: [candidate-execution/types.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L100)
+Defined in: [candidate-execution/types.ts:112](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L112)
 
 #### Methods
 
@@ -2421,7 +2383,7 @@ Defined in: [candidate-execution/types.ts:100](https://github.com/tangle-network
 
 > **resolve**(`input`): `Promise`\<[`ResolvedAgentCandidateContainer`](#resolvedagentcandidatecontainer)\>
 
-Defined in: [candidate-execution/types.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L101)
+Defined in: [candidate-execution/types.ts:113](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L113)
 
 ###### Parameters
 
@@ -2443,7 +2405,7 @@ Defined in: [candidate-execution/types.ts:101](https://github.com/tangle-network
 
 ### AgentCandidateModelPort
 
-Defined in: [candidate-execution/types.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L107)
+Defined in: [candidate-execution/types.ts:119](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L119)
 
 #### Methods
 
@@ -2451,7 +2413,7 @@ Defined in: [candidate-execution/types.ts:107](https://github.com/tangle-network
 
 > **resolve**(`input`): `Promise`\<`AgentCandidateResolvedModel`\>
 
-Defined in: [candidate-execution/types.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L108)
+Defined in: [candidate-execution/types.ts:120](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L120)
 
 ###### Parameters
 
@@ -2477,7 +2439,7 @@ Defined in: [candidate-execution/types.ts:108](https://github.com/tangle-network
 
 > **reserveGrant**(`input`): `Promise`\<[`AgentCandidateProtectedModelReservation`](#agentcandidateprotectedmodelreservation)\>
 
-Defined in: [candidate-execution/types.ts:118](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L118)
+Defined in: [candidate-execution/types.ts:130](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L130)
 
 Reserve a stable access identity without creating a live credential.
 The reservation is scoped to `preparationId` and must automatically expire
@@ -2523,7 +2485,7 @@ at `expiresAtMs`, even if this call returns ambiguously to the runtime.
 
 > **activateGrant**(`input`): `Promise`\<[`AgentCandidateProtectedModelActivation`](#agentcandidateprotectedmodelactivation)\>
 
-Defined in: [candidate-execution/types.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L128)
+Defined in: [candidate-execution/types.ts:140](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L140)
 
 Create the live scoped credential only after the execution attempt is durably claimed.
 
@@ -2559,7 +2521,7 @@ Create the live scoped credential only after the execution attempt is durably cl
 
 > **settleGrant**(`input`): `Promise`\<[`AgentCandidateProtectedModelSettlement`](#agentcandidateprotectedmodelsettlement)\>
 
-Defined in: [candidate-execution/types.ts:141](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L141)
+Defined in: [candidate-execution/types.ts:153](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L153)
 
 Atomically revoke the grant, drain in-flight calls, and return its immutable final ledger.
 This operation must be idempotent for the exact preparation and must also
@@ -2598,7 +2560,7 @@ different preparation, even when both reservations report the same digest.
 
 ### AgentCandidateBenchmarkGraderIdentity
 
-Defined in: [candidate-execution/types.ts:156](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L156)
+Defined in: [candidate-execution/types.ts:168](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L168)
 
 #### Properties
 
@@ -2606,25 +2568,25 @@ Defined in: [candidate-execution/types.ts:156](https://github.com/tangle-network
 
 > **name**: `string`
 
-Defined in: [candidate-execution/types.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L157)
+Defined in: [candidate-execution/types.ts:169](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L169)
 
 ##### version
 
 > **version**: `string`
 
-Defined in: [candidate-execution/types.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L158)
+Defined in: [candidate-execution/types.ts:170](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L170)
 
 ##### artifact
 
 > **artifact**: `AgentCandidateArtifactRef`
 
-Defined in: [candidate-execution/types.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L159)
+Defined in: [candidate-execution/types.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L171)
 
 ***
 
 ### AgentCandidateProtectedModelReservation
 
-Defined in: [candidate-execution/types.ts:162](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L162)
+Defined in: [candidate-execution/types.ts:174](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L174)
 
 #### Properties
 
@@ -2632,19 +2594,19 @@ Defined in: [candidate-execution/types.ts:162](https://github.com/tangle-network
 
 > **preparationId**: `string`
 
-Defined in: [candidate-execution/types.ts:163](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L163)
+Defined in: [candidate-execution/types.ts:175](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L175)
 
 ##### digest
 
 > **digest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:164](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L164)
+Defined in: [candidate-execution/types.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L176)
 
 ##### expiresAtMs
 
 > **expiresAtMs**: `number`
 
-Defined in: [candidate-execution/types.ts:166](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L166)
+Defined in: [candidate-execution/types.ts:178](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L178)
 
 Evaluator service must expire and revoke this reservation at this epoch millisecond.
 
@@ -2652,7 +2614,7 @@ Evaluator service must expire and revoke this reservation at this epoch millisec
 
 > **enforcedLimits**: [`AgentCandidateModelLimits`](#agentcandidatemodellimits)
 
-Defined in: [candidate-execution/types.ts:168](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L168)
+Defined in: [candidate-execution/types.ts:180](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L180)
 
 The gateway must stop calls before any one of these limits is exceeded.
 
@@ -2660,7 +2622,7 @@ The gateway must stop calls before any one of these limits is exceeded.
 
 > **network**: `AgentCandidateModelAccessNetwork`
 
-Defined in: [candidate-execution/types.ts:170](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L170)
+Defined in: [candidate-execution/types.ts:182](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L182)
 
 Exact public endpoint exception; every other candidate destination stays blocked.
 
@@ -2668,7 +2630,7 @@ Exact public endpoint exception; every other candidate destination stays blocked
 
 ### AgentCandidateProtectedModelActivation
 
-Defined in: [candidate-execution/types.ts:173](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L173)
+Defined in: [candidate-execution/types.ts:185](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L185)
 
 #### Properties
 
@@ -2676,103 +2638,15 @@ Defined in: [candidate-execution/types.ts:173](https://github.com/tangle-network
 
 > **env**: `Readonly`\<`Record`\<`string`, `string`\>\>
 
-Defined in: [candidate-execution/types.ts:175](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L175)
-
-Injected only into the trusted executor after all pre-launch checks pass.
-
-***
-
-### AgentCandidateProtectedModelCall
-
-Defined in: [candidate-execution/types.ts:179](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L179)
-
-One evaluator-gateway call in the final, revoked model-access ledger.
-
-#### Properties
-
-##### callId
-
-> **callId**: `string`
-
-Defined in: [candidate-execution/types.ts:180](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L180)
-
-##### generationId
-
-> **generationId**: `string`
-
-Defined in: [candidate-execution/types.ts:182](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L182)
-
-Router-generated public response identity.
-
-##### traceSpanId
-
-> **traceSpanId**: `string`
-
-Defined in: [candidate-execution/types.ts:184](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L184)
-
-Exact protected agent-eval LLM span produced from the router ledger.
-
-##### status
-
-> **status**: `"failed"` \| `"succeeded"`
-
-Defined in: [candidate-execution/types.ts:185](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L185)
-
-##### model
-
-> **model**: `string`
-
-Defined in: [candidate-execution/types.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L186)
-
-##### startedAtMs
-
-> **startedAtMs**: `number`
-
 Defined in: [candidate-execution/types.ts:187](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L187)
 
-##### endedAtMs
-
-> **endedAtMs**: `number`
-
-Defined in: [candidate-execution/types.ts:188](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L188)
-
-##### inputTokens
-
-> **inputTokens**: `number`
-
-Defined in: [candidate-execution/types.ts:189](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L189)
-
-##### outputTokens
-
-> **outputTokens**: `number`
-
-Defined in: [candidate-execution/types.ts:190](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L190)
-
-##### cachedInputTokens
-
-> **cachedInputTokens**: `number`
-
-Defined in: [candidate-execution/types.ts:191](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L191)
-
-##### reasoningTokens
-
-> **reasoningTokens**: `number`
-
-Defined in: [candidate-execution/types.ts:192](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L192)
-
-##### costUsdNanos
-
-> **costUsdNanos**: `number`
-
-Defined in: [candidate-execution/types.ts:194](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L194)
-
-Integer billionths of one US dollar; avoids floating-point ledger drift.
+Injected only into the trusted executor after all pre-launch checks pass.
 
 ***
 
 ### AgentCandidateProtectedModelSettlement
 
-Defined in: [candidate-execution/types.ts:197](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L197)
+Defined in: [candidate-execution/types.ts:190](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L190)
 
 #### Properties
 
@@ -2780,31 +2654,31 @@ Defined in: [candidate-execution/types.ts:197](https://github.com/tangle-network
 
 > **preparationId**: `string`
 
-Defined in: [candidate-execution/types.ts:198](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L198)
+Defined in: [candidate-execution/types.ts:191](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L191)
 
 ##### grantDigest
 
 > **grantDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L199)
+Defined in: [candidate-execution/types.ts:192](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L192)
 
 ##### closed
 
 > **closed**: `true`
 
-Defined in: [candidate-execution/types.ts:200](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L200)
+Defined in: [candidate-execution/types.ts:193](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L193)
 
 ##### calls
 
-> **calls**: readonly [`AgentCandidateProtectedModelCall`](#agentcandidateprotectedmodelcall)[]
+> **calls**: readonly `AgentCandidateModelSettlementCall`[]
 
-Defined in: [candidate-execution/types.ts:201](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L201)
+Defined in: [candidate-execution/types.ts:194](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L194)
 
 ***
 
 ### AgentCandidateMemoryResetResult
 
-Defined in: [candidate-execution/types.ts:204](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L204)
+Defined in: [candidate-execution/types.ts:197](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L197)
 
 #### Properties
 
@@ -2812,43 +2686,43 @@ Defined in: [candidate-execution/types.ts:204](https://github.com/tangle-network
 
 > **preparationId**: `string`
 
-Defined in: [candidate-execution/types.ts:205](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L205)
+Defined in: [candidate-execution/types.ts:198](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L198)
 
 ##### accessDigest
 
 > **accessDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:206](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L206)
+Defined in: [candidate-execution/types.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L199)
 
 ##### expiresAtMs
 
 > **expiresAtMs**: `number`
 
-Defined in: [candidate-execution/types.ts:207](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L207)
+Defined in: [candidate-execution/types.ts:200](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L200)
 
 ##### evidence
 
 > **evidence**: `AgentCandidateCapturedArtifact`
 
-Defined in: [candidate-execution/types.ts:208](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L208)
+Defined in: [candidate-execution/types.ts:201](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L201)
 
 ##### emptyStateDigest
 
 > **emptyStateDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:209](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L209)
+Defined in: [candidate-execution/types.ts:202](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L202)
 
 ##### beforeState
 
 > **beforeState**: `AgentCandidateWorkspaceSnapshotEvidence`
 
-Defined in: [candidate-execution/types.ts:210](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L210)
+Defined in: [candidate-execution/types.ts:203](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L203)
 
 ***
 
 ### AgentCandidateMemoryPort
 
-Defined in: [candidate-execution/types.ts:213](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L213)
+Defined in: [candidate-execution/types.ts:206](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L206)
 
 #### Methods
 
@@ -2856,7 +2730,7 @@ Defined in: [candidate-execution/types.ts:213](https://github.com/tangle-network
 
 > **reset**(`input`): `Promise`\<[`AgentCandidateMemoryResetResult`](#agentcandidatememoryresetresult)\>
 
-Defined in: [candidate-execution/types.ts:219](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L219)
+Defined in: [candidate-execution/types.ts:212](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L212)
 
 Reset and reserve exact task memory without returning live access.
 The service must scope the reservation to `preparationId`, automatically
@@ -2898,7 +2772,7 @@ revoke it at `expiresAtMs`, and never reuse it for another preparation.
 
 > **activate**(`input`): `Promise`\<\{ `env`: `Readonly`\<`Record`\<`string`, `string`\>\>; \}\>
 
-Defined in: [candidate-execution/types.ts:231](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L231)
+Defined in: [candidate-execution/types.ts:224](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L224)
 
 Create live scoped access only after the execution attempt is durably claimed.
 Activation must match the exact preparation/access pair and may not extend expiry.
@@ -2935,7 +2809,7 @@ Activation must match the exact preparation/access pair and may not extend expir
 
 > **close**(`input`): `Promise`\<\{ `closed`: `true`; \}\>
 
-Defined in: [candidate-execution/types.ts:243](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L243)
+Defined in: [candidate-execution/types.ts:236](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L236)
 
 Revoke evaluator-owned access after process death or a failed preparation.
 Must be idempotent and concurrency-safe for the exact preparation/access
@@ -2973,7 +2847,7 @@ pair and must never close a different preparation.
 
 ### AgentCandidateExecutionPorts
 
-Defined in: [candidate-execution/types.ts:252](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L252)
+Defined in: [candidate-execution/types.ts:245](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L245)
 
 #### Extends
 
@@ -2985,7 +2859,7 @@ Defined in: [candidate-execution/types.ts:252](https://github.com/tangle-network
 
 > **artifacts**: [`AgentCandidateArtifactPort`](#agentcandidateartifactport)
 
-Defined in: [candidate-execution/types.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L72)
+Defined in: [candidate-execution/types.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L84)
 
 ###### Inherited from
 
@@ -2995,7 +2869,7 @@ Defined in: [candidate-execution/types.ts:72](https://github.com/tangle-network/
 
 > **repositories**: [`AgentCandidateRepositoryPort`](#agentcandidaterepositoryport)
 
-Defined in: [candidate-execution/types.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L73)
+Defined in: [candidate-execution/types.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L85)
 
 ###### Inherited from
 
@@ -3005,31 +2879,33 @@ Defined in: [candidate-execution/types.ts:73](https://github.com/tangle-network/
 
 > **workspaces**: [`AgentCandidateWorkspacePort`](#agentcandidateworkspaceport)
 
-Defined in: [candidate-execution/types.ts:253](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L253)
+Defined in: [candidate-execution/types.ts:246](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L246)
 
 ##### containers
 
 > **containers**: [`AgentCandidateContainerPort`](#agentcandidatecontainerport)
 
-Defined in: [candidate-execution/types.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L254)
+Defined in: [candidate-execution/types.ts:247](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L247)
 
 ##### models
 
 > **models**: [`AgentCandidateModelPort`](#agentcandidatemodelport)
 
-Defined in: [candidate-execution/types.ts:255](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L255)
+Defined in: [candidate-execution/types.ts:248](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L248)
 
 ##### memory
 
 > **memory**: [`AgentCandidateMemoryPort`](#agentcandidatememoryport)
 
-Defined in: [candidate-execution/types.ts:256](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L256)
+Defined in: [candidate-execution/types.ts:249](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L249)
 
 ***
 
 ### AgentCandidateTaskExecution
 
-Defined in: [candidate-execution/types.ts:259](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L259)
+Defined in: [candidate-execution/types.ts:253](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L253)
+
+One signed benchmark task and the exact result shape its executor must capture.
 
 #### Properties
 
@@ -3037,73 +2913,65 @@ Defined in: [candidate-execution/types.ts:259](https://github.com/tangle-network
 
 > **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:260](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L260)
+Defined in: [candidate-execution/types.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L254)
 
 ##### benchmark
 
 > **benchmark**: `string`
 
-Defined in: [candidate-execution/types.ts:261](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L261)
+Defined in: [candidate-execution/types.ts:255](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L255)
 
 ##### benchmarkVersion
 
 > **benchmarkVersion**: `string`
 
-Defined in: [candidate-execution/types.ts:262](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L262)
+Defined in: [candidate-execution/types.ts:256](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L256)
 
 ##### taskId
 
 > **taskId**: `string`
 
-Defined in: [candidate-execution/types.ts:263](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L263)
+Defined in: [candidate-execution/types.ts:257](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L257)
 
 ##### splitDigest
 
 > **splitDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L264)
+Defined in: [candidate-execution/types.ts:258](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L258)
 
 ##### instruction
 
 > **instruction**: `string`
 
-Defined in: [candidate-execution/types.ts:266](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L266)
+Defined in: [candidate-execution/types.ts:260](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L260)
 
 Exact agent-visible task instruction. The runtime rejects malformed Unicode.
 
-##### repository
-
-> **repository**: `object`
-
-Defined in: [candidate-execution/types.ts:267](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L267)
-
-###### identity
-
-> **identity**: `string`
+##### repository?
 
-###### rootIdentity
+> `optional` **repository?**: `AgentCandidateTaskRepository`
 
-> **rootIdentity**: `string`
+Defined in: [candidate-execution/types.ts:262](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L262)
 
-###### baseCommit
+Optional source identity, required when the expected outcome is a workspace.
 
-> **baseCommit**: `string`
+##### outcome
 
-###### baseTree
+> **outcome**: `AgentCandidateTaskOutcomeSpec`
 
-> **baseTree**: `string`
+Defined in: [candidate-execution/types.ts:263](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L263)
 
 ##### attempt
 
 > **attempt**: `AgentCandidateAttemptPolicy`
 
-Defined in: [candidate-execution/types.ts:273](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L273)
+Defined in: [candidate-execution/types.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L264)
 
 ##### model
 
 > **model**: `object`
 
-Defined in: [candidate-execution/types.ts:274](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L274)
+Defined in: [candidate-execution/types.ts:265](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L265)
 
 ###### requested
 
@@ -3117,13 +2985,13 @@ Defined in: [candidate-execution/types.ts:274](https://github.com/tangle-network
 
 > **grader**: [`AgentCandidateBenchmarkGraderIdentity`](#agentcandidatebenchmarkgraderidentity)
 
-Defined in: [candidate-execution/types.ts:278](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L278)
+Defined in: [candidate-execution/types.ts:269](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L269)
 
 ##### executionRoots
 
 > **executionRoots**: `object`
 
-Defined in: [candidate-execution/types.ts:280](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L280)
+Defined in: [candidate-execution/types.ts:271](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L271)
 
 Absolute paths inside the evaluator-owned execution environment.
 
@@ -3139,7 +3007,7 @@ Absolute paths inside the evaluator-owned execution environment.
 
 > **stagingRoots**: `object`
 
-Defined in: [candidate-execution/types.ts:285](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L285)
+Defined in: [candidate-execution/types.ts:276](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L276)
 
 Host-side staging roots. These are verified but never signed as container paths.
 
@@ -3159,51 +3027,51 @@ Host-side staging roots. These are verified but never signed as container paths.
 
 > **workspace**: `AgentCandidateWorkspaceSnapshotEvidence`
 
-Defined in: [candidate-execution/types.ts:290](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L290)
+Defined in: [candidate-execution/types.ts:281](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L281)
 
 ##### evaluatorTaskContainer?
 
 > `optional` **evaluatorTaskContainer?**: [`ResolvedAgentCandidateContainer`](#resolvedagentcandidatecontainer)
 
-Defined in: [candidate-execution/types.ts:291](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L291)
+Defined in: [candidate-execution/types.ts:282](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L282)
 
 ##### limits
 
 > **limits**: `AgentCandidateExecutionLimits`
 
-Defined in: [candidate-execution/types.ts:292](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L292)
+Defined in: [candidate-execution/types.ts:283](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L283)
 
 ***
 
 ### VerifiedAgentCandidate
 
-Defined in: [candidate-execution/types.ts:295](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L295)
+Defined in: [candidate-execution/types.ts:286](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L286)
 
 #### Properties
 
 ##### bundle
 
-> `readonly` **bundle**: `AgentCandidateBundleV1`
+> `readonly` **bundle**: `AgentCandidateBundle`
 
-Defined in: [candidate-execution/types.ts:296](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L296)
+Defined in: [candidate-execution/types.ts:287](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L287)
 
 ##### materializedTree?
 
 > `readonly` `optional` **materializedTree?**: `string`
 
-Defined in: [candidate-execution/types.ts:297](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L297)
+Defined in: [candidate-execution/types.ts:288](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L288)
 
 ##### \[verifiedCandidateBrand\]
 
 > `readonly` **\[verifiedCandidateBrand\]**: `true`
 
-Defined in: [candidate-execution/types.ts:298](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L298)
+Defined in: [candidate-execution/types.ts:289](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L289)
 
 ***
 
 ### CanonicalCandidateDocument
 
-Defined in: [candidate-execution/types.ts:301](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L301)
+Defined in: [candidate-execution/types.ts:292](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L292)
 
 #### Type Parameters
 
@@ -3217,13 +3085,13 @@ Defined in: [candidate-execution/types.ts:301](https://github.com/tangle-network
 
 > `readonly` **value**: `T`
 
-Defined in: [candidate-execution/types.ts:302](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L302)
+Defined in: [candidate-execution/types.ts:293](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L293)
 
 ##### bytes
 
 > `readonly` **bytes**: `Uint8Array`
 
-Defined in: [candidate-execution/types.ts:304](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L304)
+Defined in: [candidate-execution/types.ts:295](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L295)
 
 Canonical UTF-8 bytes of `value` with its top-level digest omitted.
 
@@ -3231,13 +3099,13 @@ Canonical UTF-8 bytes of `value` with its top-level digest omitted.
 
 > `readonly` **digest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:305](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L305)
+Defined in: [candidate-execution/types.ts:296](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L296)
 
 ***
 
 ### PreparedAgentCandidateLaunch
 
-Defined in: [candidate-execution/types.ts:308](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L308)
+Defined in: [candidate-execution/types.ts:299](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L299)
 
 #### Properties
 
@@ -3245,13 +3113,13 @@ Defined in: [candidate-execution/types.ts:308](https://github.com/tangle-network
 
 > **executable**: `string`
 
-Defined in: [candidate-execution/types.ts:309](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L309)
+Defined in: [candidate-execution/types.ts:300](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L300)
 
 ##### args
 
 > **args**: readonly `string`[]
 
-Defined in: [candidate-execution/types.ts:311](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L311)
+Defined in: [candidate-execution/types.ts:302](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L302)
 
 Complete fixed argv, including profile materializer flags but excluding task delivery.
 
@@ -3259,13 +3127,13 @@ Complete fixed argv, including profile materializer flags but excluding task del
 
 > **env**: `Readonly`\<`Record`\<`string`, `string`\>\>
 
-Defined in: [candidate-execution/types.ts:312](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L312)
+Defined in: [candidate-execution/types.ts:303](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L303)
 
 ##### flags
 
 > **flags**: readonly `string`[]
 
-Defined in: [candidate-execution/types.ts:314](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L314)
+Defined in: [candidate-execution/types.ts:305](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L305)
 
 Informational subset already present at the tail of `args`; executors must not append twice.
 
@@ -3273,13 +3141,13 @@ Informational subset already present at the tail of `args`; executors must not a
 
 > **cwd**: `string`
 
-Defined in: [candidate-execution/types.ts:315](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L315)
+Defined in: [candidate-execution/types.ts:306](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L306)
 
 ***
 
 ### PreparedAgentCandidateInstruction
 
-Defined in: [candidate-execution/types.ts:318](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L318)
+Defined in: [candidate-execution/types.ts:309](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L309)
 
 #### Properties
 
@@ -3287,19 +3155,19 @@ Defined in: [candidate-execution/types.ts:318](https://github.com/tangle-network
 
 > **bytes**: `Uint8Array`
 
-Defined in: [candidate-execution/types.ts:319](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L319)
+Defined in: [candidate-execution/types.ts:310](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L310)
 
 ##### delivery
 
 > **delivery**: `AgentCandidateInstructionDelivery`
 
-Defined in: [candidate-execution/types.ts:320](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L320)
+Defined in: [candidate-execution/types.ts:311](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L311)
 
 ***
 
 ### PreparedAgentCandidateTrace
 
-Defined in: [candidate-execution/types.ts:323](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L323)
+Defined in: [candidate-execution/types.ts:314](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L314)
 
 #### Properties
 
@@ -3307,45 +3175,45 @@ Defined in: [candidate-execution/types.ts:323](https://github.com/tangle-network
 
 > **runId**: `string`
 
-Defined in: [candidate-execution/types.ts:324](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L324)
+Defined in: [candidate-execution/types.ts:315](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L315)
 
 ##### tags
 
 > **tags**: `Readonly`\<`Record`\<`string`, `string`\>\>
 
-Defined in: [candidate-execution/types.ts:325](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L325)
+Defined in: [candidate-execution/types.ts:316](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L316)
 
 ##### env
 
 > **env**: `Readonly`\<`Record`\<`string`, `string`\>\>
 
-Defined in: [candidate-execution/types.ts:326](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L326)
+Defined in: [candidate-execution/types.ts:317](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L317)
 
 ***
 
 ### PreparedAgentCandidateExecution
 
-Defined in: [candidate-execution/types.ts:329](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L329)
+Defined in: [candidate-execution/types.ts:320](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L320)
 
 #### Properties
 
 ##### bundle
 
-> `readonly` **bundle**: `AgentCandidateBundleV1`
+> `readonly` **bundle**: `AgentCandidateBundle`
 
-Defined in: [candidate-execution/types.ts:330](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L330)
+Defined in: [candidate-execution/types.ts:321](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L321)
 
 ##### executionId
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:331](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L331)
+Defined in: [candidate-execution/types.ts:322](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L322)
 
 ##### roots
 
 > `readonly` **roots**: `object`
 
-Defined in: [candidate-execution/types.ts:332](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L332)
+Defined in: [candidate-execution/types.ts:323](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L323)
 
 ###### execution
 
@@ -3379,7 +3247,7 @@ Defined in: [candidate-execution/types.ts:332](https://github.com/tangle-network
 
 > `readonly` **profilePlan**: `object`
 
-Defined in: [candidate-execution/types.ts:343](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L343)
+Defined in: [candidate-execution/types.ts:334](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L334)
 
 ###### value
 
@@ -3393,11 +3261,17 @@ Defined in: [candidate-execution/types.ts:343](https://github.com/tangle-network
 
 > **written**: readonly `string`[]
 
+##### profileActivation
+
+> `readonly` **profileActivation**: `AgentCandidateProfileActivation`
+
+Defined in: [candidate-execution/types.ts:339](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L339)
+
 ##### executionPlan
 
 > `readonly` **executionPlan**: `object`
 
-Defined in: [candidate-execution/types.ts:348](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L348)
+Defined in: [candidate-execution/types.ts:340](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L340)
 
 ###### value
 
@@ -3409,69 +3283,51 @@ Defined in: [candidate-execution/types.ts:348](https://github.com/tangle-network
 
 ##### materializationReceipt
 
-> `readonly` **materializationReceipt**: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateMaterializationReceiptV1`\>
+> `readonly` **materializationReceipt**: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateMaterializationReceipt`\>
 
-Defined in: [candidate-execution/types.ts:352](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L352)
+Defined in: [candidate-execution/types.ts:344](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L344)
 
 ##### launch
 
 > `readonly` **launch**: [`PreparedAgentCandidateLaunch`](#preparedagentcandidatelaunch)
 
-Defined in: [candidate-execution/types.ts:353](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L353)
+Defined in: [candidate-execution/types.ts:345](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L345)
 
 ##### instruction
 
 > `readonly` **instruction**: [`PreparedAgentCandidateInstruction`](#preparedagentcandidateinstruction)
 
-Defined in: [candidate-execution/types.ts:354](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L354)
+Defined in: [candidate-execution/types.ts:346](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L346)
 
 ##### resolvedModel
 
 > `readonly` **resolvedModel**: `AgentCandidateResolvedModel`
 
-Defined in: [candidate-execution/types.ts:355](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L355)
-
-##### knowledge?
-
-> `readonly` `optional` **knowledge?**: `object`
-
-Defined in: [candidate-execution/types.ts:356](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L356)
-
-###### snapshotId
-
-> **snapshotId**: `string`
-
-###### manifestDigest
-
-> **manifestDigest**: `` `sha256:${string}` ``
-
-###### manifest
-
-> **manifest**: `Uint8Array`
+Defined in: [candidate-execution/types.ts:347](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L347)
 
 ##### trace
 
 > `readonly` **trace**: [`PreparedAgentCandidateTrace`](#preparedagentcandidatetrace)
 
-Defined in: [candidate-execution/types.ts:361](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L361)
+Defined in: [candidate-execution/types.ts:348](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L348)
 
 ##### memory
 
 > `readonly` **memory**: `AgentCandidateEffectiveMemory`
 
-Defined in: [candidate-execution/types.ts:362](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L362)
+Defined in: [candidate-execution/types.ts:349](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L349)
 
 ##### \[preparedCandidateBrand\]
 
 > `readonly` **\[preparedCandidateBrand\]**: `true`
 
-Defined in: [candidate-execution/types.ts:363](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L363)
+Defined in: [candidate-execution/types.ts:350](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L350)
 
 ***
 
 ### AgentCandidateProtectedRunCapture
 
-Defined in: [candidate-execution/types.ts:366](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L366)
+Defined in: [candidate-execution/types.ts:353](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L353)
 
 #### Properties
 
@@ -3479,61 +3335,19 @@ Defined in: [candidate-execution/types.ts:366](https://github.com/tangle-network
 
 > **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:367](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L367)
+Defined in: [candidate-execution/types.ts:354](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L354)
 
 ##### termination
 
 > **termination**: `AgentCandidateTermination`
 
-Defined in: [candidate-execution/types.ts:368](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L368)
-
-***
-
-### AgentCandidateExecutorTaskOutcomeCapture
-
-Defined in: [candidate-execution/types.ts:372](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L372)
-
-Raw evaluator capture made only after the candidate process is dead.
-
-#### Properties
-
-##### resultTree
-
-> **resultTree**: `string`
-
-Defined in: [candidate-execution/types.ts:374](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L374)
-
-Claimed final tree. The runtime recomputes it independently from `gitDiff`.
-
-##### afterState
-
-> **afterState**: `AgentCandidateWorkspaceManifestMaterialV1`
-
-Defined in: [candidate-execution/types.ts:376](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L376)
-
-Complete evaluator-captured workspace description after candidate execution.
-
-##### archive
-
-> **archive**: `Uint8Array`
-
-Defined in: [candidate-execution/types.ts:378](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L378)
-
-Reproducible workspace archive corresponding to `afterState`.
-
-##### gitDiff
-
-> **gitDiff**: `Uint8Array`
-
-Defined in: [candidate-execution/types.ts:380](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L380)
-
-Exact binary patch from the signed task base to `afterState`.
+Defined in: [candidate-execution/types.ts:355](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L355)
 
 ***
 
 ### AgentCandidateExecutorMemoryCapture
 
-Defined in: [candidate-execution/types.ts:384](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L384)
+Defined in: [candidate-execution/types.ts:378](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L378)
 
 Raw isolated-memory capture made only after access has been revoked.
 
@@ -3541,85 +3355,53 @@ Raw isolated-memory capture made only after access has been revoked.
 
 ##### afterState
 
-> `readonly` **afterState**: `AgentCandidateWorkspaceManifestMaterialV1`
+> `readonly` **afterState**: `AgentCandidateWorkspaceManifestMaterial`
 
-Defined in: [candidate-execution/types.ts:385](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L385)
+Defined in: [candidate-execution/types.ts:379](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L379)
 
 ##### archive
 
 > `readonly` **archive**: `Uint8Array`
 
-Defined in: [candidate-execution/types.ts:386](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L386)
+Defined in: [candidate-execution/types.ts:380](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L380)
 
 ***
 
 ### AgentCandidateExecutorFinalCapture
 
-Defined in: [candidate-execution/types.ts:390](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L390)
+Defined in: [candidate-execution/types.ts:384](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L384)
 
-Idempotent executor result after process death and trace drain.
+Replayable evaluator result captured only after process death and trace drain.
 
 #### Properties
 
-##### stopped
-
-> `readonly` **stopped**: `true`
-
-Defined in: [candidate-execution/types.ts:391](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L391)
-
 ##### taskOutcome?
 
 > `readonly` `optional` **taskOutcome?**: [`AgentCandidateExecutorTaskOutcomeCapture`](#agentcandidateexecutortaskoutcomecapture)
 
-Defined in: [candidate-execution/types.ts:392](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L392)
+Defined in: [candidate-execution/types.ts:385](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L385)
 
 ##### memoryAfter?
 
 > `readonly` `optional` **memoryAfter?**: [`AgentCandidateExecutorMemoryCapture`](#agentcandidateexecutormemorycapture)
 
-Defined in: [candidate-execution/types.ts:394](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L394)
+Defined in: [candidate-execution/types.ts:387](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L387)
 
 Required only when the prepared candidate uses isolated task memory.
 
-***
-
-### VerifiedAgentCandidateTaskOutcome
-
-Defined in: [candidate-execution/types.ts:398](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L398)
-
-Branded task outcome that has survived independent patch and tree verification.
-
-#### Properties
-
-##### evidence
-
-> `readonly` **evidence**: `AgentCandidateTaskOutcomeEvidence` & `object`
-
-Defined in: [candidate-execution/types.ts:399](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L399)
-
-###### Type Declaration
+##### evidence?
 
-###### artifact
+> `readonly` `optional` **evidence?**: `Uint8Array`\<`ArrayBufferLike`\>
 
-> `readonly` **artifact**: `AgentCandidateArtifactRef`
-
-##### patch
+Defined in: [candidate-execution/types.ts:389](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L389)
 
-> `readonly` **patch**: `Uint8Array`
-
-Defined in: [candidate-execution/types.ts:402](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L402)
-
-##### \[verifiedTaskOutcomeBrand\]
-
-> `readonly` **\[verifiedTaskOutcomeBrand\]**: `true`
-
-Defined in: [candidate-execution/types.ts:403](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L403)
+Executor-native bytes preserved when a fresh worker cannot reconstruct a verified outcome.
 
 ***
 
 ### AgentCandidateBenchmarkGraderPort
 
-Defined in: [candidate-execution/types.ts:415](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L415)
+Defined in: [candidate-execution/types.ts:426](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L426)
 
 Evaluator-owned executable grader, pinned by immutable implementation bytes.
 
@@ -3635,19 +3417,19 @@ copying an expected digest from ambient configuration.
 
 > `readonly` **name**: `string`
 
-Defined in: [candidate-execution/types.ts:416](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L416)
+Defined in: [candidate-execution/types.ts:427](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L427)
 
 ##### version
 
 > `readonly` **version**: `string`
 
-Defined in: [candidate-execution/types.ts:417](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L417)
+Defined in: [candidate-execution/types.ts:428](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L428)
 
 ##### artifact
 
 > `readonly` **artifact**: `AgentCandidateArtifactRef`
 
-Defined in: [candidate-execution/types.ts:418](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L418)
+Defined in: [candidate-execution/types.ts:429](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L429)
 
 #### Methods
 
@@ -3655,7 +3437,7 @@ Defined in: [candidate-execution/types.ts:418](https://github.com/tangle-network
 
 > **run**(`input`): `Promise`\<\{ `evaluation`: `BenchmarkEvaluation`; `evidence`: `Uint8Array`; `binding`: \{ `implementationDigest`: `` `sha256:${string}` ``; `taskOutcomeDigest`: `` `sha256:${string}` ``; `outputDigest`: `` `sha256:${string}` ``; \}; \}\>
 
-Defined in: [candidate-execution/types.ts:419](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L419)
+Defined in: [candidate-execution/types.ts:430](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L430)
 
 ###### Parameters
 
@@ -3701,7 +3483,7 @@ Frozen result deadline; runners must stop work and side effects when aborted.
 
 ### AgentCandidateExecutorRequest
 
-Defined in: [candidate-execution/types.ts:447](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L447)
+Defined in: [candidate-execution/types.ts:458](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L458)
 
 One detached request passed to the trusted environment-specific executor.
 
@@ -3711,13 +3493,13 @@ One detached request passed to the trusted environment-specific executor.
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:448](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L448)
+Defined in: [candidate-execution/types.ts:459](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L459)
 
 ##### inputs
 
 > `readonly` **inputs**: `object`
 
-Defined in: [candidate-execution/types.ts:450](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L450)
+Defined in: [candidate-execution/types.ts:461](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L461)
 
 Immutable bytes from which the executor creates fresh isolated workspaces.
 
@@ -3729,19 +3511,11 @@ Immutable bytes from which the executor creates fresh isolated workspaces.
 
 > `readonly` `optional` **candidate?**: [`AgentCandidateExecutorWorkspaceInput`](#agentcandidateexecutorworkspaceinput)
 
-###### profile
-
-> `readonly` **profile**: `object`
-
-###### profile.files
-
-> `readonly` **files**: readonly [`AgentCandidateExecutorProfileFile`](#agentcandidateexecutorprofilefile)[]
-
 ##### roots
 
 > `readonly` **roots**: `object`
 
-Defined in: [candidate-execution/types.ts:457](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L457)
+Defined in: [candidate-execution/types.ts:465](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L465)
 
 ###### taskRoot
 
@@ -3755,7 +3529,7 @@ Defined in: [candidate-execution/types.ts:457](https://github.com/tangle-network
 
 > `readonly` **profilePlan**: `object`
 
-Defined in: [candidate-execution/types.ts:458](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L458)
+Defined in: [candidate-execution/types.ts:466](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L466)
 
 ###### value
 
@@ -3769,11 +3543,17 @@ Defined in: [candidate-execution/types.ts:458](https://github.com/tangle-network
 
 > **written**: readonly `string`[]
 
+##### profileActivation
+
+> `readonly` **profileActivation**: `AgentCandidateProfileActivation`
+
+Defined in: [candidate-execution/types.ts:467](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L467)
+
 ##### executionPlan
 
 > `readonly` **executionPlan**: `object`
 
-Defined in: [candidate-execution/types.ts:459](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L459)
+Defined in: [candidate-execution/types.ts:468](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L468)
 
 ###### value
 
@@ -3785,33 +3565,33 @@ Defined in: [candidate-execution/types.ts:459](https://github.com/tangle-network
 
 ##### materializationReceipt
 
-> `readonly` **materializationReceipt**: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateMaterializationReceiptV1`\>
+> `readonly` **materializationReceipt**: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateMaterializationReceipt`\>
 
-Defined in: [candidate-execution/types.ts:460](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L460)
+Defined in: [candidate-execution/types.ts:469](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L469)
 
 ##### launch
 
 > `readonly` **launch**: [`PreparedAgentCandidateLaunch`](#preparedagentcandidatelaunch)
 
-Defined in: [candidate-execution/types.ts:461](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L461)
+Defined in: [candidate-execution/types.ts:470](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L470)
 
 ##### instruction
 
 > `readonly` **instruction**: [`PreparedAgentCandidateInstruction`](#preparedagentcandidateinstruction)
 
-Defined in: [candidate-execution/types.ts:462](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L462)
+Defined in: [candidate-execution/types.ts:471](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L471)
 
 ##### resolvedModel
 
 > `readonly` **resolvedModel**: `AgentCandidateResolvedModel`
 
-Defined in: [candidate-execution/types.ts:463](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L463)
+Defined in: [candidate-execution/types.ts:472](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L472)
 
 ##### hardLimits
 
 > `readonly` **hardLimits**: `Pick`\<`AgentCandidateExecutionLimits`, `"timeoutMs"`\>
 
-Defined in: [candidate-execution/types.ts:465](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L465)
+Defined in: [candidate-execution/types.ts:474](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L474)
 
 Mechanically enforced by the runtime plus executor process-death acknowledgement.
 
@@ -3819,45 +3599,27 @@ Mechanically enforced by the runtime plus executor process-death acknowledgement
 
 > `readonly` **observedLimits**: `Pick`\<`AgentCandidateExecutionLimits`, `"maxSteps"`\>
 
-Defined in: [candidate-execution/types.ts:467](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L467)
+Defined in: [candidate-execution/types.ts:476](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L476)
 
 Validity bound checked against protected traces; generic black-box executors cannot preempt it.
 
-##### knowledge?
-
-> `readonly` `optional` **knowledge?**: `object`
-
-Defined in: [candidate-execution/types.ts:468](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L468)
-
-###### snapshotId
-
-> **snapshotId**: `string`
-
-###### manifestDigest
-
-> **manifestDigest**: `` `sha256:${string}` ``
-
-###### manifest
-
-> **manifest**: `Uint8Array`
-
 ##### trace
 
 > `readonly` **trace**: [`PreparedAgentCandidateTrace`](#preparedagentcandidatetrace)
 
-Defined in: [candidate-execution/types.ts:469](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L469)
+Defined in: [candidate-execution/types.ts:477](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L477)
 
 ##### memory
 
 > `readonly` **memory**: `AgentCandidateEffectiveMemory`
 
-Defined in: [candidate-execution/types.ts:470](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L470)
+Defined in: [candidate-execution/types.ts:478](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L478)
 
 ***
 
 ### AgentCandidateExecutorPort
 
-Defined in: [candidate-execution/types.ts:481](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L481)
+Defined in: [candidate-execution/types.ts:489](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L489)
 
 Executes one prepared request inside an evaluator-owned isolation boundary.
 
@@ -3872,7 +3634,7 @@ no candidate-authored usage or score fields.
 
 > **execute**(`request`, `context`): `Promise`\<[`AgentCandidateProtectedRunCapture`](#agentcandidateprotectedruncapture)\>
 
-Defined in: [candidate-execution/types.ts:482](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L482)
+Defined in: [candidate-execution/types.ts:490](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L490)
 
 ###### Parameters
 
@@ -3890,29 +3652,65 @@ Defined in: [candidate-execution/types.ts:482](https://github.com/tangle-network
 
 `AbortSignal`
 
-Aborted by the runtime at the exact frozen wall-time deadline.
+Aborted by the runtime at the exact frozen wall-time deadline.
+
+###### deadlineAtMs
+
+`number`
+
+Absolute epoch-millisecond deadline owned by the runtime.
+
+###### Returns
+
+`Promise`\<[`AgentCandidateProtectedRunCapture`](#agentcandidateprotectedruncapture)\>
+
+##### stop()
+
+> **stop**(`request`, `context`): `Promise`\<\{ `stopped`: `true`; \}\>
+
+Defined in: [candidate-execution/types.ts:501](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L501)
+
+Kill the exact process/container and drain trace writes. Must be idempotent.
+
+###### Parameters
+
+###### request
+
+[`AgentCandidateExecutorStopRequest`](#agentcandidateexecutorstoprequest)
+
+###### context
+
+###### traceStore
+
+`TraceStore`
+
+###### reason
+
+`"completed"` \| `"failed"` \| `"timeout"`
+
+###### signal
+
+`AbortSignal`
+
+Aborted at the frozen execution deadline or evaluator cleanup deadline.
 
 ###### deadlineAtMs
 
 `number`
 
-Absolute epoch-millisecond deadline owned by the runtime.
+Absolute execution deadline; a later stop acknowledgement cannot produce success.
 
 ###### Returns
 
-`Promise`\<[`AgentCandidateProtectedRunCapture`](#agentcandidateprotectedruncapture)\>
+`Promise`\<\{ `stopped`: `true`; \}\>
 
-##### stopAndCapture()
+##### capture()
 
-> **stopAndCapture**(`request`, `context`): `Promise`\<[`AgentCandidateExecutorFinalCapture`](#agentcandidateexecutorfinalcapture)\>
+> **capture**(`request`, `context`): `Promise`\<[`AgentCandidateExecutorFinalCapture`](#agentcandidateexecutorfinalcapture)\>
 
-Defined in: [candidate-execution/types.ts:499](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L499)
+Defined in: [candidate-execution/types.ts:513](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L513)
 
-Kill any process/container still associated with the request, drain trace
-writes, and capture the final task workspace before teardown.
-The runtime calls this on success, failure, and timeout before model settlement.
-Implementations must be idempotent and concurrency-safe for this exact
-execution/plan pair because a fresh worker may repeat crash recovery.
+Capture immutable final evidence after stop. Must be replayable by a fresh worker.
 
 ###### Parameters
 
@@ -3926,22 +3724,12 @@ execution/plan pair because a fresh worker may repeat crash recovery.
 
 `TraceStore`
 
-###### reason
-
-`"completed"` \| `"failed"` \| `"timeout"`
-
 ###### signal
 
 `AbortSignal`
 
 Aborted at the frozen execution deadline or evaluator cleanup deadline.
 
-###### deadlineAtMs
-
-`number`
-
-Absolute execution deadline; a later stop acknowledgement cannot produce success.
-
 ###### Returns
 
 `Promise`\<[`AgentCandidateExecutorFinalCapture`](#agentcandidateexecutorfinalcapture)\>
@@ -3950,7 +3738,7 @@ Absolute execution deadline; a later stop acknowledgement cannot produce success
 
 ### AgentCandidateExecutorStopRequest
 
-Defined in: [candidate-execution/types.ts:513](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L513)
+Defined in: [candidate-execution/types.ts:524](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L524)
 
 Opaque process identity used for termination without re-exposing launch credentials.
 
@@ -3960,19 +3748,19 @@ Opaque process identity used for termination without re-exposing launch credenti
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:514](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L514)
+Defined in: [candidate-execution/types.ts:525](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L525)
 
 ##### executionPlanDigest
 
 > `readonly` **executionPlanDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:515](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L515)
+Defined in: [candidate-execution/types.ts:526](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L526)
 
 ***
 
 ### AgentCandidateExecutorWorkspaceInput
 
-Defined in: [candidate-execution/types.ts:518](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L518)
+Defined in: [candidate-execution/types.ts:529](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L529)
 
 #### Properties
 
@@ -3980,45 +3768,19 @@ Defined in: [candidate-execution/types.ts:518](https://github.com/tangle-network
 
 > `readonly` **snapshot**: `AgentCandidateWorkspaceSnapshotEvidence`
 
-Defined in: [candidate-execution/types.ts:519](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L519)
+Defined in: [candidate-execution/types.ts:530](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L530)
 
 ##### files
 
 > `readonly` **files**: readonly [`AgentCandidateExecutorWorkspaceFile`](#agentcandidateexecutorworkspacefile)[]
 
-Defined in: [candidate-execution/types.ts:520](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L520)
+Defined in: [candidate-execution/types.ts:531](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L531)
 
 ***
 
 ### AgentCandidateExecutorWorkspaceFile
 
-Defined in: [candidate-execution/types.ts:523](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L523)
-
-#### Properties
-
-##### path
-
-> `readonly` **path**: `string`
-
-Defined in: [candidate-execution/types.ts:524](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L524)
-
-##### mode
-
-> `readonly` **mode**: `420` \| `493`
-
-Defined in: [candidate-execution/types.ts:525](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L525)
-
-##### bytes
-
-> `readonly` **bytes**: `Uint8Array`
-
-Defined in: [candidate-execution/types.ts:526](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L526)
-
-***
-
-### AgentCandidateExecutorProfileFile
-
-Defined in: [candidate-execution/types.ts:529](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L529)
+Defined in: [candidate-execution/types.ts:534](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L534)
 
 #### Properties
 
@@ -4026,19 +3788,19 @@ Defined in: [candidate-execution/types.ts:529](https://github.com/tangle-network
 
 > `readonly` **path**: `string`
 
-Defined in: [candidate-execution/types.ts:530](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L530)
+Defined in: [candidate-execution/types.ts:535](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L535)
 
 ##### mode
 
-> `readonly` **mode**: `420` \| `493`
+> `readonly` **mode**: `number`
 
-Defined in: [candidate-execution/types.ts:531](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L531)
+Defined in: [candidate-execution/types.ts:536](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L536)
 
 ##### bytes
 
 > `readonly` **bytes**: `Uint8Array`
 
-Defined in: [candidate-execution/types.ts:532](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L532)
+Defined in: [candidate-execution/types.ts:537](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L537)
 
 ***
 
@@ -6327,7 +6089,7 @@ Defined in: [improvement/reflective-generator.ts:22](https://github.com/tangle-n
 
 ### RunKnowledgeImprovementJobOptions
 
-Defined in: [knowledge/improvement-job.ts:21](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L21)
+Defined in: [knowledge/improvement-job.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L42)
 
 #### Extends
 
@@ -6339,61 +6101,61 @@ Defined in: [knowledge/improvement-job.ts:21](https://github.com/tangle-network/
 
 > **budget**: [`Budget`](runtime.md#budget-12)
 
-Defined in: [knowledge/improvement-job.ts:23](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L23)
+Defined in: [knowledge/improvement-job.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L44)
 
 ##### readinessCheck?
 
 > `optional` **readinessCheck?**: [`KnowledgeReadinessCheck`](#knowledgereadinesscheck)
 
-Defined in: [knowledge/improvement-job.ts:24](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L24)
+Defined in: [knowledge/improvement-job.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L45)
 
 ##### backend?
 
 > `optional` **backend?**: [`ExecutorConfig`](runtime.md#executorconfig)
 
-Defined in: [knowledge/improvement-job.ts:25](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L25)
+Defined in: [knowledge/improvement-job.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L46)
 
 ##### makeWorkerAgent?
 
 > `optional` **makeWorkerAgent?**: [`MakeWorkerAgent`](runtime.md#makeworkeragent)
 
-Defined in: [knowledge/improvement-job.ts:26](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L26)
+Defined in: [knowledge/improvement-job.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L47)
 
 ##### harness?
 
 > `optional` **harness?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:27](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L27)
+Defined in: [knowledge/improvement-job.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L48)
 
 ##### supervisorModel?
 
 > `optional` **supervisorModel?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:28](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L28)
+Defined in: [knowledge/improvement-job.ts:49](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L49)
 
 ##### supervisorSystemPrompt?
 
 > `optional` **supervisorSystemPrompt?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:29](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L29)
+Defined in: [knowledge/improvement-job.ts:50](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L50)
 
 ##### superviseOptions?
 
 > `optional` **superviseOptions?**: `Partial`\<`Omit`\<[`SuperviseOptions`](runtime.md#superviseoptions), `"backend"` \| `"budget"` \| `"makeWorkerAgent"` \| `"deliverable"` \| `"allowedModels"`\>\>
 
-Defined in: [knowledge/improvement-job.ts:30](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L30)
+Defined in: [knowledge/improvement-job.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L51)
 
 ##### allowedModels?
 
 > `optional` **allowedModels?**: readonly `string`[]
 
-Defined in: [knowledge/improvement-job.ts:36](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L36)
+Defined in: [knowledge/improvement-job.ts:57](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L57)
 
 ##### runSupervised?
 
 > `optional` **runSupervised?**: (`profile`, `task`, `opts`) => `Promise`\<[`SupervisedResult`](runtime.md#supervisedresult)\<`unknown`\>\>
 
-Defined in: [knowledge/improvement-job.ts:37](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L37)
+Defined in: [knowledge/improvement-job.ts:58](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L58)
 
 ###### Parameters
 
@@ -6413,11 +6175,23 @@ Defined in: [knowledge/improvement-job.ts:37](https://github.com/tangle-network/
 
 `Promise`\<[`SupervisedResult`](runtime.md#supervisedresult)\<`unknown`\>\>
 
+##### candidateArtifacts?
+
+> `optional` **candidateArtifacts?**: [`AgentCandidateOutputArtifactPort`](#agentcandidateoutputartifactport)
+
+Defined in: [knowledge/improvement-job.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L63)
+
+##### approval?
+
+> `optional` **approval?**: [`ApprovedKnowledgeImprovementCandidate`](#approvedknowledgeimprovementcandidate)
+
+Defined in: [knowledge/improvement-job.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L64)
+
 ##### onMeasurement?
 
 > `optional` **onMeasurement?**: (`measurement`) => `void` \| `Promise`\<`void`\>
 
-Defined in: [knowledge/improvement-job.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L42)
+Defined in: [knowledge/improvement-job.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L65)
 
 ###### Parameters
 
@@ -6431,9 +6205,49 @@ Defined in: [knowledge/improvement-job.ts:42](https://github.com/tangle-network/
 
 ***
 
+### ApprovedKnowledgeImprovementCandidate
+
+Defined in: [knowledge/improvement-job.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L68)
+
+#### Properties
+
+##### proposal
+
+> **proposal**: `AgentImprovementProposal`
+
+Defined in: [knowledge/improvement-job.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L69)
+
+##### review
+
+> **review**: `AgentImprovementReview`
+
+Defined in: [knowledge/improvement-job.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L70)
+
+##### authorizeReview
+
+> **authorizeReview**: (`review`, `proposal`) => `boolean` \| `Promise`\<`boolean`\>
+
+Defined in: [knowledge/improvement-job.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L71)
+
+###### Parameters
+
+###### review
+
+`AgentImprovementReview`
+
+###### proposal
+
+`AgentImprovementProposal`
+
+###### Returns
+
+`boolean` \| `Promise`\<`boolean`\>
+
+***
+
 ### KnowledgeImprovementJobMeasurement
 
-Defined in: [knowledge/improvement-job.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L45)
+Defined in: [knowledge/improvement-job.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L77)
 
 #### Properties
 
@@ -6441,37 +6255,37 @@ Defined in: [knowledge/improvement-job.ts:45](https://github.com/tangle-network/
 
 > **startedAt**: `string`
 
-Defined in: [knowledge/improvement-job.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L46)
+Defined in: [knowledge/improvement-job.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L78)
 
 ##### finishedAt
 
 > **finishedAt**: `string`
 
-Defined in: [knowledge/improvement-job.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L47)
+Defined in: [knowledge/improvement-job.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L79)
 
 ##### durationMs
 
 > **durationMs**: `number`
 
-Defined in: [knowledge/improvement-job.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L48)
+Defined in: [knowledge/improvement-job.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L80)
 
 ##### updateCalls
 
 > **updateCalls**: `number`
 
-Defined in: [knowledge/improvement-job.ts:49](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L49)
+Defined in: [knowledge/improvement-job.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L81)
 
 ##### updateDurationMs
 
 > **updateDurationMs**: `number`
 
-Defined in: [knowledge/improvement-job.ts:50](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L50)
+Defined in: [knowledge/improvement-job.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L82)
 
 ##### supervisedSpent
 
 > **supervisedSpent**: `object`
 
-Defined in: [knowledge/improvement-job.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L51)
+Defined in: [knowledge/improvement-job.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L83)
 
 ###### iterations
 
@@ -6497,7 +6311,7 @@ Defined in: [knowledge/improvement-job.ts:51](https://github.com/tangle-network/
 
 ### KnowledgeImprovementJobResult
 
-Defined in: [knowledge/improvement-job.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L60)
+Defined in: [knowledge/improvement-job.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L92)
 
 #### Properties
 
@@ -6505,31 +6319,37 @@ Defined in: [knowledge/improvement-job.ts:60](https://github.com/tangle-network/
 
 > **improvement**: `KnowledgeImprovementResult`
 
-Defined in: [knowledge/improvement-job.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L61)
+Defined in: [knowledge/improvement-job.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L93)
+
+##### candidateKnowledge?
+
+> `optional` **candidateKnowledge?**: `AgentCandidateKnowledge`
+
+Defined in: [knowledge/improvement-job.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L94)
 
 ##### measurement
 
 > **measurement**: [`KnowledgeImprovementJobMeasurement`](#knowledgeimprovementjobmeasurement)
 
-Defined in: [knowledge/improvement-job.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L62)
+Defined in: [knowledge/improvement-job.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L95)
 
 ##### promoted
 
 > **promoted**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L63)
+Defined in: [knowledge/improvement-job.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L96)
 
 ##### blocked
 
 > **blocked**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L64)
+Defined in: [knowledge/improvement-job.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L97)
 
 ***
 
 ### AgentKnowledgeReadinessCheckOptions
 
-Defined in: [knowledge/improvement-job.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L67)
+Defined in: [knowledge/improvement-job.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L100)
 
 #### Properties
 
@@ -6537,43 +6357,43 @@ Defined in: [knowledge/improvement-job.ts:67](https://github.com/tangle-network/
 
 > **goal**: `string`
 
-Defined in: [knowledge/improvement-job.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L68)
+Defined in: [knowledge/improvement-job.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L101)
 
 ##### readinessSpecs?
 
 > `optional` **readinessSpecs?**: readonly `KnowledgeReadinessSpec`[]
 
-Defined in: [knowledge/improvement-job.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L69)
+Defined in: [knowledge/improvement-job.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L102)
 
 ##### readinessTaskId?
 
 > `optional` **readinessTaskId?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L70)
+Defined in: [knowledge/improvement-job.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L103)
 
 ##### readiness?
 
 > `optional` **readiness?**: `Omit`\<`BuildEvalKnowledgeBundleOptions`, `"taskId"` \| `"index"` \| `"specs"`\>
 
-Defined in: [knowledge/improvement-job.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L71)
+Defined in: [knowledge/improvement-job.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L104)
 
 ##### strict?
 
 > `optional` **strict?**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L72)
+Defined in: [knowledge/improvement-job.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L105)
 
 ##### kbQuality?
 
 > `optional` **kbQuality?**: `KnowledgeBaseQualityOptions`
 
-Defined in: [knowledge/improvement-job.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L73)
+Defined in: [knowledge/improvement-job.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L106)
 
 ***
 
 ### KnowledgeReadinessCheckInput
 
-Defined in: [knowledge/supervised-update.ts:22](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L22)
+Defined in: [knowledge/supervised-update.ts:23](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L23)
 
 #### Properties
 
@@ -6581,37 +6401,37 @@ Defined in: [knowledge/supervised-update.ts:22](https://github.com/tangle-networ
 
 > **root**: `string`
 
-Defined in: [knowledge/supervised-update.ts:23](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L23)
+Defined in: [knowledge/supervised-update.ts:24](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L24)
 
 ##### goal
 
 > **goal**: `string`
 
-Defined in: [knowledge/supervised-update.ts:24](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L24)
+Defined in: [knowledge/supervised-update.ts:25](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L25)
 
 ##### readinessSpecs?
 
 > `optional` **readinessSpecs?**: readonly `unknown`[]
 
-Defined in: [knowledge/supervised-update.ts:25](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L25)
+Defined in: [knowledge/supervised-update.ts:26](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L26)
 
 ##### readinessTaskId?
 
 > `optional` **readinessTaskId?**: `string`
 
-Defined in: [knowledge/supervised-update.ts:26](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L26)
+Defined in: [knowledge/supervised-update.ts:27](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L27)
 
 ##### readiness?
 
 > `optional` **readiness?**: `unknown`
 
-Defined in: [knowledge/supervised-update.ts:27](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L27)
+Defined in: [knowledge/supervised-update.ts:28](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L28)
 
 ***
 
 ### SupervisedKnowledgeUpdateInput
 
-Defined in: [knowledge/supervised-update.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L42)
+Defined in: [knowledge/supervised-update.ts:43](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L43)
 
 #### Properties
 
@@ -6619,37 +6439,37 @@ Defined in: [knowledge/supervised-update.ts:42](https://github.com/tangle-networ
 
 > `optional` **goal?**: `string`
 
-Defined in: [knowledge/supervised-update.ts:43](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L43)
+Defined in: [knowledge/supervised-update.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L44)
 
 ##### root?
 
 > `optional` **root?**: `string`
 
-Defined in: [knowledge/supervised-update.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L44)
+Defined in: [knowledge/supervised-update.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L45)
 
 ##### candidateRoot?
 
 > `optional` **candidateRoot?**: `string`
 
-Defined in: [knowledge/supervised-update.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L45)
+Defined in: [knowledge/supervised-update.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L46)
 
 ##### findings?
 
 > `optional` **findings?**: readonly `unknown`[]
 
-Defined in: [knowledge/supervised-update.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L46)
+Defined in: [knowledge/supervised-update.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L47)
 
 ##### metadata?
 
 > `optional` **metadata?**: `Record`\<`string`, `unknown`\>
 
-Defined in: [knowledge/supervised-update.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L47)
+Defined in: [knowledge/supervised-update.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L48)
 
 ***
 
 ### SupervisedKnowledgeUpdateResult
 
-Defined in: [knowledge/supervised-update.ts:50](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L50)
+Defined in: [knowledge/supervised-update.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L51)
 
 #### Properties
 
@@ -6657,31 +6477,31 @@ Defined in: [knowledge/supervised-update.ts:50](https://github.com/tangle-networ
 
 > **applied**: `boolean`
 
-Defined in: [knowledge/supervised-update.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L51)
+Defined in: [knowledge/supervised-update.ts:52](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L52)
 
 ##### summary
 
 > **summary**: `string`
 
-Defined in: [knowledge/supervised-update.ts:52](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L52)
+Defined in: [knowledge/supervised-update.ts:53](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L53)
 
 ##### supervised
 
 > **supervised**: [`SupervisedResult`](runtime.md#supervisedresult)\<`unknown`\>
 
-Defined in: [knowledge/supervised-update.ts:53](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L53)
+Defined in: [knowledge/supervised-update.ts:54](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L54)
 
 ##### metadata
 
-> **metadata**: `Record`\<`string`, `unknown`\>
+> **metadata**: `NonNullable`\<`RagKnowledgeUpdateResult`\[`"metadata"`\]\>
 
-Defined in: [knowledge/supervised-update.ts:54](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L54)
+Defined in: [knowledge/supervised-update.ts:55](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L55)
 
 ***
 
 ### SupervisedKnowledgeUpdateOptions
 
-Defined in: [knowledge/supervised-update.ts:57](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L57)
+Defined in: [knowledge/supervised-update.ts:58](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L58)
 
 #### Properties
 
@@ -6689,103 +6509,103 @@ Defined in: [knowledge/supervised-update.ts:57](https://github.com/tangle-networ
 
 > **root**: `string`
 
-Defined in: [knowledge/supervised-update.ts:58](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L58)
+Defined in: [knowledge/supervised-update.ts:59](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L59)
 
 ##### goal
 
 > **goal**: `string`
 
-Defined in: [knowledge/supervised-update.ts:59](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L59)
+Defined in: [knowledge/supervised-update.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L60)
 
 ##### readiness
 
 > **readiness**: [`KnowledgeReadinessCheck`](#knowledgereadinesscheck)
 
-Defined in: [knowledge/supervised-update.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L60)
+Defined in: [knowledge/supervised-update.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L61)
 
 ##### readinessSpecs?
 
 > `optional` **readinessSpecs?**: readonly `unknown`[]
 
-Defined in: [knowledge/supervised-update.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L61)
+Defined in: [knowledge/supervised-update.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L62)
 
 ##### readinessTaskId?
 
 > `optional` **readinessTaskId?**: `string`
 
-Defined in: [knowledge/supervised-update.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L62)
+Defined in: [knowledge/supervised-update.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L63)
 
 ##### readinessOptions?
 
 > `optional` **readinessOptions?**: `unknown`
 
-Defined in: [knowledge/supervised-update.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L63)
+Defined in: [knowledge/supervised-update.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L64)
 
 ##### findings?
 
 > `optional` **findings?**: readonly `unknown`[]
 
-Defined in: [knowledge/supervised-update.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L64)
+Defined in: [knowledge/supervised-update.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L65)
 
 ##### metadata?
 
 > `optional` **metadata?**: `Record`\<`string`, `unknown`\>
 
-Defined in: [knowledge/supervised-update.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L65)
+Defined in: [knowledge/supervised-update.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L66)
 
 ##### budget
 
 > **budget**: [`Budget`](runtime.md#budget-12)
 
-Defined in: [knowledge/supervised-update.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L66)
+Defined in: [knowledge/supervised-update.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L67)
 
 ##### backend?
 
 > `optional` **backend?**: [`ExecutorConfig`](runtime.md#executorconfig)
 
-Defined in: [knowledge/supervised-update.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L67)
+Defined in: [knowledge/supervised-update.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L68)
 
 ##### makeWorkerAgent?
 
 > `optional` **makeWorkerAgent?**: [`MakeWorkerAgent`](runtime.md#makeworkeragent)
 
-Defined in: [knowledge/supervised-update.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L68)
+Defined in: [knowledge/supervised-update.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L69)
 
 ##### harness?
 
 > `optional` **harness?**: `string`
 
-Defined in: [knowledge/supervised-update.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L69)
+Defined in: [knowledge/supervised-update.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L70)
 
 ##### supervisorModel?
 
 > `optional` **supervisorModel?**: `string`
 
-Defined in: [knowledge/supervised-update.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L70)
+Defined in: [knowledge/supervised-update.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L71)
 
 ##### supervisorSystemPrompt?
 
 > `optional` **supervisorSystemPrompt?**: `string`
 
-Defined in: [knowledge/supervised-update.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L71)
+Defined in: [knowledge/supervised-update.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L72)
 
 ##### superviseOptions?
 
 > `optional` **superviseOptions?**: `Partial`\<`Omit`\<[`SuperviseOptions`](runtime.md#superviseoptions), `"backend"` \| `"budget"` \| `"makeWorkerAgent"` \| `"deliverable"` \| `"allowedModels"`\>\>
 
-Defined in: [knowledge/supervised-update.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L72)
+Defined in: [knowledge/supervised-update.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L73)
 
 ##### allowedModels?
 
 > `optional` **allowedModels?**: readonly `string`[]
 
-Defined in: [knowledge/supervised-update.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L78)
+Defined in: [knowledge/supervised-update.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L79)
 
 ##### runSupervised?
 
 > `optional` **runSupervised?**: (`profile`, `task`, `opts`) => `Promise`\<[`SupervisedResult`](runtime.md#supervisedresult)\<`unknown`\>\>
 
-Defined in: [knowledge/supervised-update.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L79)
+Defined in: [knowledge/supervised-update.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L80)
 
 ###### Parameters
 
@@ -10422,7 +10242,7 @@ Exact candidate wire shape before the runtime computes its canonical digest.
 
 > **AgentCandidateExecutionFailureClass** = `"pre-model-infrastructure"` \| `"execution"` \| `"post-model-infrastructure"` \| `"unknown"`
 
-Defined in: [candidate-execution/claim.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L78)
+Defined in: [candidate-execution/claim.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L83)
 
 Only the first class is retryable, and only when the closed model ledger has zero calls.
 
@@ -10430,9 +10250,9 @@ Only the first class is retryable, and only when the closed model ledger has zer
 
 ### AgentCandidateExecutionTerminalResult
 
-> **AgentCandidateExecutionTerminalResult** = \{ `schemaVersion`: `1`; `status`: `"succeeded"`; `usage`: [`AgentCandidateExecutionUsage`](#agentcandidateexecutionusage); `modelSettlement`: `AgentCandidateArtifactRef`; `taskOutcome`: `AgentCandidateArtifactRef`; `benchmarkResult`: `AgentCandidateArtifactRef`; `runReceipt`: `AgentCandidateArtifactRef`; \} \| \{ `schemaVersion`: `1`; `status`: `"failed"`; `failureClass`: [`AgentCandidateExecutionFailureClass`](#agentcandidateexecutionfailureclass); `usage`: [`AgentCandidateExecutionUsage`](#agentcandidateexecutionusage); `modelSettlement`: `AgentCandidateArtifactRef`; `failureEvidence?`: `AgentCandidateArtifactRef`; \}
+> **AgentCandidateExecutionTerminalResult** = \{ `schemaVersion`: `1`; `status`: `"succeeded"`; `usage`: `AgentCandidateFixedSpend`; `modelSettlement`: `AgentCandidateArtifactRef`; `taskOutcome`: `AgentCandidateArtifactRef`; `benchmarkResult`: `AgentCandidateArtifactRef`; `runReceipt`: `AgentCandidateArtifactRef`; \} \| \{ `schemaVersion`: `1`; `status`: `"failed"`; `failureClass`: [`AgentCandidateExecutionFailureClass`](#agentcandidateexecutionfailureclass); `usage`: `AgentCandidateFixedSpend`; `modelSettlement`: `AgentCandidateArtifactRef`; `failureEvidence?`: `AgentCandidateArtifactRef`; \}
 
-Defined in: [candidate-execution/claim.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L95)
+Defined in: [candidate-execution/claim.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L90)
 
 Evaluator-owned terminal facts staged durably before the terminal CAS.
 
@@ -10442,7 +10262,7 @@ Evaluator-owned terminal facts staged durably before the terminal CAS.
 
 > **AgentCandidateExecutionTerminalRecord** = [`AgentCandidateExecutionTerminalResult`](#agentcandidateexecutionterminalresult) & `object`
 
-Defined in: [candidate-execution/claim.ts:115](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L115)
+Defined in: [candidate-execution/claim.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L110)
 
 Durable terminal record for one acquired execution attempt.
 
@@ -10464,6 +10284,10 @@ Durable terminal record for one acquired execution attempt.
 
 > `readonly` **executionPlanDigest**: `Sha256Digest`
 
+##### preparationEvidence
+
+> `readonly` **preparationEvidence**: [`AgentCandidateExecutionClaim`](#agentcandidateexecutionclaim)\[`"preparationEvidence"`\]
+
 ##### terminalDigest
 
 > `readonly` **terminalDigest**: `Sha256Digest`
@@ -10476,7 +10300,7 @@ RFC 8785 SHA-256 of this record with `terminalDigest` omitted.
 
 > **AgentCandidateExecutionPhase** = `"claimed"` \| `"candidate-may-run"`
 
-Defined in: [candidate-execution/claim.ts:125](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L125)
+Defined in: [candidate-execution/claim.ts:121](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L121)
 
 Monotonic durable phase: the second value means candidate code could have started.
 
@@ -10486,7 +10310,7 @@ Monotonic durable phase: the second value means candidate code could have starte
 
 > **AgentCandidateExecutionClaimResult** = \{ `acquired`: `true`; `claim`: [`AgentCandidateExecutionClaim`](#agentcandidateexecutionclaim); `lease`: [`AgentCandidateExecutionLease`](#agentcandidateexecutionlease); \} \| \{ `acquired`: `false`; `reason`: `"already-claimed"`; `claim`: [`AgentCandidateExecutionClaim`](#agentcandidateexecutionclaim); `exactReplay`: `boolean`; \} \| \{ `acquired`: `false`; `reason`: `"retry-not-eligible"`; `claim`: [`AgentCandidateExecutionClaim`](#agentcandidateexecutionclaim); `detail`: [`AgentCandidateRetryRejection`](#agentcandidateretryrejection); \}
 
-Defined in: [candidate-execution/claim.ts:165](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L165)
+Defined in: [candidate-execution/claim.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L161)
 
 Result of atomically claiming one execution attempt.
 
@@ -10534,7 +10358,7 @@ True only when every signed claim field matches the durable winner.
 
 > **AgentCandidateExecutionFinishResult** = \{ `finished`: `true`; `terminal`: [`AgentCandidateExecutionTerminalRecord`](#agentcandidateexecutionterminalrecord); \} \| \{ `finished`: `false`; `terminal`: [`AgentCandidateExecutionTerminalRecord`](#agentcandidateexecutionterminalrecord); `exactReplay`: `boolean`; \}
 
-Defined in: [candidate-execution/claim.ts:187](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L187)
+Defined in: [candidate-execution/claim.ts:183](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L183)
 
 Result of atomically recording an attempt's terminal facts.
 
@@ -10570,7 +10394,7 @@ True when a repeated finish supplied the same terminal digest.
 
 > **AgentCandidateExecutionStageResult** = \{ `staged`: `true`; `terminal`: [`AgentCandidateExecutionTerminalRecord`](#agentcandidateexecutionterminalrecord); \} \| \{ `staged`: `false`; `terminal`: [`AgentCandidateExecutionTerminalRecord`](#agentcandidateexecutionterminalrecord); `exactReplay`: `boolean`; \}
 
-Defined in: [candidate-execution/claim.ts:200](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L200)
+Defined in: [candidate-execution/claim.ts:196](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L196)
 
 Result of durably staging the one immutable terminal outbox entry.
 
@@ -10580,7 +10404,7 @@ Result of durably staging the one immutable terminal outbox entry.
 
 > **AgentCandidateExecutionPhaseResult** = \{ `marked`: `true`; `phase`: `"candidate-may-run"`; \} \| \{ `marked`: `false`; `phase`: `"candidate-may-run"`; \}
 
-Defined in: [candidate-execution/claim.ts:212](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L212)
+Defined in: [candidate-execution/claim.ts:208](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L208)
 
 Result of crossing the irreversible candidate-may-run boundary.
 
@@ -10590,7 +10414,7 @@ Result of crossing the irreversible candidate-may-run boundary.
 
 > **AgentCandidateRetryRejection** = `"prior-attempt-missing"` \| `"prior-attempt-running"` \| `"prior-attempt-succeeded"` \| `"prior-attempt-spent-model-calls"` \| `"prior-attempt-not-pre-model-infrastructure"` \| `"retry-lineage-mismatch"`
 
-Defined in: [candidate-execution/claim.ts:216](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L216)
+Defined in: [candidate-execution/claim.ts:212](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L212)
 
 ***
 
@@ -10630,9 +10454,9 @@ Secret-free response from the service's reservation endpoint.
 
 ### AgentCandidateOutputPurpose
 
-> **AgentCandidateOutputPurpose** = `"candidate-workspace-manifest"` \| `"candidate-workspace-archive"` \| `"task-manifest"` \| `"task-archive"` \| `"task-patch"` \| `"task-outcome"` \| `"memory-after-manifest"` \| `"memory-after-archive"` \| `"grader-evidence"` \| `"benchmark-result"` \| `"model-settlement"` \| `"trace"` \| `"run-receipt"` \| `"failure-evidence"`
+> **AgentCandidateOutputPurpose** = `"execution-plan"` \| `"materialization-receipt"` \| `"candidate-workspace-manifest"` \| `"candidate-workspace-archive"` \| `"task-manifest"` \| `"task-archive"` \| `"task-patch"` \| `"task-output"` \| `"task-outcome"` \| `"memory-after-manifest"` \| `"memory-after-archive"` \| `"grader-evidence"` \| `"benchmark-result"` \| `"model-settlement"` \| `"trace"` \| `"run-receipt"` \| `"executor-capture"` \| `"knowledge-retrieval-config"` \| `"knowledge-evaluation"` \| `"failure-evidence"`
 
-Defined in: [candidate-execution/types.ts:38](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L38)
+Defined in: [candidate-execution/types.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L44)
 
 ***
 
@@ -10640,29 +10464,99 @@ Defined in: [candidate-execution/types.ts:38](https://github.com/tangle-network/
 
 > **AgentCandidateModelLimits** = `Pick`\<`AgentCandidateExecutionLimits`, `"maxModelCalls"` \| `"maxInputTokens"` \| `"maxOutputTokens"` \| `"maxCostUsd"`\>
 
-Defined in: [candidate-execution/types.ts:151](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L151)
+Defined in: [candidate-execution/types.ts:163](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L163)
 
 Limits mechanically enforced by the evaluator-owned model gateway.
 
 ***
 
+### AgentCandidateExecutorTaskOutcomeCapture
+
+> **AgentCandidateExecutorTaskOutcomeCapture** = \{ `kind`: `"workspace"`; `resultTree`: `string`; `afterState`: `AgentCandidateWorkspaceManifestMaterial`; `archive`: `Uint8Array`; `gitDiff`: `Uint8Array`; \} \| \{ `kind`: `"output"`; `bytes`: `Uint8Array`; \}
+
+Defined in: [candidate-execution/types.ts:359](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L359)
+
+Raw evaluator capture made only after the candidate process is dead.
+
+#### Union Members
+
+##### Type Literal
+
+\{ `kind`: `"workspace"`; `resultTree`: `string`; `afterState`: `AgentCandidateWorkspaceManifestMaterial`; `archive`: `Uint8Array`; `gitDiff`: `Uint8Array`; \}
+
+###### kind
+
+> `readonly` **kind**: `"workspace"`
+
+###### resultTree
+
+> `readonly` **resultTree**: `string`
+
+Claimed final tree. The runtime recomputes it independently from `gitDiff`.
+
+###### afterState
+
+> `readonly` **afterState**: `AgentCandidateWorkspaceManifestMaterial`
+
+Complete evaluator-captured workspace description after candidate execution.
+
+###### archive
+
+> `readonly` **archive**: `Uint8Array`
+
+Reproducible workspace archive corresponding to `afterState`.
+
+###### gitDiff
+
+> `readonly` **gitDiff**: `Uint8Array`
+
+Exact binary patch from the signed task base to `afterState`.
+
+***
+
+##### Type Literal
+
+\{ `kind`: `"output"`; `bytes`: `Uint8Array`; \}
+
+###### kind
+
+> `readonly` **kind**: `"output"`
+
+###### bytes
+
+> `readonly` **bytes**: `Uint8Array`
+
+Exact evaluator-captured final output bytes.
+
+***
+
+### VerifiedAgentCandidateTaskOutcome
+
+> **VerifiedAgentCandidateTaskOutcome** = \{ `kind`: `"workspace"`; `evidence`: `PersistedTaskOutcomeEvidence`\<`"workspace"`\>; `patch`: `Uint8Array`; `[verifiedTaskOutcomeBrand]`: `true`; \} \| \{ `kind`: `"output"`; `evidence`: `PersistedTaskOutcomeEvidence`\<`"output"`\>; `spec`: `AgentCandidateTaskOutputSpec`; `bytes`: `Uint8Array`; `[verifiedTaskOutcomeBrand]`: `true`; \}
+
+Defined in: [candidate-execution/types.ts:402](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L402)
+
+Branded task outcome that has survived independent evaluator verification.
+
+***
+
 ### AgentCandidateRunFinalization
 
-> **AgentCandidateRunFinalization** = \{ `succeeded`: `true`; `receipt`: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateRunReceiptV2`\>; `artifacts`: \{ `modelSettlement`: `AgentCandidateArtifactRef`; `taskOutcome`: `AgentCandidateArtifactRef`; `benchmarkResult`: `AgentCandidateArtifactRef`; `runReceipt`: `AgentCandidateArtifactRef`; \}; \} \| \{ `succeeded`: `false`; `reason`: `string`; `partial`: \{ `executionId`: `string`; `bundleDigest`: `Sha256Digest`; `executionPlanDigest`: `Sha256Digest`; `materializationReceiptDigest`: `Sha256Digest`; `termination?`: `AgentCandidateTermination`; \}; `usage`: `AgentCandidateSpend` \| `null`; \}
+> **AgentCandidateRunFinalization** = \{ `succeeded`: `true`; `receipt`: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateRunReceipt`\>; `artifacts`: \{ `modelSettlement`: `AgentCandidateArtifactRef`; `executorCapture`: `AgentCandidateArtifactRef`; `taskOutcome`: `AgentCandidateArtifactRef`; `benchmarkResult`: `AgentCandidateArtifactRef`; `runReceipt`: `AgentCandidateArtifactRef`; \}; \} \| \{ `succeeded`: `false`; `reason`: `string`; `partial`: \{ `executionId`: `string`; `bundleDigest`: `Sha256Digest`; `executionPlanDigest`: `Sha256Digest`; `materializationReceiptDigest`: `Sha256Digest`; `termination?`: `AgentCandidateTermination`; \}; `usage`: `AgentCandidateFixedSpend` \| `null`; \}
 
-Defined in: [candidate-execution/types.ts:535](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L535)
+Defined in: [candidate-execution/types.ts:540](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L540)
 
 #### Union Members
 
 ##### Type Literal
 
-\{ `succeeded`: `true`; `receipt`: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateRunReceiptV2`\>; `artifacts`: \{ `modelSettlement`: `AgentCandidateArtifactRef`; `taskOutcome`: `AgentCandidateArtifactRef`; `benchmarkResult`: `AgentCandidateArtifactRef`; `runReceipt`: `AgentCandidateArtifactRef`; \}; \}
+\{ `succeeded`: `true`; `receipt`: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateRunReceipt`\>; `artifacts`: \{ `modelSettlement`: `AgentCandidateArtifactRef`; `executorCapture`: `AgentCandidateArtifactRef`; `taskOutcome`: `AgentCandidateArtifactRef`; `benchmarkResult`: `AgentCandidateArtifactRef`; `runReceipt`: `AgentCandidateArtifactRef`; \}; \}
 
 ***
 
 ##### Type Literal
 
-\{ `succeeded`: `false`; `reason`: `string`; `partial`: \{ `executionId`: `string`; `bundleDigest`: `Sha256Digest`; `executionPlanDigest`: `Sha256Digest`; `materializationReceiptDigest`: `Sha256Digest`; `termination?`: `AgentCandidateTermination`; \}; `usage`: `AgentCandidateSpend` \| `null`; \}
+\{ `succeeded`: `false`; `reason`: `string`; `partial`: \{ `executionId`: `string`; `bundleDigest`: `Sha256Digest`; `executionPlanDigest`: `Sha256Digest`; `materializationReceiptDigest`: `Sha256Digest`; `termination?`: `AgentCandidateTermination`; \}; `usage`: `AgentCandidateFixedSpend` \| `null`; \}
 
 ###### succeeded
 
@@ -10698,7 +10592,7 @@ Defined in: [candidate-execution/types.ts:535](https://github.com/tangle-network
 
 ###### usage
 
-> **usage**: `AgentCandidateSpend` \| `null`
+> **usage**: `AgentCandidateFixedSpend` \| `null`
 
 Independent evaluator-gateway usage, even when execution or trace capture failed.
 
@@ -10998,7 +10892,7 @@ Defined in: [improvement/profile-diff-proposer.ts:26](https://github.com/tangle-
 
 > **KnowledgeReadinessCheckResult** = `boolean` \| \{ `ready`: `boolean`; `summary?`: `string`; `metadata?`: `Record`\<`string`, `unknown`\>; \}
 
-Defined in: [knowledge/supervised-update.ts:30](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L30)
+Defined in: [knowledge/supervised-update.ts:31](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L31)
 
 ***
 
@@ -11006,7 +10900,7 @@ Defined in: [knowledge/supervised-update.ts:30](https://github.com/tangle-networ
 
 > **KnowledgeReadinessCheck** = (`input`) => `Promise`\<[`KnowledgeReadinessCheckResult`](#knowledgereadinesscheckresult)\> \| [`KnowledgeReadinessCheckResult`](#knowledgereadinesscheckresult)
 
-Defined in: [knowledge/supervised-update.ts:38](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L38)
+Defined in: [knowledge/supervised-update.ts:39](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L39)
 
 #### Parameters
 
@@ -11024,7 +10918,7 @@ Defined in: [knowledge/supervised-update.ts:38](https://github.com/tangle-networ
 
 > **SupervisedKnowledgeUpdater** = (`input`) => `Promise`\<[`SupervisedKnowledgeUpdateResult`](#supervisedknowledgeupdateresult)\>
 
-Defined in: [knowledge/supervised-update.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L86)
+Defined in: [knowledge/supervised-update.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L87)
 
 #### Parameters
 
@@ -11557,7 +11451,7 @@ MUST map this to `RunRecord.error` rather than recording silent
 
 > `const` **CANDIDATE\_TRACE\_TAGS**: `object`
 
-Defined in: [candidate-execution/types.ts:561](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L561)
+Defined in: [candidate-execution/types.ts:567](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L567)
 
 Protected trace tags that bind a run to one prepared candidate execution.
 
@@ -11585,7 +11479,7 @@ Protected trace tags that bind a run to one prepared candidate execution.
 
 > `const` **CANDIDATE\_TRACE\_ENV**: `object`
 
-Defined in: [candidate-execution/types.ts:569](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L569)
+Defined in: [candidate-execution/types.ts:575](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L575)
 
 Environment keys used to propagate immutable candidate trace identity.
 
@@ -11613,6 +11507,16 @@ Environment keys used to propagate immutable candidate trace identity.
 
 ***
 
+### AGENT\_CANDIDATE\_EXECUTION\_SUPPORT
+
+> `const` **AGENT\_CANDIDATE\_EXECUTION\_SUPPORT**: `Readonly`\<\{ `outcomes`: readonly \[`"workspace"`, `"output"`\]; `code`: readonly \[`"disabled"`, `"no-op"`, `"git-patch"`\]; `memory`: readonly \[`"disabled"`, `"isolated"`\]; `knowledge`: `false`; `profile`: `Readonly`\<\{ `mcpTransports`: readonly \[`"stdio"`\]; `remoteMcp`: `false`; `tools`: `false`; `permissions`: `false`; `modes`: `false`; `confidential`: `false`; \}\>; \}\>
+
+Defined in: [candidate-execution/verify.ts:38](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/verify.ts#L38)
+
+Surfaces admitted by Runtime's verifier before an environment adapter is selected.
+
+***
+
 ### defaultIsRetryable
 
 > `const` **defaultIsRetryable**: [`RetryableErrorPredicate`](#retryableerrorpredicate)
@@ -11687,7 +11591,7 @@ Hard cap on chained gateway hops; refused beyond this. Default keeps recursion b
 
 > `const` **RESEARCH\_SUPERVISOR\_SYSTEM\_PROMPT**: `string`
 
-Defined in: [knowledge/supervised-update.ts:9](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L9)
+Defined in: [knowledge/supervised-update.ts:10](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L10)
 
 Standing prompt for a supervisor that grows a shared knowledge base through spawned researchers.
 
@@ -11934,7 +11838,7 @@ Maximum completion tokens, sent as OpenAI-compatible `max_tokens`. Omit for prov
 
 ### buildAgentCandidateBundle()
 
-> **buildAgentCandidateBundle**(`input`): `AgentCandidateBundleV1`
+> **buildAgentCandidateBundle**(`input`): `AgentCandidateBundle`
 
 Defined in: [candidate-execution/builder.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/builder.ts#L76)
 
@@ -11952,13 +11856,13 @@ to re-read external knowledge, memory, repository, and workspace artifacts.
 
 #### Returns
 
-`AgentCandidateBundleV1`
+`AgentCandidateBundle`
 
 ***
 
 ### sealAgentCandidateBundle()
 
-> **sealAgentCandidateBundle**(`input`): `AgentCandidateBundleV1`
+> **sealAgentCandidateBundle**(`input`): `AgentCandidateBundle`
 
 Defined in: [candidate-execution/bundle.ts:10](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/bundle.ts#L10)
 
@@ -11972,13 +11876,13 @@ Validate and content-address a candidate bundle before it crosses an approval bo
 
 #### Returns
 
-`AgentCandidateBundleV1`
+`AgentCandidateBundle`
 
 ***
 
 ### candidateExecutionClaim()
 
-> **candidateExecutionClaim**(`prepared`): [`AgentCandidateExecutionClaim`](#agentcandidateexecutionclaim)
+> **candidateExecutionClaim**(`prepared`, `preparationEvidence`): [`AgentCandidateExecutionClaim`](#agentcandidateexecutionclaim)
 
 Defined in: [candidate-execution/claim-plan.ts:10](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim-plan.ts#L10)
 
@@ -11990,6 +11894,16 @@ Extract the complete durable claim from a prepared execution.
 
 [`PreparedAgentCandidateExecution`](#preparedagentcandidateexecution)
 
+##### preparationEvidence
+
+###### executionPlan
+
+`AgentCandidateArtifactRef`
+
+###### materializationReceipt
+
+`AgentCandidateArtifactRef`
+
 #### Returns
 
 [`AgentCandidateExecutionClaim`](#agentcandidateexecutionclaim)
@@ -12024,7 +11938,7 @@ Revoke reservations held by a prepared candidate that will not be executed.
 
 > **executePreparedAgentCandidate**(`prepared`, `options`): `Promise`\<[`AgentCandidateRunFinalization`](#agentcandidaterunfinalization)\>
 
-Defined in: [candidate-execution/execute.ts:75](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L75)
+Defined in: [candidate-execution/execute.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L77)
 
 Executes and finalizes one durably claimed candidate without exposing an unproven result.
 
@@ -12086,7 +12000,7 @@ Persist evaluator evidence, read it back, and bind the returned locator to the e
 
 > **prepareAgentCandidateExecution**(`candidate`, `task`, `ports`, `options?`): `Promise`\<[`PreparedAgentCandidateExecution`](#preparedagentcandidateexecution)\>
 
-Defined in: [candidate-execution/prepare.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L98)
+Defined in: [candidate-execution/prepare.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L82)
 
 Materializes a verified candidate into one immutable evaluator-owned execution plan.
 
@@ -12118,7 +12032,7 @@ Materializes a verified candidate into one immutable evaluator-owned execution p
 
 > **parseExactAgentProfile**(`input`, `label`): `AgentProfile`
 
-Defined in: [candidate-execution/profile.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L96)
+Defined in: [candidate-execution/profile.ts:185](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L185)
 
 Parse a complete profile without silently discarding unsupported fields.
 
@@ -12142,7 +12056,7 @@ Parse a complete profile without silently discarding unsupported fields.
 
 > **parseExactAgentProfileDiff**(`input`, `label`): `AgentProfileDiff`
 
-Defined in: [candidate-execution/profile.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L103)
+Defined in: [candidate-execution/profile.ts:192](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L192)
 
 Parse a profile diff without silently discarding unsupported fields.
 
@@ -12166,7 +12080,7 @@ Parse a profile diff without silently discarding unsupported fields.
 
 > **applyExactAgentProfileDiff**(`baseInput`, `diffInput`, `label`): `AgentProfile`
 
-Defined in: [candidate-execution/profile.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L110)
+Defined in: [candidate-execution/profile.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L199)
 
 Apply one exact diff and reject any value that cannot be preserved canonically.
 
@@ -12218,7 +12132,7 @@ it to cross into candidate execution or durable receipt finalization.
 
 > **recoverExpiredAgentCandidateExecution**(`options`): `Promise`\<[`AgentCandidateExecutionFinishResult`](#agentcandidateexecutionfinishresult)\>
 
-Defined in: [candidate-execution/recover.ts:36](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L36)
+Defined in: [candidate-execution/recover.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L51)
 
 Close an expired crashed attempt from persisted non-secret handles, then record failure.
 
@@ -12238,7 +12152,7 @@ Close an expired crashed attempt from persisted non-secret handles, then record
 
 > **verifyAgentCandidateBundle**(`input`, `ports`): `Promise`\<[`VerifiedAgentCandidate`](#verifiedagentcandidate)\>
 
-Defined in: [candidate-execution/verify.ts:37](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/verify.ts#L37)
+Defined in: [candidate-execution/verify.ts:54](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/verify.ts#L54)
 
 Verifies every digest, resource, workspace, and Git object in a candidate bundle.
 
@@ -13119,7 +13033,7 @@ Cheap no-sandbox `CandidateGenerator` (the `shots=1` setting): draft surface edi
 
 > **createAgentKnowledgeReadinessCheck**(`options`): [`KnowledgeReadinessCheck`](#knowledgereadinesscheck)
 
-Defined in: [knowledge/improvement-job.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L77)
+Defined in: [knowledge/improvement-job.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L110)
 
 Build the default readiness check backed by `@tangle-network/agent-knowledge` validation and scoring.
 
@@ -13139,9 +13053,9 @@ Build the default readiness check backed by `@tangle-network/agent-knowledge` va
 
 > **runKnowledgeImprovementJob**(`options`): `Promise`\<[`KnowledgeImprovementJobResult`](#knowledgeimprovementjobresult)\>
 
-Defined in: [knowledge/improvement-job.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L108)
+Defined in: [knowledge/improvement-job.ts:141](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L141)
 
-Run the full KB improvement job: candidate workspace, runtime supervisor update, readiness check, and promotion.
+Produce a frozen KB candidate, and promote it only when an exact signed review is supplied.
 
 #### Parameters
 
@@ -13159,7 +13073,7 @@ Run the full KB improvement job: candidate workspace, runtime supervisor update,
 
 > **knowledgeReadinessDeliverable**(`options`): [`DeliverableSpec`](runtime.md#deliverablespec)\<`unknown`\>
 
-Defined in: [knowledge/supervised-update.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L91)
+Defined in: [knowledge/supervised-update.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L92)
 
 Build the completion check a supervised KB update uses to stop only when the KB is ready.
 
@@ -13179,7 +13093,7 @@ Build the completion check a supervised KB update uses to stop only when the KB
 
 > **createSupervisedKnowledgeUpdater**(`options`): [`SupervisedKnowledgeUpdater`](#supervisedknowledgeupdater)
 
-Defined in: [knowledge/supervised-update.ts:113](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L113)
+Defined in: [knowledge/supervised-update.ts:114](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L114)
 
 Create an `improveKnowledgeBase` update callback backed by runtime supervision.
 
@@ -13199,7 +13113,7 @@ Create an `improveKnowledgeBase` update callback backed by runtime supervision.
 
 > **runSupervisedKnowledgeUpdate**(`options`): `Promise`\<[`SupervisedKnowledgeUpdateResult`](#supervisedknowledgeupdateresult)\>
 
-Defined in: [knowledge/supervised-update.ts:127](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L127)
+Defined in: [knowledge/supervised-update.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L128)
 
 Run a runtime supervisor that updates one candidate knowledge base and stops on readiness.
 
@@ -13219,7 +13133,7 @@ Run a runtime supervisor that updates one candidate knowledge base and stops on
 
 > **formatSupervisedKnowledgeTask**(`options`): `string`
 
-Defined in: [knowledge/supervised-update.ts:169](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L169)
+Defined in: [knowledge/supervised-update.ts:170](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/supervised-update.ts#L170)
 
 Format the supervisor task with the KB root, readiness requirements, current findings, and metadata.
 
diff --git a/docs/api/intelligence.md b/docs/api/intelligence.md
index 92260637..a7040331 100644
--- a/docs/api/intelligence.md
+++ b/docs/api/intelligence.md
@@ -1076,229 +1076,9 @@ Intelligence-class spend ceiling. `0` refuses every intelligence spawn; `null` u
 
 ***
 
-### AgentImprovementProposal
-
-Defined in: [intelligence/improvement-cycle.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L62)
-
-#### Type Parameters
-
-##### TScenario
-
-`TScenario` *extends* `Scenario` = `Scenario`
-
-##### TArtifact
-
-`TArtifact` = `unknown`
-
-#### Properties
-
-##### schemaVersion
-
-> **schemaVersion**: `1`
-
-Defined in: [intelligence/improvement-cycle.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L66)
-
-##### kind
-
-> **kind**: `"agent-improvement-proposal"`
-
-Defined in: [intelligence/improvement-cycle.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L67)
-
-##### runId
-
-> **runId**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L68)
-
-##### surface
-
-> **surface**: [`ImproveSurface`](index.md#improvesurface)
-
-Defined in: [intelligence/improvement-cycle.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L69)
-
-##### proposedAt
-
-> **proposedAt**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L70)
-
-##### baselineProfile
-
-> **baselineProfile**: `AgentProfile`
-
-Defined in: [intelligence/improvement-cycle.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L71)
-
-##### baselineProfileHash
-
-> **baselineProfileHash**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L72)
-
-##### candidateProfile
-
-> **candidateProfile**: `AgentProfile`
-
-Defined in: [intelligence/improvement-cycle.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L73)
-
-##### candidateProfileHash
-
-> **candidateProfileHash**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:74](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L74)
-
-##### findings
-
-> **findings**: `AnalystFinding`[]
-
-Defined in: [intelligence/improvement-cycle.ts:75](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L75)
-
-##### evaluation
-
-> **evaluation**: [`AgentImprovementEvaluation`](#agentimprovementevaluation)\<`TScenario`, `TArtifact`\>
-
-Defined in: [intelligence/improvement-cycle.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L76)
-
-##### candidateBundle?
-
-> `optional` **candidateBundle?**: `AgentCandidateBundleV1`
-
-Defined in: [intelligence/improvement-cycle.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L77)
-
-##### digest
-
-> **digest**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L78)
-
-***
-
-### AgentImprovementReview
-
-Defined in: [intelligence/improvement-cycle.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L83)
-
-#### Properties
-
-##### schemaVersion
-
-> **schemaVersion**: `1`
-
-Defined in: [intelligence/improvement-cycle.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L84)
-
-##### kind
-
-> **kind**: `"agent-improvement-review"`
-
-Defined in: [intelligence/improvement-cycle.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L85)
-
-##### proposalDigest
-
-> **proposalDigest**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L86)
-
-##### candidateBundleDigest?
-
-> `optional` **candidateBundleDigest?**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L87)
-
-##### decision
-
-> **decision**: [`AgentImprovementReviewDecision`](#agentimprovementreviewdecision)
-
-Defined in: [intelligence/improvement-cycle.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L88)
-
-##### reviewedBy
-
-> **reviewedBy**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L89)
-
-##### reviewedAt
-
-> **reviewedAt**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L90)
-
-##### reason
-
-> **reason**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L91)
-
-##### feedback?
-
-> `optional` **feedback?**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L92)
-
-##### digest
-
-> **digest**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L93)
-
-***
-
-### CandidateExecutionEvidence
-
-Defined in: [intelligence/improvement-cycle.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L96)
-
-#### Properties
-
-##### proposalDigest
-
-> **proposalDigest**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L97)
-
-##### reviewDigest
-
-> **reviewDigest**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L98)
-
-##### bundleDigest
-
-> **bundleDigest**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:99](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L99)
-
-##### executionId
-
-> **executionId**: `string`
-
-Defined in: [intelligence/improvement-cycle.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L100)
-
-##### executionPlanDigest
-
-> **executionPlanDigest**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L101)
-
-##### materializationReceiptDigest
-
-> **materializationReceiptDigest**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L102)
-
-##### succeeded
-
-> **succeeded**: `boolean`
-
-Defined in: [intelligence/improvement-cycle.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L103)
-
-##### runReceiptDigest?
-
-> `optional` **runReceiptDigest?**: `` `sha256:${string}` ``
-
-Defined in: [intelligence/improvement-cycle.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L104)
-
-***
-
 ### ProposeAgentImprovementOptions
 
-Defined in: [intelligence/improvement-cycle.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L107)
+Defined in: [intelligence/improvement-cycle.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L84)
 
 #### Type Parameters
 
@@ -1316,35 +1096,33 @@ Defined in: [intelligence/improvement-cycle.ts:107](https://github.com/tangle-ne
 
 > **runId**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L108)
+Defined in: [intelligence/improvement-cycle.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L85)
 
 ##### profile
 
 > **profile**: `AgentProfile`
 
-Defined in: [intelligence/improvement-cycle.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L109)
+Defined in: [intelligence/improvement-cycle.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L86)
 
 ##### analysis
 
 > **analysis**: `Omit`\<[`RunAnalystLoopOpts`](analyst-loop.md#runanalystloopopts), `"runId"` \| `"improvementAdapter"` \| `"autoApply"`\>
 
-Defined in: [intelligence/improvement-cycle.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L110)
+Defined in: [intelligence/improvement-cycle.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L87)
 
 ##### improvement
 
 > **improvement**: [`ImproveOptions`](index.md#improveoptions)\<`TScenario`, `TArtifact`\>
 
-Defined in: [intelligence/improvement-cycle.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L111)
+Defined in: [intelligence/improvement-cycle.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L88)
 
-##### buildCandidate?
+##### buildCandidate
 
-> `optional` **buildCandidate?**: (`input`) => `AgentCandidateBundleV1` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput) \| `Promise`\<`AgentCandidateBundleV1` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)\>
+> **buildCandidate**: (`input`) => `AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput) \| `Promise`\<`AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)\>
 
-Defined in: [intelligence/improvement-cycle.ts:117](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L117)
+Defined in: [intelligence/improvement-cycle.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L90)
 
-Optional environment adapter that freezes an executable bundle after the
-measured comparison recommends the candidate. Return the sealed output of
-`buildAgentCandidateBundle` directly, or a low-level digest-free input.
+Freezes the measured winner into the exact bundle reviewed for execution.
 
 ###### Parameters
 
@@ -1360,13 +1138,13 @@ measured comparison recommends the candidate. Return the sealed output of
 
 ###### Returns
 
-`AgentCandidateBundleV1` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput) \| `Promise`\<`AgentCandidateBundleV1` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)\>
+`AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput) \| `Promise`\<`AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)\>
 
 ##### now?
 
 > `optional` **now?**: () => `Date`
 
-Defined in: [intelligence/improvement-cycle.ts:124](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L124)
+Defined in: [intelligence/improvement-cycle.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L97)
 
 ###### Returns
 
@@ -1376,7 +1154,7 @@ Defined in: [intelligence/improvement-cycle.ts:124](https://github.com/tangle-ne
 
 ### ProposeAgentImprovementResult
 
-Defined in: [intelligence/improvement-cycle.ts:127](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L127)
+Defined in: [intelligence/improvement-cycle.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L100)
 
 #### Type Parameters
 
@@ -1394,35 +1172,25 @@ Defined in: [intelligence/improvement-cycle.ts:127](https://github.com/tangle-ne
 
 > **analysis**: [`RunAnalystLoopResult`](analyst-loop.md#runanalystloopresult)
 
-Defined in: [intelligence/improvement-cycle.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L128)
+Defined in: [intelligence/improvement-cycle.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L101)
 
 ##### improvement
 
 > **improvement**: [`ImproveResult`](index.md#improveresult)\<`TScenario`, `TArtifact`\>
 
-Defined in: [intelligence/improvement-cycle.ts:129](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L129)
+Defined in: [intelligence/improvement-cycle.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L102)
 
 ##### proposal
 
-> **proposal**: [`AgentImprovementProposal`](#agentimprovementproposal)\<`TScenario`, `TArtifact`\>
+> **proposal**: `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:130](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L130)
+Defined in: [intelligence/improvement-cycle.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L103)
 
 ***
 
 ### CreateAgentImprovementProposalOptions
 
-Defined in: [intelligence/improvement-cycle.ts:133](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L133)
-
-#### Type Parameters
-
-##### TScenario
-
-`TScenario` *extends* `Scenario`
-
-##### TArtifact
-
-`TArtifact`
+Defined in: [intelligence/improvement-cycle.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L106)
 
 #### Properties
 
@@ -1430,49 +1198,37 @@ Defined in: [intelligence/improvement-cycle.ts:133](https://github.com/tangle-ne
 
 > **runId**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:134](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L134)
-
-##### surface
-
-> **surface**: [`ImproveSurface`](index.md#improvesurface)
-
-Defined in: [intelligence/improvement-cycle.ts:135](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L135)
+Defined in: [intelligence/improvement-cycle.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L107)
 
 ##### baselineProfile
 
 > **baselineProfile**: `AgentProfile`
 
-Defined in: [intelligence/improvement-cycle.ts:136](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L136)
-
-##### candidateProfile
-
-> **candidateProfile**: `AgentProfile`
-
-Defined in: [intelligence/improvement-cycle.ts:137](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L137)
+Defined in: [intelligence/improvement-cycle.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L108)
 
 ##### findings
 
 > **findings**: readonly `AnalystFinding`[]
 
-Defined in: [intelligence/improvement-cycle.ts:138](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L138)
+Defined in: [intelligence/improvement-cycle.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L109)
 
 ##### evaluation
 
-> **evaluation**: `SelfImproveResult`\<`TScenario`, `TArtifact`\>
+> **evaluation**: `AgentImprovementMeasuredComparison`
 
-Defined in: [intelligence/improvement-cycle.ts:139](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L139)
+Defined in: [intelligence/improvement-cycle.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L110)
 
-##### candidateBundle?
+##### candidateBundle
 
-> `optional` **candidateBundle?**: `AgentCandidateBundleV1` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)
+> **candidateBundle**: `AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)
 
-Defined in: [intelligence/improvement-cycle.ts:140](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L140)
+Defined in: [intelligence/improvement-cycle.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L111)
 
 ##### now?
 
 > `optional` **now?**: () => `Date`
 
-Defined in: [intelligence/improvement-cycle.ts:141](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L141)
+Defined in: [intelligence/improvement-cycle.ts:112](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L112)
 
 ###### Returns
 
@@ -1482,39 +1238,39 @@ Defined in: [intelligence/improvement-cycle.ts:141](https://github.com/tangle-ne
 
 ### ReviewAgentImprovementInput
 
-Defined in: [intelligence/improvement-cycle.ts:144](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L144)
+Defined in: [intelligence/improvement-cycle.ts:125](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L125)
 
 #### Properties
 
 ##### decision
 
-> **decision**: [`AgentImprovementReviewDecision`](#agentimprovementreviewdecision)
+> **decision**: `AgentImprovementReviewDecision`
 
-Defined in: [intelligence/improvement-cycle.ts:145](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L145)
+Defined in: [intelligence/improvement-cycle.ts:126](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L126)
 
 ##### reviewedBy
 
 > **reviewedBy**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:146](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L146)
+Defined in: [intelligence/improvement-cycle.ts:127](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L127)
 
 ##### reason
 
 > **reason**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:147](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L147)
+Defined in: [intelligence/improvement-cycle.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L128)
 
 ##### feedback?
 
 > `optional` **feedback?**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:148](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L148)
+Defined in: [intelligence/improvement-cycle.ts:129](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L129)
 
 ##### now?
 
 > `optional` **now?**: () => `Date`
 
-Defined in: [intelligence/improvement-cycle.ts:149](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L149)
+Defined in: [intelligence/improvement-cycle.ts:130](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L130)
 
 ###### Returns
 
@@ -1524,27 +1280,27 @@ Defined in: [intelligence/improvement-cycle.ts:149](https://github.com/tangle-ne
 
 ### ExecuteApprovedAgentCandidateOptions
 
-Defined in: [intelligence/improvement-cycle.ts:152](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L152)
+Defined in: [intelligence/improvement-cycle.ts:133](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L133)
 
 #### Properties
 
 ##### proposal
 
-> **proposal**: [`AgentImprovementProposal`](#agentimprovementproposal)
+> **proposal**: `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:153](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L153)
+Defined in: [intelligence/improvement-cycle.ts:134](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L134)
 
 ##### review
 
-> **review**: [`AgentImprovementReview`](#agentimprovementreview)
+> **review**: `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:154](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L154)
+Defined in: [intelligence/improvement-cycle.ts:135](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L135)
 
 ##### authorizeReview
 
 > **authorizeReview**: (`review`, `proposal`) => `boolean` \| `Promise`\<`boolean`\>
 
-Defined in: [intelligence/improvement-cycle.ts:156](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L156)
+Defined in: [intelligence/improvement-cycle.ts:137](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L137)
 
 Product-owned authentication check for the persisted approval record.
 
@@ -1552,11 +1308,11 @@ Product-owned authentication check for the persisted approval record.
 
 ###### review
 
-[`AgentImprovementReview`](#agentimprovementreview)
+`AgentImprovementReview`
 
 ###### proposal
 
-[`AgentImprovementProposal`](#agentimprovementproposal)
+`AgentImprovementProposal`
 
 ###### Returns
 
@@ -1566,51 +1322,63 @@ Product-owned authentication check for the persisted approval record.
 
 > **task**: [`AgentCandidateTaskExecution`](index.md#agentcandidatetaskexecution)
 
-Defined in: [intelligence/improvement-cycle.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L160)
+Defined in: [intelligence/improvement-cycle.ts:141](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L141)
 
 ##### ports
 
 > **ports**: [`AgentCandidateExecutionPorts`](index.md#agentcandidateexecutionports)
 
-Defined in: [intelligence/improvement-cycle.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L161)
+Defined in: [intelligence/improvement-cycle.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L142)
 
 ##### preparation?
 
 > `optional` **preparation?**: [`PrepareAgentCandidateExecutionOptions`](index.md#prepareagentcandidateexecutionoptions)
 
-Defined in: [intelligence/improvement-cycle.ts:162](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L162)
+Defined in: [intelligence/improvement-cycle.ts:143](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L143)
 
 ##### execution
 
 > **execution**: [`ExecutePreparedAgentCandidateOptions`](index.md#executepreparedagentcandidateoptions)
 
-Defined in: [intelligence/improvement-cycle.ts:163](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L163)
+Defined in: [intelligence/improvement-cycle.ts:144](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L144)
 
 ***
 
-### ExecuteApprovedAgentCandidateResult
+### VerifyCandidateExecutionEvidenceOptions
 
-Defined in: [intelligence/improvement-cycle.ts:166](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L166)
+Defined in: [intelligence/improvement-cycle.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L157)
 
 #### Properties
 
-##### finalization
+##### proposal
+
+> **proposal**: `AgentImprovementProposal`
+
+Defined in: [intelligence/improvement-cycle.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L158)
+
+##### review
+
+> **review**: `AgentImprovementReview`
+
+Defined in: [intelligence/improvement-cycle.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L159)
+
+##### expectedCount
 
-> **finalization**: [`AgentCandidateRunFinalization`](index.md#agentcandidaterunfinalization)
+> **expectedCount**: `number`
 
-Defined in: [intelligence/improvement-cycle.ts:167](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L167)
+Defined in: [intelligence/improvement-cycle.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L160)
 
-##### evidence
+##### resolvedResources?
 
-> **evidence**: [`CandidateExecutionEvidence`](#candidateexecutionevidence)
+> `optional` **resolvedResources?**: `ReadonlyMap`\<`` `sha256:${string}` ``, `string`\>
 
-Defined in: [intelligence/improvement-cycle.ts:168](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L168)
+Defined in: [intelligence/improvement-cycle.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L161)
 
 ***
 
 ### UsageSplit
 
-Defined in: [intelligence/index.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L142)
+Defined in: [intelligence/index.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L157)
 
 The per-class cost split carried by every trace and outcome. `off` ⇒
 `intelligenceUsd: 0` by construction — there is no intelligence spawn to
@@ -1622,7 +1390,7 @@ bill. This is a classification on the trace, NOT a budget-pool split.
 
 > **inferenceUsd**: `number`
 
-Defined in: [intelligence/index.ts:144](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L144)
+Defined in: [intelligence/index.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L159)
 
 Base-stream (model) spend in USD.
 
@@ -1630,7 +1398,7 @@ Base-stream (model) spend in USD.
 
 > **intelligenceUsd**: `number`
 
-Defined in: [intelligence/index.ts:146](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L146)
+Defined in: [intelligence/index.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L161)
 
 Intelligence-spawn spend in USD. Provably `0` at the OFF tier.
 
@@ -1638,7 +1406,7 @@ Intelligence-spawn spend in USD. Provably `0` at the OFF tier.
 
 ### RunRecord
 
-Defined in: [intelligence/index.ts:156](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L156)
+Defined in: [intelligence/index.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L171)
 
 The typed record `withIntelligence` sends per call — serialized through the
 shipped OTLP builders to the plane's `/v1/otlp` ingest. `input`/`output` are
@@ -1652,43 +1420,43 @@ tree under the same `traceId`.
 
 > **runId**: `string`
 
-Defined in: [intelligence/index.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L157)
+Defined in: [intelligence/index.ts:172](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L172)
 
 ##### traceId
 
 > **traceId**: `string`
 
-Defined in: [intelligence/index.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L158)
+Defined in: [intelligence/index.ts:173](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L173)
 
 ##### project
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L159)
+Defined in: [intelligence/index.ts:174](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L174)
 
 ##### target
 
 > **target**: `string`
 
-Defined in: [intelligence/index.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L160)
+Defined in: [intelligence/index.ts:175](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L175)
 
 ##### input
 
 > **input**: `unknown`
 
-Defined in: [intelligence/index.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L161)
+Defined in: [intelligence/index.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L176)
 
 ##### output
 
 > **output**: `unknown`
 
-Defined in: [intelligence/index.ts:162](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L162)
+Defined in: [intelligence/index.ts:177](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L177)
 
 ##### outcome
 
 > **outcome**: `object`
 
-Defined in: [intelligence/index.ts:163](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L163)
+Defined in: [intelligence/index.ts:178](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L178)
 
 ###### success?
 
@@ -1706,61 +1474,61 @@ Defined in: [intelligence/index.ts:163](https://github.com/tangle-network/agent-
 
 > `optional` **model?**: `string`
 
-Defined in: [intelligence/index.ts:168](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L168)
+Defined in: [intelligence/index.ts:183](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L183)
 
 ##### provider?
 
 > `optional` **provider?**: `string`
 
-Defined in: [intelligence/index.ts:169](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L169)
+Defined in: [intelligence/index.ts:184](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L184)
 
 ##### loopEvents?
 
 > `optional` **loopEvents?**: [`LoopTraceEvent`](runtime.md#looptraceevent)[]
 
-Defined in: [intelligence/index.ts:170](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L170)
+Defined in: [intelligence/index.ts:185](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L185)
 
 ##### runtimeEvents?
 
 > `optional` **runtimeEvents?**: [`RuntimeStreamEvent`](index.md#runtimestreamevent)[]
 
-Defined in: [intelligence/index.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L171)
+Defined in: [intelligence/index.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L186)
 
 ##### profile?
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [intelligence/index.ts:172](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L172)
+Defined in: [intelligence/index.ts:187](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L187)
 
 ##### sessionId?
 
 > `optional` **sessionId?**: `string`
 
-Defined in: [intelligence/index.ts:173](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L173)
+Defined in: [intelligence/index.ts:188](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L188)
 
 ##### harness?
 
 > `optional` **harness?**: `string`
 
-Defined in: [intelligence/index.ts:174](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L174)
+Defined in: [intelligence/index.ts:189](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L189)
 
 ##### repository?
 
 > `optional` **repository?**: `string`
 
-Defined in: [intelligence/index.ts:175](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L175)
+Defined in: [intelligence/index.ts:190](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L190)
 
 ##### commitSha?
 
 > `optional` **commitSha?**: `string`
 
-Defined in: [intelligence/index.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L176)
+Defined in: [intelligence/index.ts:191](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L191)
 
 ##### timing?
 
 > `optional` **timing?**: `object`
 
-Defined in: [intelligence/index.ts:177](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L177)
+Defined in: [intelligence/index.ts:192](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L192)
 
 ###### startedAt
 
@@ -1778,7 +1546,7 @@ Defined in: [intelligence/index.ts:177](https://github.com/tangle-network/agent-
 
 > `optional` **tokens?**: `object`
 
-Defined in: [intelligence/index.ts:178](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L178)
+Defined in: [intelligence/index.ts:193](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L193)
 
 ###### input
 
@@ -1800,7 +1568,7 @@ Defined in: [intelligence/index.ts:178](https://github.com/tangle-network/agent-
 
 > `optional` **error?**: `object`
 
-Defined in: [intelligence/index.ts:184](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L184)
+Defined in: [intelligence/index.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L199)
 
 ###### name
 
@@ -1816,9 +1584,9 @@ Defined in: [intelligence/index.ts:184](https://github.com/tangle-network/agent-
 
 ##### candidateExecution?
 
-> `optional` **candidateExecution?**: [`CandidateExecutionEvidence`](#candidateexecutionevidence)
+> `optional` **candidateExecution?**: `CandidateExecutionEvidence`
 
-Defined in: [intelligence/index.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L186)
+Defined in: [intelligence/index.ts:201](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L201)
 
 Exact proposal → review → execution → receipt linkage for candidate runs.
 
@@ -1826,7 +1594,7 @@ Exact proposal → review → execution → receipt linkage for candidate runs.
 
 ### RunReport
 
-Defined in: [intelligence/index.ts:195](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L195)
+Defined in: [intelligence/index.ts:210](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L210)
 
 What an agent reports (via `applied.record`) to enrich the [RunRecord](#runrecord)
 sent for its call. All optional — an un-recorded run still sends input/output
@@ -1839,79 +1607,79 @@ as pure inference (the base stream).
 
 > `optional` **success?**: `boolean`
 
-Defined in: [intelligence/index.ts:196](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L196)
+Defined in: [intelligence/index.ts:211](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L211)
 
 ##### score?
 
 > `optional` **score?**: `number`
 
-Defined in: [intelligence/index.ts:197](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L197)
+Defined in: [intelligence/index.ts:212](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L212)
 
 ##### usage?
 
 > `optional` **usage?**: `Partial`\<[`UsageSplit`](#usagesplit)\>
 
-Defined in: [intelligence/index.ts:198](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L198)
+Defined in: [intelligence/index.ts:213](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L213)
 
 ##### costUsd?
 
 > `optional` **costUsd?**: `number`
 
-Defined in: [intelligence/index.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L199)
+Defined in: [intelligence/index.ts:214](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L214)
 
 ##### model?
 
 > `optional` **model?**: `string`
 
-Defined in: [intelligence/index.ts:200](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L200)
+Defined in: [intelligence/index.ts:215](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L215)
 
 ##### provider?
 
 > `optional` **provider?**: `string`
 
-Defined in: [intelligence/index.ts:201](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L201)
+Defined in: [intelligence/index.ts:216](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L216)
 
 ##### loopEvents?
 
 > `optional` **loopEvents?**: [`LoopTraceEvent`](runtime.md#looptraceevent)[]
 
-Defined in: [intelligence/index.ts:202](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L202)
+Defined in: [intelligence/index.ts:217](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L217)
 
 ##### runtimeEvents?
 
 > `optional` **runtimeEvents?**: [`RuntimeStreamEvent`](index.md#runtimestreamevent)[]
 
-Defined in: [intelligence/index.ts:203](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L203)
+Defined in: [intelligence/index.ts:218](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L218)
 
 ##### profile?
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [intelligence/index.ts:204](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L204)
+Defined in: [intelligence/index.ts:219](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L219)
 
 ##### sessionId?
 
 > `optional` **sessionId?**: `string`
 
-Defined in: [intelligence/index.ts:205](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L205)
+Defined in: [intelligence/index.ts:220](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L220)
 
 ##### harness?
 
 > `optional` **harness?**: `string`
 
-Defined in: [intelligence/index.ts:206](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L206)
+Defined in: [intelligence/index.ts:221](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L221)
 
 ##### commitSha?
 
 > `optional` **commitSha?**: `string`
 
-Defined in: [intelligence/index.ts:207](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L207)
+Defined in: [intelligence/index.ts:222](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L222)
 
 ##### tokens?
 
 > `optional` **tokens?**: `object`
 
-Defined in: [intelligence/index.ts:208](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L208)
+Defined in: [intelligence/index.ts:223](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L223)
 
 ###### input
 
@@ -1933,7 +1701,7 @@ Defined in: [intelligence/index.ts:208](https://github.com/tangle-network/agent-
 
 > `optional` **error?**: `object`
 
-Defined in: [intelligence/index.ts:209](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L209)
+Defined in: [intelligence/index.ts:224](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L224)
 
 ###### name
 
@@ -1949,15 +1717,15 @@ Defined in: [intelligence/index.ts:209](https://github.com/tangle-network/agent-
 
 ##### candidateExecution?
 
-> `optional` **candidateExecution?**: [`CandidateExecutionEvidence`](#candidateexecutionevidence)
+> `optional` **candidateExecution?**: `CandidateExecutionEvidence`
 
-Defined in: [intelligence/index.ts:210](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L210)
+Defined in: [intelligence/index.ts:225](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L225)
 
 ***
 
 ### RepoConfig
 
-Defined in: [intelligence/index.ts:216](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L216)
+Defined in: [intelligence/index.ts:231](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L231)
 
 Repo coordinates a product may declare for the (later) Gated-PR mode. The
  Observe slice only records their PRESENCE for `doctor()`; it never touches
@@ -1969,25 +1737,25 @@ Repo coordinates a product may declare for the (later) Gated-PR mode. The
 
 > **owner**: `string`
 
-Defined in: [intelligence/index.ts:217](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L217)
+Defined in: [intelligence/index.ts:232](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L232)
 
 ##### name
 
 > **name**: `string`
 
-Defined in: [intelligence/index.ts:218](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L218)
+Defined in: [intelligence/index.ts:233](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L233)
 
 ##### baseBranch
 
 > **baseBranch**: `string`
 
-Defined in: [intelligence/index.ts:219](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L219)
+Defined in: [intelligence/index.ts:234](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L234)
 
 ***
 
 ### IntelligenceConfig
 
-Defined in: [intelligence/index.ts:225](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L225)
+Defined in: [intelligence/index.ts:240](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L240)
 
 Client configuration. `project` + `apiKey` are the Observe minimum; the
  rest tune effort, endpoint, redaction, and (for `doctor()` readiness)
@@ -2003,7 +1771,7 @@ Client configuration. `project` + `apiKey` are the Observe minimum; the
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:227](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L227)
+Defined in: [intelligence/index.ts:242](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L242)
 
 Stable project id — the tenant dimension every trace is tagged with.
 
@@ -2011,7 +1779,7 @@ Stable project id — the tenant dimension every trace is tagged with.
 
 > `optional` **apiKey?**: `string`
 
-Defined in: [intelligence/index.ts:229](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L229)
+Defined in: [intelligence/index.ts:244](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L244)
 
 Bearer key for the Intelligence ingest. Reads `TANGLE_API_KEY` when omitted.
 
@@ -2019,7 +1787,7 @@ Bearer key for the Intelligence ingest. Reads `TANGLE_API_KEY` when omitted.
 
 > `optional` **effort?**: [`EffortTier`](#efforttier) \| \{ `tier`: [`EffortTier`](#efforttier); `overrides?`: `Partial`\<[`EffortSettings`](#effortsettings)\>; \}
 
-Defined in: [intelligence/index.ts:231](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L231)
+Defined in: [intelligence/index.ts:246](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L246)
 
 Effort tier (default `'standard'`) plus optional per-field overrides.
 
@@ -2027,7 +1795,7 @@ Effort tier (default `'standard'`) plus optional per-field overrides.
 
 > `optional` **baseUrl?**: `string`
 
-Defined in: [intelligence/index.ts:239](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L239)
+Defined in: [intelligence/index.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L254)
 
 The ONE Tangle Intelligence base URL — both the send (OTLP `/v1/otlp`) and
 receive (`/v1/profiles/:target/composed`) paths derive from it. Reads
@@ -2039,7 +1807,7 @@ key the ingest requires); absent a key, export is a no-op.
 
 > `optional` **redact?**: `false` \| [`Redactor`](#redactor)
 
-Defined in: [intelligence/index.ts:245](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L245)
+Defined in: [intelligence/index.ts:260](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L260)
 
 Redaction hook run over every exported input/output. A function replaces
 the default scrubber; `false` opts out entirely (raw fidelity, caller has
@@ -2049,7 +1817,7 @@ sanitized upstream); omitted ⇒ the built-in `defaultRedactor`.
 
 > `optional` **surfaces?**: `string`[]
 
-Defined in: [intelligence/index.ts:247](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L247)
+Defined in: [intelligence/index.ts:262](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L262)
 
 Mutable surfaces a later PR mode would edit. Recorded for `doctor()` only.
 
@@ -2057,7 +1825,7 @@ Mutable surfaces a later PR mode would edit. Recorded for `doctor()` only.
 
 > `optional` **checks?**: `string`[]
 
-Defined in: [intelligence/index.ts:249](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L249)
+Defined in: [intelligence/index.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L264)
 
 Verification checks a later PR mode would gate on. Recorded for `doctor()` only.
 
@@ -2065,7 +1833,7 @@ Verification checks a later PR mode would gate on. Recorded for `doctor()` only.
 
 > `optional` **repo?**: [`RepoConfig`](#repoconfig)
 
-Defined in: [intelligence/index.ts:251](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L251)
+Defined in: [intelligence/index.ts:266](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L266)
 
 Repo access a later PR mode would need. Recorded for `doctor()` only.
 
@@ -2073,7 +1841,7 @@ Repo access a later PR mode would need. Recorded for `doctor()` only.
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [intelligence/index.ts:253](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L253)
+Defined in: [intelligence/index.ts:268](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L268)
 
 Full canonical profile used for this agent. Exported redacted with a stable hash.
 
@@ -2081,7 +1849,7 @@ Full canonical profile used for this agent. Exported redacted with a stable hash
 
 > `optional` **commitSha?**: `string`
 
-Defined in: [intelligence/index.ts:255](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L255)
+Defined in: [intelligence/index.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L270)
 
 Commit that produced the running agent, when known.
 
@@ -2089,7 +1857,7 @@ Commit that produced the running agent, when known.
 
 > `optional` **runtimeTelemetry?**: [`RuntimeTelemetryOptions`](index.md#runtimetelemetryoptions)
 
-Defined in: [intelligence/index.ts:257](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L257)
+Defined in: [intelligence/index.ts:272](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L272)
 
 Runtime-event payload policy. Tool inputs/results remain off unless explicitly enabled.
 
@@ -2097,7 +1865,7 @@ Runtime-event payload policy. Tool inputs/results remain off unless explicitly e
 
 > `optional` **payloadAttributes?**: `"metadata"` \| `"full"`
 
-Defined in: [intelligence/index.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L264)
+Defined in: [intelligence/index.ts:279](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L279)
 
 Payloads are metadata-only by default: the run span carries a stable hash
 and UTF-8 byte count, but not the redacted content. Set `full` only when
@@ -2108,7 +1876,7 @@ inputs, outputs, and profiles.
 
 ### TraceMeta
 
-Defined in: [intelligence/index.ts:268](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L268)
+Defined in: [intelligence/index.ts:283](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L283)
 
 Metadata describing one traced run. `runId`/`traceId` default to fresh ids.
 
@@ -2118,7 +1886,7 @@ Metadata describing one traced run. `runId`/`traceId` default to fresh ids.
 
 > `optional` **input?**: `unknown`
 
-Defined in: [intelligence/index.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L270)
+Defined in: [intelligence/index.ts:285](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L285)
 
 The run's input — exported through the redactor.
 
@@ -2126,7 +1894,7 @@ The run's input — exported through the redactor.
 
 > `optional` **runId?**: `string`
 
-Defined in: [intelligence/index.ts:272](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L272)
+Defined in: [intelligence/index.ts:287](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L287)
 
 Stable run id. Defaults to a fresh id.
 
@@ -2134,7 +1902,7 @@ Stable run id. Defaults to a fresh id.
 
 > `optional` **traceId?**: `string`
 
-Defined in: [intelligence/index.ts:274](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L274)
+Defined in: [intelligence/index.ts:289](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L289)
 
 32-hex trace id. Defaults to a fresh id.
 
@@ -2142,7 +1910,7 @@ Defined in: [intelligence/index.ts:274](https://github.com/tangle-network/agent-
 
 > `optional` **model?**: `string`
 
-Defined in: [intelligence/index.ts:276](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L276)
+Defined in: [intelligence/index.ts:291](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L291)
 
 Model id, when known — stamped on the span.
 
@@ -2150,7 +1918,7 @@ Model id, when known — stamped on the span.
 
 > `optional` **provider?**: `string`
 
-Defined in: [intelligence/index.ts:278](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L278)
+Defined in: [intelligence/index.ts:293](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L293)
 
 Provider name, when known — stamped on the span.
 
@@ -2158,7 +1926,7 @@ Provider name, when known — stamped on the span.
 
 > `optional` **labels?**: `Record`\<`string`, `string` \| `number` \| `boolean`\>
 
-Defined in: [intelligence/index.ts:280](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L280)
+Defined in: [intelligence/index.ts:295](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L295)
 
 Arbitrary extra labels (string/number/boolean) stamped on the span.
 
@@ -2166,7 +1934,7 @@ Arbitrary extra labels (string/number/boolean) stamped on the span.
 
 ### TraceHandle
 
-Defined in: [intelligence/index.ts:289](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L289)
+Defined in: [intelligence/index.ts:304](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L304)
 
 The trace handle a `traceRun` body records into. `recordOutput` captures the
 agent's result (redacted on export); `recordOutcome` captures the scored
@@ -2179,7 +1947,7 @@ an un-recorded run still exports a span with whatever was set.
 
 > **recordOutput**(`output`): `void`
 
-Defined in: [intelligence/index.ts:291](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L291)
+Defined in: [intelligence/index.ts:306](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L306)
 
 Capture the run's output. Exported through the redactor.
 
@@ -2197,7 +1965,7 @@ Capture the run's output. Exported through the redactor.
 
 > **recordOutcome**(`outcome`): `void`
 
-Defined in: [intelligence/index.ts:298](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L298)
+Defined in: [intelligence/index.ts:313](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L313)
 
 Capture the run's outcome. `usage` defaults to inference-only
 (`intelligenceUsd: 0`) — the OFF baseline; an intelligence-enabled run
@@ -2232,7 +2000,7 @@ treated as pure inference.
 
 ### RecordTraceMeta
 
-Defined in: [intelligence/index.ts:307](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L307)
+Defined in: [intelligence/index.ts:322](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L322)
 
 Metadata for [IntelligenceClient.recordTrace](#recordtrace).
 
@@ -2242,7 +2010,7 @@ Metadata for [IntelligenceClient.recordTrace](#recordtrace).
 
 > `optional` **traceId?**: `string`
 
-Defined in: [intelligence/index.ts:309](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L309)
+Defined in: [intelligence/index.ts:324](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L324)
 
 32-hex trace id to anchor every span to. Defaults to a fresh id.
 
@@ -2250,7 +2018,7 @@ Defined in: [intelligence/index.ts:309](https://github.com/tangle-network/agent-
 
 > `optional` **rootParentSpanId?**: `string`
 
-Defined in: [intelligence/index.ts:312](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L312)
+Defined in: [intelligence/index.ts:327](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L327)
 
 Span id of an enclosing span the loop root should parent under (e.g. a
  `traceRun` span). Omitted ⇒ the loop root is the trace root.
@@ -2259,7 +2027,7 @@ Span id of an enclosing span the loop root should parent under (e.g. a
 
 ### TraceOutcome
 
-Defined in: [intelligence/index.ts:317](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L317)
+Defined in: [intelligence/index.ts:332](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L332)
 
 The resolved outcome of one traced run, surfaced on the export span and
  available to the caller for downstream billing assertions.
@@ -2270,25 +2038,25 @@ The resolved outcome of one traced run, surfaced on the export span and
 
 > **runId**: `string`
 
-Defined in: [intelligence/index.ts:318](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L318)
+Defined in: [intelligence/index.ts:333](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L333)
 
 ##### traceId
 
 > **traceId**: `string`
 
-Defined in: [intelligence/index.ts:319](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L319)
+Defined in: [intelligence/index.ts:334](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L334)
 
 ##### project
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:320](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L320)
+Defined in: [intelligence/index.ts:335](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L335)
 
 ##### effort
 
 > **effort**: [`EffortSettings`](#effortsettings)
 
-Defined in: [intelligence/index.ts:322](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L322)
+Defined in: [intelligence/index.ts:337](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L337)
 
 The resolved effort settings this run executed under.
 
@@ -2296,7 +2064,7 @@ The resolved effort settings this run executed under.
 
 > **intelligenceOff**: `boolean`
 
-Defined in: [intelligence/index.ts:324](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L324)
+Defined in: [intelligence/index.ts:339](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L339)
 
 True when this run ran as pure passthrough (the OFF floor).
 
@@ -2304,19 +2072,19 @@ True when this run ran as pure passthrough (the OFF floor).
 
 > `optional` **success?**: `boolean`
 
-Defined in: [intelligence/index.ts:325](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L325)
+Defined in: [intelligence/index.ts:340](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L340)
 
 ##### score?
 
 > `optional` **score?**: `number`
 
-Defined in: [intelligence/index.ts:326](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L326)
+Defined in: [intelligence/index.ts:341](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L341)
 
 ##### usage
 
 > **usage**: [`UsageSplit`](#usagesplit)
 
-Defined in: [intelligence/index.ts:328](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L328)
+Defined in: [intelligence/index.ts:343](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L343)
 
 Per-class billing split. `intelligenceUsd` is `0` at the OFF tier.
 
@@ -2324,7 +2092,7 @@ Per-class billing split. `intelligenceUsd` is `0` at the OFF tier.
 
 ### IntelligenceClient
 
-Defined in: [intelligence/index.ts:332](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L332)
+Defined in: [intelligence/index.ts:347](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L347)
 
 The Observe-mode Intelligence client.
 
@@ -2334,7 +2102,7 @@ The Observe-mode Intelligence client.
 
 > `readonly` **project**: `string`
 
-Defined in: [intelligence/index.ts:334](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L334)
+Defined in: [intelligence/index.ts:349](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L349)
 
 The resolved project id.
 
@@ -2342,7 +2110,7 @@ The resolved project id.
 
 > `readonly` **effort**: [`EffortSettings`](#effortsettings)
 
-Defined in: [intelligence/index.ts:336](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L336)
+Defined in: [intelligence/index.ts:351](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L351)
 
 The resolved effort settings.
 
@@ -2352,7 +2120,7 @@ The resolved effort settings.
 
 > **traceRun**\<`T`\>(`meta`, `fn`): `Promise`\<`T`\>
 
-Defined in: [intelligence/index.ts:342](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L342)
+Defined in: [intelligence/index.ts:357](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L357)
 
 Run `fn` under a trace, export one span best-effort, and return whatever
 `fn` returns. Telemetry-export failures are swallowed; an error THROWN by
@@ -2382,7 +2150,7 @@ Run `fn` under a trace, export one span best-effort, and return whatever
 
 > **recordTrace**(`events`, `meta?`): `string`
 
-Defined in: [intelligence/index.ts:352](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L352)
+Defined in: [intelligence/index.ts:367](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L367)
 
 Export a run's full loop topology — the ordered `LoopTraceEvent` stream a
 `runLoop`/`Supervisor` run emits — as a nested OTLP span tree (loop → round →
@@ -2410,7 +2178,7 @@ readonly [`LoopTraceEvent`](runtime.md#looptraceevent)[]
 
 > **exportRunRecord**(`record`): `string`
 
-Defined in: [intelligence/index.ts:360](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L360)
+Defined in: [intelligence/index.ts:375](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L375)
 
 Send one typed [RunRecord](#runrecord) — the run's flat span (input/output/outcome/
 usage/model/provider, redacted) plus, when `loopEvents` are present, the
@@ -2432,7 +2200,7 @@ Best-effort: export failures are swallowed. Returns the record's `traceId`.
 
 > **freshRunId**(): `string`
 
-Defined in: [intelligence/index.ts:362](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L362)
+Defined in: [intelligence/index.ts:377](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L377)
 
 Mint a fresh run id (`run-<hex>`).
 
@@ -2444,7 +2212,7 @@ Mint a fresh run id (`run-<hex>`).
 
 > **freshTraceId**(): `string`
 
-Defined in: [intelligence/index.ts:364](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L364)
+Defined in: [intelligence/index.ts:379](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L379)
 
 Mint a fresh 32-hex trace id.
 
@@ -2456,7 +2224,7 @@ Mint a fresh 32-hex trace id.
 
 > **doctor**(): [`DoctorReport`](#doctorreport)
 
-Defined in: [intelligence/index.ts:370](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L370)
+Defined in: [intelligence/index.ts:385](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L385)
 
 Network-free readiness report: which adoption modes are reachable given
 this config. Observe is always reachable; Recommend needs outcomes; PR
@@ -2470,7 +2238,7 @@ needs checks + surfaces + repo.
 
 > **flush**(): `Promise`\<`void`\>
 
-Defined in: [intelligence/index.ts:372](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L372)
+Defined in: [intelligence/index.ts:387](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L387)
 
 Flush any pending export spans. Best-effort; resolves even if export fails.
 
@@ -2482,7 +2250,7 @@ Flush any pending export spans. Best-effort; resolves even if export fails.
 
 ### ModeReadiness
 
-Defined in: [intelligence/index.ts:376](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L376)
+Defined in: [intelligence/index.ts:391](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L391)
 
 One mode's readiness verdict.
 
@@ -2492,13 +2260,13 @@ One mode's readiness verdict.
 
 > **ready**: `boolean`
 
-Defined in: [intelligence/index.ts:377](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L377)
+Defined in: [intelligence/index.ts:392](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L392)
 
 ##### missing
 
 > **missing**: `string`[]
 
-Defined in: [intelligence/index.ts:379](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L379)
+Defined in: [intelligence/index.ts:394](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L394)
 
 Inputs this mode still needs, when not ready. Empty when ready.
 
@@ -2506,7 +2274,7 @@ Inputs this mode still needs, when not ready. Empty when ready.
 
 ### DoctorReport
 
-Defined in: [intelligence/index.ts:383](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L383)
+Defined in: [intelligence/index.ts:398](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L398)
 
 The `doctor()` readiness report — Mode-readiness without any network call.
 
@@ -2516,19 +2284,19 @@ The `doctor()` readiness report — Mode-readiness without any network call.
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:384](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L384)
+Defined in: [intelligence/index.ts:399](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L399)
 
 ##### effort
 
 > **effort**: [`EffortSettings`](#effortsettings)
 
-Defined in: [intelligence/index.ts:385](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L385)
+Defined in: [intelligence/index.ts:400](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L400)
 
 ##### exportConfigured
 
 > **exportConfigured**: `boolean`
 
-Defined in: [intelligence/index.ts:387](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L387)
+Defined in: [intelligence/index.ts:402](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L402)
 
 True when an OTLP endpoint is configured (export will actually ship).
 
@@ -2536,7 +2304,7 @@ True when an OTLP endpoint is configured (export will actually ship).
 
 > **modes**: `object`
 
-Defined in: [intelligence/index.ts:388](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L388)
+Defined in: [intelligence/index.ts:403](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L403)
 
 ###### observe
 
@@ -2760,6 +2528,170 @@ carrying an un-admitted binding kind is a hard error, not a soft drop).
 
 ***
 
+### CreateSandboxApprovedCandidateExecutorOptions
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L60)
+
+#### Properties
+
+##### client
+
+> **client**: `SandboxClientPort`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L61)
+
+##### ports
+
+> **ports**: [`AgentCandidateExecutionPorts`](index.md#agentcandidateexecutionports)
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L62)
+
+##### grader
+
+> **grader**: [`AgentCandidateBenchmarkGraderPort`](index.md#agentcandidatebenchmarkgraderport)
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L63)
+
+##### outputArtifacts
+
+> **outputArtifacts**: [`AgentCandidateOutputArtifactPort`](index.md#agentcandidateoutputartifactport)
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L64)
+
+##### traceStore
+
+> **traceStore**: `TraceStore`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L65)
+
+##### claimStore
+
+> **claimStore**: [`AgentCandidateExecutionClaimStore`](index.md#agentcandidateexecutionclaimstore)
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L66)
+
+##### authorizeReview
+
+> **authorizeReview**: (`review`, `proposal`) => `boolean` \| `Promise`\<`boolean`\>
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L67)
+
+###### Parameters
+
+###### review
+
+`AgentImprovementReview`
+
+###### proposal
+
+`AgentImprovementProposal`
+
+###### Returns
+
+`boolean` \| `Promise`\<`boolean`\>
+
+##### sandbox?
+
+> `optional` **sandbox?**: `object`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L71)
+
+###### teamId?
+
+> `optional` **teamId?**: `string`
+
+###### resources?
+
+> `optional` **resources?**: `SandboxResources`
+
+###### createTimeoutMs?
+
+> `optional` **createTimeoutMs?**: `number`
+
+###### evidenceRetentionSeconds?
+
+> `optional` **evidenceRetentionSeconds?**: `number`
+
+##### cleanupTimeoutMs?
+
+> `optional` **cleanupTimeoutMs?**: `number`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L77)
+
+##### resultTimeoutMs?
+
+> `optional` **resultTimeoutMs?**: `number`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L78)
+
+***
+
+### SandboxApprovedCandidateExecution
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L81)
+
+#### Properties
+
+##### proposal
+
+> **proposal**: `AgentImprovementProposal`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L82)
+
+##### review
+
+> **review**: `AgentImprovementReview`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L83)
+
+##### task
+
+> **task**: [`AgentCandidateTaskExecution`](index.md#agentcandidatetaskexecution)
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L84)
+
+##### preparation?
+
+> `optional` **preparation?**: [`PrepareAgentCandidateExecutionOptions`](index.md#prepareagentcandidateexecutionoptions)
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L85)
+
+***
+
+### SandboxApprovedCandidateExecutor
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L88)
+
+#### Properties
+
+##### executor
+
+> `readonly` **executor**: [`AgentCandidateExecutorPort`](index.md#agentcandidateexecutorport)
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L90)
+
+The same port is usable by Runtime's expired-claim recovery path.
+
+#### Methods
+
+##### execute()
+
+> **execute**(`input`): `Promise`\<`CandidateExecutionEvidence`\>
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L91)
+
+###### Parameters
+
+###### input
+
+[`SandboxApprovedCandidateExecution`](#sandboxapprovedcandidateexecution)
+
+###### Returns
+
+`Promise`\<`CandidateExecutionEvidence`\>
+
+***
+
 ### AppliedIntelligence
 
 Defined in: [intelligence/with-intelligence.ts:53](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/with-intelligence.ts#L53)
@@ -2884,7 +2816,7 @@ Defined in: [intelligence/with-intelligence.ts:83](https://github.com/tangle-net
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:227](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L227)
+Defined in: [intelligence/index.ts:242](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L242)
 
 Stable project id — the tenant dimension every trace is tagged with.
 
@@ -2896,7 +2828,7 @@ Stable project id — the tenant dimension every trace is tagged with.
 
 > `optional` **apiKey?**: `string`
 
-Defined in: [intelligence/index.ts:229](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L229)
+Defined in: [intelligence/index.ts:244](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L244)
 
 Bearer key for the Intelligence ingest. Reads `TANGLE_API_KEY` when omitted.
 
@@ -2908,7 +2840,7 @@ Bearer key for the Intelligence ingest. Reads `TANGLE_API_KEY` when omitted.
 
 > `optional` **effort?**: [`EffortTier`](#efforttier) \| \{ `tier`: [`EffortTier`](#efforttier); `overrides?`: `Partial`\<[`EffortSettings`](#effortsettings)\>; \}
 
-Defined in: [intelligence/index.ts:231](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L231)
+Defined in: [intelligence/index.ts:246](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L246)
 
 Effort tier (default `'standard'`) plus optional per-field overrides.
 
@@ -2920,7 +2852,7 @@ Effort tier (default `'standard'`) plus optional per-field overrides.
 
 > `optional` **baseUrl?**: `string`
 
-Defined in: [intelligence/index.ts:239](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L239)
+Defined in: [intelligence/index.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L254)
 
 The ONE Tangle Intelligence base URL — both the send (OTLP `/v1/otlp`) and
 receive (`/v1/profiles/:target/composed`) paths derive from it. Reads
@@ -2936,7 +2868,7 @@ key the ingest requires); absent a key, export is a no-op.
 
 > `optional` **redact?**: `false` \| [`Redactor`](#redactor)
 
-Defined in: [intelligence/index.ts:245](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L245)
+Defined in: [intelligence/index.ts:260](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L260)
 
 Redaction hook run over every exported input/output. A function replaces
 the default scrubber; `false` opts out entirely (raw fidelity, caller has
@@ -2950,7 +2882,7 @@ sanitized upstream); omitted ⇒ the built-in `defaultRedactor`.
 
 > `optional` **surfaces?**: `string`[]
 
-Defined in: [intelligence/index.ts:247](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L247)
+Defined in: [intelligence/index.ts:262](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L262)
 
 Mutable surfaces a later PR mode would edit. Recorded for `doctor()` only.
 
@@ -2962,7 +2894,7 @@ Mutable surfaces a later PR mode would edit. Recorded for `doctor()` only.
 
 > `optional` **checks?**: `string`[]
 
-Defined in: [intelligence/index.ts:249](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L249)
+Defined in: [intelligence/index.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L264)
 
 Verification checks a later PR mode would gate on. Recorded for `doctor()` only.
 
@@ -2974,7 +2906,7 @@ Verification checks a later PR mode would gate on. Recorded for `doctor()` only.
 
 > `optional` **repo?**: [`RepoConfig`](#repoconfig)
 
-Defined in: [intelligence/index.ts:251](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L251)
+Defined in: [intelligence/index.ts:266](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L266)
 
 Repo access a later PR mode would need. Recorded for `doctor()` only.
 
@@ -2986,7 +2918,7 @@ Repo access a later PR mode would need. Recorded for `doctor()` only.
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [intelligence/index.ts:253](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L253)
+Defined in: [intelligence/index.ts:268](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L268)
 
 Full canonical profile used for this agent. Exported redacted with a stable hash.
 
@@ -2998,7 +2930,7 @@ Full canonical profile used for this agent. Exported redacted with a stable hash
 
 > `optional` **commitSha?**: `string`
 
-Defined in: [intelligence/index.ts:255](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L255)
+Defined in: [intelligence/index.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L270)
 
 Commit that produced the running agent, when known.
 
@@ -3010,7 +2942,7 @@ Commit that produced the running agent, when known.
 
 > `optional` **runtimeTelemetry?**: [`RuntimeTelemetryOptions`](index.md#runtimetelemetryoptions)
 
-Defined in: [intelligence/index.ts:257](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L257)
+Defined in: [intelligence/index.ts:272](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L272)
 
 Runtime-event payload policy. Tool inputs/results remain off unless explicitly enabled.
 
@@ -3022,7 +2954,7 @@ Runtime-event payload policy. Tool inputs/results remain off unless explicitly e
 
 > `optional` **payloadAttributes?**: `"metadata"` \| `"full"`
 
-Defined in: [intelligence/index.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L264)
+Defined in: [intelligence/index.ts:279](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L279)
 
 Payloads are metadata-only by default: the run span carries a stable hash
 and UTF-8 byte count, but not the redacted content. Set `full` only when
@@ -3222,29 +3154,11 @@ Per-field overrides applied on top of a tier preset. Any subset of the
 
 ***
 
-### AgentImprovementEvaluation
-
-> **AgentImprovementEvaluation**\<`TScenario`, `TArtifact`\> = `Pick`\<`SelfImproveResult`\<`TScenario`, `TArtifact`\>, `"baseline"` \| `"winner"` \| `"lift"` \| `"diff"` \| `"provenance"` \| `"gateDecision"` \| `"generationsExplored"` \| `"durationMs"` \| `"totalCostUsd"` \| `"insight"` \| `"power"`\>
-
-Defined in: [intelligence/improvement-cycle.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L47)
-
-#### Type Parameters
-
-##### TScenario
-
-`TScenario` *extends* `Scenario`
-
-##### TArtifact
-
-`TArtifact`
-
-***
-
-### AgentImprovementReviewDecision
+### ExecuteApprovedAgentCandidateResult
 
-> **AgentImprovementReviewDecision** = `"approve"` \| `"reject"` \| `"request-changes"`
+> **ExecuteApprovedAgentCandidateResult** = \{ `finalization`: `Extract`\<[`AgentCandidateRunFinalization`](index.md#agentcandidaterunfinalization), \{ `succeeded`: `false`; \}\>; `evidence?`: `never`; \} \| \{ `finalization`: `Extract`\<[`AgentCandidateRunFinalization`](index.md#agentcandidaterunfinalization), \{ `succeeded`: `true`; \}\>; `evidence`: `CandidateExecutionEvidence`; \}
 
-Defined in: [intelligence/improvement-cycle.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L81)
+Defined in: [intelligence/improvement-cycle.ts:147](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L147)
 
 ***
 
@@ -3252,7 +3166,7 @@ Defined in: [intelligence/improvement-cycle.ts:81](https://github.com/tangle-net
 
 > **UsageClass** = `"inference"` \| `"intelligence"`
 
-Defined in: [intelligence/index.ts:135](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L135)
+Defined in: [intelligence/index.ts:150](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L150)
 
 Usage class for billing. Base-stream tokens bill `'inference'`; every
  intelligence spawn (analyst, corpus, loop) bills `'intelligence'`. The
@@ -3375,8 +3289,42 @@ The default tier when a client declares no effort. `'standard'` turns
  intelligence on with sensible knobs; opt down to `'off'`/`'eco'` or up to
  `'thorough'`/`'max'`.
 
+***
+
+### sandboxApprovedCandidateExecutionSupport
+
+> `const` **sandboxApprovedCandidateExecutionSupport**: `Readonly`\<\{ `outcomes`: readonly \[`"output"`\]; `outputMediaTypes`: readonly \[`"text/*"`, `"application/json"`, `"*+json"`\]; `code`: readonly \[`"disabled"`\]; `memory`: readonly \[`"disabled"`\]; `knowledge`: `false`; `profile`: `Readonly`\<\{ `mcpTransports`: readonly \[`"stdio"`\]; `remoteMcp`: `false`; `tools`: `false`; `permissions`: `false`; `modes`: `false`; `confidential`: `false`; \}\>; `isolation`: `Readonly`\<\{ `freshSandbox`: `true`; `exactProcess`: `true`; `egress`: readonly \[`"blocked"`, `"strict"`\]; \}\>; \}\>
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L46)
+
+Declares the exact candidate surfaces the sandbox executor can run.
+
 ## Functions
 
+### parseAgentCandidateProfileActivation()
+
+> **parseAgentCandidateProfileActivation**(`input`, `expectedProfilePlanDigest?`): `AgentCandidateProfileActivation`
+
+Defined in: [candidate-execution/profile.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L91)
+
+Parse and check every native file hash plus both canonical document digests.
+
+#### Parameters
+
+##### input
+
+`unknown`
+
+##### expectedProfilePlanDigest?
+
+`` `sha256:${string}` ``
+
+#### Returns
+
+`AgentCandidateProfileActivation`
+
+***
+
 ### manifestFromProfile()
 
 > **manifestFromProfile**(`profile`): [`CapabilityManifest`](#capabilitymanifest)
@@ -3610,9 +3558,9 @@ compile to `withAnalyst: true`, the tier's `fanout`, and `withLoops: true`.
 
 > **proposeAgentImprovement**\<`TScenario`, `TArtifact`\>(`options`): `Promise`\<[`ProposeAgentImprovementResult`](#proposeagentimprovementresult)\<`TScenario`, `TArtifact`\>\>
 
-Defined in: [intelligence/improvement-cycle.ts:172](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L172)
+Defined in: [intelligence/improvement-cycle.ts:165](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L165)
 
-Analyze one run and produce one measured, review-only improvement proposal.
+Analyze one run and freeze its measured winner into one exact proposal.
 
 #### Type Parameters
 
@@ -3638,41 +3586,31 @@ Analyze one run and produce one measured, review-only improvement proposal.
 
 ### createAgentImprovementProposal()
 
-> **createAgentImprovementProposal**\<`TScenario`, `TArtifact`\>(`options`): [`AgentImprovementProposal`](#agentimprovementproposal)\<`TScenario`, `TArtifact`\>
+> **createAgentImprovementProposal**(`options`): `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:216](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L216)
+Defined in: [intelligence/improvement-cycle.ts:213](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L213)
 
 Freeze an already-measured improvement into the one reviewable proposal
 contract. Products that run analysis or evaluation in separate workers use
 this constructor instead of rerunning either phase or rebuilding digests.
 
-#### Type Parameters
-
-##### TScenario
-
-`TScenario` *extends* `Scenario`
-
-##### TArtifact
-
-`TArtifact`
-
 #### Parameters
 
 ##### options
 
-[`CreateAgentImprovementProposalOptions`](#createagentimprovementproposaloptions)\<`TScenario`, `TArtifact`\>
+[`CreateAgentImprovementProposalOptions`](#createagentimprovementproposaloptions)
 
 #### Returns
 
-[`AgentImprovementProposal`](#agentimprovementproposal)\<`TScenario`, `TArtifact`\>
+`AgentImprovementProposal`
 
 ***
 
 ### reviewAgentImprovementProposal()
 
-> **reviewAgentImprovementProposal**(`inputProposal`, `input`): [`AgentImprovementReview`](#agentimprovementreview)
+> **reviewAgentImprovementProposal**(`inputProposal`, `input`): `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:266](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L266)
+Defined in: [intelligence/improvement-cycle.ts:252](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L252)
 
 Persist an approve/reject/change-request decision bound to one exact proposal.
 
@@ -3680,7 +3618,7 @@ Persist an approve/reject/change-request decision bound to one exact proposal.
 
 ##### inputProposal
 
-[`AgentImprovementProposal`](#agentimprovementproposal)
+`AgentImprovementProposal`
 
 ##### input
 
@@ -3688,7 +3626,7 @@ Persist an approve/reject/change-request decision bound to one exact proposal.
 
 #### Returns
 
-[`AgentImprovementReview`](#agentimprovementreview)
+`AgentImprovementReview`
 
 ***
 
@@ -3696,7 +3634,7 @@ Persist an approve/reject/change-request decision bound to one exact proposal.
 
 > **executeApprovedAgentCandidate**(`options`): `Promise`\<[`ExecuteApprovedAgentCandidateResult`](#executeapprovedagentcandidateresult)\>
 
-Defined in: [intelligence/improvement-cycle.ts:296](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L296)
+Defined in: [intelligence/improvement-cycle.ts:281](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L281)
 
 Verify, materialize, run, grade, and receipt only the exact approved bundle.
 
@@ -3712,11 +3650,35 @@ Verify, materialize, run, grade, and receipt only the exact approved bundle.
 
 ***
 
+### verifyCandidateExecutionEvidence()
+
+> **verifyCandidateExecutionEvidence**(`input`, `options`): `CandidateExecutionEvidence`[]
+
+Defined in: [intelligence/improvement-cycle.ts:331](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L331)
+
+Verify approval, receipts, uniqueness, and the exact native profile files executed.
+
+#### Parameters
+
+##### input
+
+`unknown`
+
+##### options
+
+[`VerifyCandidateExecutionEvidenceOptions`](#verifycandidateexecutionevidenceoptions)
+
+#### Returns
+
+`CandidateExecutionEvidence`[]
+
+***
+
 ### verifyAgentImprovementProposal()
 
-> **verifyAgentImprovementProposal**(`input`): [`AgentImprovementProposal`](#agentimprovementproposal)
+> **verifyAgentImprovementProposal**(`input`): `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:338](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L338)
+Defined in: [intelligence/improvement-cycle.ts:434](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L434)
 
 Validate a proposal's schema, profile, sealed bundle, and canonical digest.
 
@@ -3728,15 +3690,15 @@ Validate a proposal's schema, profile, sealed bundle, and canonical digest.
 
 #### Returns
 
-[`AgentImprovementProposal`](#agentimprovementproposal)
+`AgentImprovementProposal`
 
 ***
 
 ### verifyAgentImprovementReview()
 
-> **verifyAgentImprovementReview**(`input`): [`AgentImprovementReview`](#agentimprovementreview)
+> **verifyAgentImprovementReview**(`input`): `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:455](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L455)
+Defined in: [intelligence/improvement-cycle.ts:668](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L668)
 
 Validate a review's decision fields and canonical digest.
 
@@ -3748,7 +3710,37 @@ Validate a review's decision fields and canonical digest.
 
 #### Returns
 
-[`AgentImprovementReview`](#agentimprovementreview)
+`AgentImprovementReview`
+
+***
+
+### createAgentImprovementMeasuredComparison()
+
+> **createAgentImprovementMeasuredComparison**\<`TScenario`, `TArtifact`\>(`options`): `AgentImprovementMeasuredComparison`
+
+Defined in: [intelligence/improvement-cycle.ts:674](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L674)
+
+Convert agent-eval's paired result into the portable Interface comparison.
+
+#### Type Parameters
+
+##### TScenario
+
+`TScenario` *extends* `Scenario`
+
+##### TArtifact
+
+`TArtifact`
+
+#### Parameters
+
+##### options
+
+`CreateAgentImprovementMeasuredComparisonOptions`\<`TScenario`, `TArtifact`\>
+
+#### Returns
+
+`AgentImprovementMeasuredComparison`
 
 ***
 
@@ -3756,7 +3748,7 @@ Validate a review's decision fields and canonical digest.
 
 > **createIntelligenceClient**(`config`): [`IntelligenceClient`](#intelligenceclient)
 
-Defined in: [intelligence/index.ts:454](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L454)
+Defined in: [intelligence/index.ts:469](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L469)
 
 Create an Observe-mode Intelligence client. Resolves effort, the base URL, and
 the redactor up front; the exporter is built lazily and is `undefined` when no
@@ -3890,6 +3882,26 @@ Lower a plane `CertifiedProfile` straight into a `ResolvedSurface` via
 
 ***
 
+### createSandboxApprovedCandidateExecutor()
+
+> **createSandboxApprovedCandidateExecutor**(`options`): [`SandboxApprovedCandidateExecutor`](#sandboxapprovedcandidateexecutor)
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L95)
+
+Compose approved-candidate execution directly onto fresh Tangle sandboxes.
+
+#### Parameters
+
+##### options
+
+[`CreateSandboxApprovedCandidateExecutorOptions`](#createsandboxapprovedcandidateexecutoroptions)
+
+#### Returns
+
+[`SandboxApprovedCandidateExecutor`](#sandboxapprovedcandidateexecutor)
+
+***
+
 ### withIntelligence()
 
 > **withIntelligence**\<`I`, `O`\>(`agent`, `config`): [`IntelligenceWrapped`](#intelligencewrapped)\<`I`, `O`\>
diff --git a/docs/api/knowledge.md b/docs/api/knowledge.md
index 2deb753d..cfef26aa 100644
--- a/docs/api/knowledge.md
+++ b/docs/api/knowledge.md
@@ -14,6 +14,12 @@ Re-exports [AgentKnowledgeReadinessCheckOptions](index.md#agentknowledgereadines
 
 ***
 
+### ApprovedKnowledgeImprovementCandidate
+
+Re-exports [ApprovedKnowledgeImprovementCandidate](index.md#approvedknowledgeimprovementcandidate)
+
+***
+
 ### createAgentKnowledgeReadinessCheck
 
 Re-exports [createAgentKnowledgeReadinessCheck](index.md#createagentknowledgereadinesscheck)
diff --git a/docs/api/primitive-catalog.md b/docs/api/primitive-catalog.md
index d9263446..2a006e66 100644
--- a/docs/api/primitive-catalog.md
+++ b/docs/api/primitive-catalog.md
@@ -15,7 +15,7 @@ Every subpath this package declares in `package.json` `exports`. Reach for these
 
 ### Root — task lifecycle, conversation, RSI verbs, observability
 
-Import from `@tangle-network/agent-runtime` — 337 exports.
+Import from `@tangle-network/agent-runtime` — 336 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -91,7 +91,7 @@ Import from `@tangle-network/agent-runtime` — 337 exports.
 | `runConversation` | function | Conversation orchestrator. Drives N participants in turn through their own |
 | `runConversationStream` | function | Streaming conversation orchestrator: drives N participants in turn through their own backends, enforcing `maxTurns` / `maxCreditsCents` / `haltOn`, yielding per-event stream markers. |
 | `runDelegatedLoop` | function | Dispatch a configured loop by mode. Fails loud (throws `ConfigError`) when no |
-| `runKnowledgeImprovementJob` | function | Run the full KB improvement job: candidate workspace, runtime supervisor update, readiness check, and promotion. |
+| `runKnowledgeImprovementJob` | function | Produce a frozen KB candidate, and promote it only when an exact signed review is supplied. |
 | `runLoopRunnerCli` | function | Pure CLI core (no process / argv / IO) so it's unit-testable: validate the |
 | `runPersonaConversation` | function | Run one worker profile against one persona as a multi-round conversation. |
 | `runPersonaDispatch` | function | Wrap {@link runPersonaConversation} as a `ProfileDispatchFn` for |
@@ -112,6 +112,7 @@ Import from `@tangle-network/agent-runtime` — 337 exports.
 | `validateChatModelId` | function | Validate a caller-supplied chat-model id. Rejects non-strings, malformed |
 | `verifyAgentCandidateBundle` | function | Verifies every digest, resource, workspace, and Git object in a candidate bundle. |
 | `worktreeLoopRunner` | function | `code` mode on the GENERIC recursive path: author one `AgentProfile` per harness, run them as a |
+| `AGENT_CANDIDATE_EXECUTION_SUPPORT` | const | Surfaces admitted by Runtime's verifier before an environment adapter is selected. |
 | `CANDIDATE_TRACE_ENV` | const | Environment keys used to propagate immutable candidate trace identity. |
 | `CANDIDATE_TRACE_TAGS` | const | Protected trace tags that bind a run to one prepared candidate execution. |
 | `DEFAULT_MAX_DEPTH` | const | Hard cap on chained gateway hops; refused beyond this. Default keeps recursion bounded. |
@@ -147,17 +148,15 @@ Import from `@tangle-network/agent-runtime` — 337 exports.
 | `AgentCandidateExecutionCleanupHandles` | interface | Non-secret identities a trusted recovery worker needs to close an abandoned attempt. |
 | `AgentCandidateExecutionLease` | interface | Secret capability required to finish the acquired attempt. |
 | `AgentCandidateExecutionRecoveryEvidence` | interface | Trusted, independently observed closure facts for one expired winning lease. |
-| `AgentCandidateExecutionUsage` | interface | Exact fixed-point usage proven by the closed evaluator model ledger. |
-| `AgentCandidateExecutorFinalCapture` | interface | Idempotent executor result after process death and trace drain. |
+| `AgentCandidateExecutorFinalCapture` | interface | Replayable evaluator result captured only after process death and trace drain. |
 | `AgentCandidateExecutorMemoryCapture` | interface | Raw isolated-memory capture made only after access has been revoked. |
 | `AgentCandidateExecutorPort` | interface | Executes one prepared request inside an evaluator-owned isolation boundary. |
 | `AgentCandidateExecutorRequest` | interface | One detached request passed to the trusted environment-specific executor. |
 | `AgentCandidateExecutorStopRequest` | interface | Opaque process identity used for termination without re-exposing launch credentials. |
-| `AgentCandidateExecutorTaskOutcomeCapture` | interface | Raw evaluator capture made only after the candidate process is dead. |
 | `AgentCandidateModelGrantClient` | interface | Narrow transport contract for a service that owns scoped model credentials |
 | `AgentCandidateOutputArtifactPort` | interface | Durable content-addressed evidence store controlled only by the evaluator. |
-| `AgentCandidateProtectedModelCall` | interface | One evaluator-gateway call in the final, revoked model-access ledger. |
 | `AgentCandidateRepositoryPort` | interface | Resolves a declared GitHub repository to an already-present local Git object store. |
+| `AgentCandidateTaskExecution` | interface | One signed benchmark task and the exact result shape its executor must capture. |
 | `AgentCandidateWorkspacePort` | interface | Materializes an already-verified workspace archive. |
 | `BackendErrorDetail` | interface | Typed transport / backend failure detail. Carried on `backend_error` and |
 | `BuildAgentCandidateBundleInput` | interface | Complete measured surfaces and execution policy compiled into one candidate bundle. |
@@ -181,7 +180,6 @@ Import from `@tangle-network/agent-runtime` — 337 exports.
 | `SqlAdapter` | interface | Minimal SQL driver shape. Implementations forward to whichever client the |
 | `ToolLoopAssistantToolCall` | interface | One OpenAI-shaped tool-call entry carried on an assistant message. |
 | `ToolLoopCall` | interface | Bounded turn-level tool-dispatch loop. |
-| `VerifiedAgentCandidateTaskOutcome` | interface | Branded task outcome that has survived independent patch and tree verification. |
 | `VerifyResult` | interface | Outcome of verifying a candidate worktree. `feedback` (compiler errors, |
 | `AgentBackendKind` | type | The transport a chat backend runs on. |
 | `AgentCandidateBundleInput` | type | Exact candidate wire shape before the runtime computes its canonical digest. |
@@ -194,6 +192,7 @@ Import from `@tangle-network/agent-runtime` — 337 exports.
 | `AgentCandidateExecutionStageResult` | type | Result of durably staging the one immutable terminal outbox entry. |
 | `AgentCandidateExecutionTerminalRecord` | type | Durable terminal record for one acquired execution attempt. |
 | `AgentCandidateExecutionTerminalResult` | type | Evaluator-owned terminal facts staged durably before the terminal CAS. |
+| `AgentCandidateExecutorTaskOutcomeCapture` | type | Raw evaluator capture made only after the candidate process is dead. |
 | `AgentCandidateModelGrantReservation` | type | Secret-free response from the service's reservation endpoint. |
 | `AgentCandidateModelLimits` | type | Limits mechanically enforced by the evaluator-owned model gateway. |
 | `AgentCandidateProfileSource` | type | A complete profile that can be frozen without losing behavior. |
@@ -209,9 +208,10 @@ Import from `@tangle-network/agent-runtime` — 337 exports.
 | `ToolCallOutcome` | type | Outcome of one tool dispatch — structurally compatible with a hub/integration |
 | `ToolLoopMessage` | type | A message in the running conversation the loop sends to `streamTurn`. |
 | `ToolLoopStopReason` | type | Why the loop stopped. `completed` = model finished naturally; `stuck-loop` = |
+| `VerifiedAgentCandidateTaskOutcome` | type | Branded task outcome that has survived independent evaluator verification. |
 | `Verifier` | type | Verifies the edited worktree. Sync or async; throws only on a setup fault |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentAdapter`, `AgentBackendContext`, `AgentBackendInput`, `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorProfileFile`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateTaskExecution`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `AgentExecutionBackend`, `AgenticGeneratorOptions`, `AgentKnowledgeProvider`, `AgentKnowledgeReadinessCheckOptions`, `AgentProfileDiffProposal`, `AgentTaskContext`, `AgentTaskRunResult`, `AgentTaskSpec`, `BackendCallPolicy`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `ChatTurnHooks`, `ChatTurnResult`, `ControlBudget`, `ControlEvalResult`, `ControlRunResult`, `ControlStep`, `Conversation`, `ConversationDriveState`, `ConversationJournal`, `ConversationParticipant`, `ConversationPolicy`, `ConversationResult`, `ConversationTurn`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `D1StmtLike`, `DataAcquisitionPlan`, `DelegatedLoopResult`, `DisposePreparedAgentCandidateOptions`, `EvalRunEvent`, `EvalRunGeneration`, `EvalRunsExportConfig`, `EvalRunsExportResult`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `HaltContext`, `HaltSignal`, `ImproveCodeOptions`, `ImproveMemoryOptions`, `ImprovementDriverOptions`, `ImproveResult`, `ImproveSkillsOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `KnowledgeReadinessReport`, `KnowledgeRequirement`, `LoopRunnerCliArgs`, `LoopRunnerCliResult`, `ManagedImprovementDriver`, `OtelAttribute`, `OtelExporter`, `OtelSpan`, `PersonaConversationResult`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `ProfileDiffProposerOptions`, `RecoverExpiredAgentCandidateOptions`, `ResearchLoopResult`, `ResearchLoopRunnerOptions`, `ResolveAgentBackendOptions`, `ResolvedAgentCandidateContainer`, `ResolvedChatModel`, `RunChatTurnInput`, `RunConversationOptions`, `RunDelegatedLoopOptions`, `RunKnowledgeImprovementJobOptions`, `RunPersonaConfig`, `RunPersonaConversationOptions`, `RuntimeDecisionEvidenceRef`, `RuntimeDecisionPoint`, `RuntimeEventCollector`, `RuntimeEventOtelOptions`, `RuntimeHookContext`, `RuntimeHookErrorContext`, `RuntimeHookEvent`, `RuntimeRunHandle`, `RuntimeRunPersistenceAdapter`, `RuntimeRunRow`, `RuntimeSessionStore`, `RuntimeStreamEventCollector`, `RuntimeTelemetryOptions`, `RunToolLoopOptions`, `SanitizedKnowledgeReadinessReport`, `StreamToolLoopOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `ToolLoopResult`, `VerifiedAgentCandidate`, `VetoedFact`, `WorktreeLoopRunnerOptions`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`, `AgentRuntimeEvent`, `AgentRuntimeEventSink`, `AgentTaskStatus`, `AuthSource`, `ControlDecision`, `ConversationStreamEvent`, `DelegatedLoopMode`, `DelegatedLoopRegistry`, `DelegatedLoopRunner`, `ForwardHeaderName`, `HaltPredicate`, `HaltReason`, `ImproveOptions`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `ProfileDiffProposerContext`, `RuntimeDecisionKind`, `RuntimeHookTarget`, `RuntimeStreamEvent`, `StreamToolLoopYield`, `SupervisedKnowledgeUpdater`, `ToolLoopEvent`, `TurnOrder`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentAdapter`, `AgentBackendContext`, `AgentBackendInput`, `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `AgentExecutionBackend`, `AgenticGeneratorOptions`, `AgentKnowledgeProvider`, `AgentKnowledgeReadinessCheckOptions`, `AgentProfileDiffProposal`, `AgentTaskContext`, `AgentTaskRunResult`, `AgentTaskSpec`, `ApprovedKnowledgeImprovementCandidate`, `BackendCallPolicy`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `ChatTurnHooks`, `ChatTurnResult`, `ControlBudget`, `ControlEvalResult`, `ControlRunResult`, `ControlStep`, `Conversation`, `ConversationDriveState`, `ConversationJournal`, `ConversationParticipant`, `ConversationPolicy`, `ConversationResult`, `ConversationTurn`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `D1StmtLike`, `DataAcquisitionPlan`, `DelegatedLoopResult`, `DisposePreparedAgentCandidateOptions`, `EvalRunEvent`, `EvalRunGeneration`, `EvalRunsExportConfig`, `EvalRunsExportResult`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `HaltContext`, `HaltSignal`, `ImproveCodeOptions`, `ImproveMemoryOptions`, `ImprovementDriverOptions`, `ImproveResult`, `ImproveSkillsOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `KnowledgeReadinessReport`, `KnowledgeRequirement`, `LoopRunnerCliArgs`, `LoopRunnerCliResult`, `ManagedImprovementDriver`, `OtelAttribute`, `OtelExporter`, `OtelSpan`, `PersonaConversationResult`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `ProfileDiffProposerOptions`, `RecoverExpiredAgentCandidateOptions`, `ResearchLoopResult`, `ResearchLoopRunnerOptions`, `ResolveAgentBackendOptions`, `ResolvedAgentCandidateContainer`, `ResolvedChatModel`, `RunChatTurnInput`, `RunConversationOptions`, `RunDelegatedLoopOptions`, `RunKnowledgeImprovementJobOptions`, `RunPersonaConfig`, `RunPersonaConversationOptions`, `RuntimeDecisionEvidenceRef`, `RuntimeDecisionPoint`, `RuntimeEventCollector`, `RuntimeEventOtelOptions`, `RuntimeHookContext`, `RuntimeHookErrorContext`, `RuntimeHookEvent`, `RuntimeRunHandle`, `RuntimeRunPersistenceAdapter`, `RuntimeRunRow`, `RuntimeSessionStore`, `RuntimeStreamEventCollector`, `RuntimeTelemetryOptions`, `RunToolLoopOptions`, `SanitizedKnowledgeReadinessReport`, `StreamToolLoopOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `ToolLoopResult`, `VerifiedAgentCandidate`, `VetoedFact`, `WorktreeLoopRunnerOptions`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`, `AgentRuntimeEvent`, `AgentRuntimeEventSink`, `AgentTaskStatus`, `AuthSource`, `ControlDecision`, `ConversationStreamEvent`, `DelegatedLoopMode`, `DelegatedLoopRegistry`, `DelegatedLoopRunner`, `ForwardHeaderName`, `HaltPredicate`, `HaltReason`, `ImproveOptions`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `ProfileDiffProposerContext`, `RuntimeDecisionKind`, `RuntimeHookTarget`, `RuntimeStreamEvent`, `StreamToolLoopYield`, `SupervisedKnowledgeUpdater`, `ToolLoopEvent`, `TurnOrder`.
 
 ### Vertical agent — manifest + improvement adapter
 
@@ -257,7 +257,7 @@ Import from `@tangle-network/agent-runtime/agent` — 48 exports.
 
 ### Intelligence SDK — Observe + provable-OFF billing
 
-Import from `@tangle-network/agent-runtime/intelligence` — 85 exports.
+Import from `@tangle-network/agent-runtime/intelligence` — 95 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -265,15 +265,18 @@ Import from `@tangle-network/agent-runtime/intelligence` — 85 exports.
 | `composeCertifiedProfile` | function | Compose a certified profile into a uniform `ResolvedSurface`. Additive over |
 | `composeCertifiedProfileFromWire` | function | Lower a plane `CertifiedProfile` straight into a `ResolvedSurface` via |
 | `composeCertifiedPrompt` | function | Fold the certified prompt surface (and any certified prompt-folding artifacts: |
+| `createAgentImprovementMeasuredComparison` | function | Convert agent-eval's paired result into the portable Interface comparison. |
 | `createAgentImprovementProposal` | function | Freeze an already-measured improvement into the one reviewable proposal |
 | `createCertifiedPromptSource` | function | Create the cached certified-prompt source — the ONE module-scope-cache + |
 | `createIntelligenceClient` | function | Create an Observe-mode Intelligence client. Resolves effort, the base URL, and |
+| `createSandboxApprovedCandidateExecutor` | function | Compose approved-candidate execution directly onto fresh Tangle sandboxes. |
 | `defaultRedactor` | function | The built-in redactor. Walks objects and arrays; replaces values under |
 | `executeApprovedAgentCandidate` | function | Verify, materialize, run, grade, and receipt only the exact approved bundle. |
 | `isIntelligenceOff` | function | True when these settings admit NO intelligence spawn — the passthrough |
 | `manifestFromProfile` | function | Lower the EXISTING plane wire (`CertifiedProfile`) into a `CapabilityManifest`. |
 | `normalizeCertifiedProfile` | function | Deserialize the composed-endpoint response into a `CertifiedProfile`. The |
-| `proposeAgentImprovement` | function | Analyze one run and produce one measured, review-only improvement proposal. |
+| `parseAgentCandidateProfileActivation` | function | Parse and check every native file hash plus both canonical document digests. |
+| `proposeAgentImprovement` | function | Analyze one run and freeze its measured winner into one exact proposal. |
 | `pullCertified` | function | Pull the certified composed profile for a target. Fail-closed: a network |
 | `resolveEffort` | function | Compile a named tier (plus optional per-field overrides) into the flat |
 | `resolveIntelligenceBaseUrl` | function | Resolve the ONE Intelligence base URL — the single knob both the send and |
@@ -281,10 +284,16 @@ Import from `@tangle-network/agent-runtime/intelligence` — 85 exports.
 | `reviewAgentImprovementProposal` | function | Persist an approve/reject/change-request decision bound to one exact proposal. |
 | `verifyAgentImprovementProposal` | function | Validate a proposal's schema, profile, sealed bundle, and canonical digest. |
 | `verifyAgentImprovementReview` | function | Validate a review's decision fields and canonical digest. |
+| `verifyCandidateExecutionEvidence` | function | Verify approval, receipts, uniqueness, and the exact native profile files executed. |
 | `withIntelligence` | function | Wrap an agent so it (a) RECEIVES the tenant's certified profile — the prompt |
 | `defaultEffortTier` | const | The default tier when a client declares no effort. `'standard'` turns |
+| `sandboxApprovedCandidateExecutionSupport` | const | Declares the exact candidate surfaces the sandbox executor can run. |
 | `CapabilityNotAdmittedError` | class | A binding kind whose resolver case is typed but not yet admitted (rag-index, |
+| `AgentCandidateProfileActivation` | interface | Exact native profile files and the canonical plan that activated them. |
+| `AgentImprovementMeasuredComparison` | interface | Portable paired held-out comparison produced by an evaluation package. |
+| `AgentImprovementReview` | interface | Human or tenant-policy decision bound to one exact proposal. |
 | `AppliedIntelligence` | interface | What the hook hands the agent each run. Additive over the prompt-only |
+| `CandidateExecutionEvidence` | interface | Successful post-approval execution, carrying the exact Runtime receipt. |
 | `CapabilityManifest` | interface | The strict generalization of `CertifiedProfile`. `promptSurface` is kept |
 | `CertifiedArtifact` | interface | A promoted, certified artifact (one entry in the composed profile). |
 | `CertifiedCapability` | interface | One certified unit of agent power. |
@@ -335,7 +344,7 @@ Import from `@tangle-network/agent-runtime/intelligence` — 85 exports.
 | `Redactor` | type | A redactor maps an arbitrary trace value to a safe-to-export value. Pure; |
 | `UsageClass` | type | Usage class for billing. Base-stream tokens bill `'inference'`; every |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentImprovementProposal`, `AgentImprovementReview`, `CandidateExecutionEvidence`, `CreateAgentImprovementProposalOptions`, `ExecuteApprovedAgentCandidateOptions`, `ExecuteApprovedAgentCandidateResult`, `ProposeAgentImprovementOptions`, `ProposeAgentImprovementResult`, `PullCertifiedOptions`, `ReviewAgentImprovementInput`, `AgentImprovementEvaluation`, `AgentImprovementReviewDecision`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentImprovementProposal`, `CreateAgentImprovementProposalOptions`, `CreateSandboxApprovedCandidateExecutorOptions`, `ExecuteApprovedAgentCandidateOptions`, `ProposeAgentImprovementOptions`, `ProposeAgentImprovementResult`, `PullCertifiedOptions`, `ReviewAgentImprovementInput`, `SandboxApprovedCandidateExecution`, `SandboxApprovedCandidateExecutor`, `VerifyCandidateExecutionEvidenceOptions`, `AgentImprovementReviewDecision`, `ExecuteApprovedAgentCandidateResult`.
 
 ### Recursive atom + loop kernel (alias of ./runtime)
 
@@ -793,7 +802,7 @@ Import from `@tangle-network/agent-runtime/lifecycle` — 59 exports.
 
 ### Knowledge orchestration — supervised KB updates
 
-Import from `@tangle-network/agent-runtime/knowledge` — 18 exports.
+Import from `@tangle-network/agent-runtime/knowledge` — 19 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -801,11 +810,11 @@ Import from `@tangle-network/agent-runtime/knowledge` — 18 exports.
 | `createSupervisedKnowledgeUpdater` | function | Create an `improveKnowledgeBase` update callback backed by runtime supervision. |
 | `formatSupervisedKnowledgeTask` | function | Format the supervisor task with the KB root, readiness requirements, current findings, and metadata. |
 | `knowledgeReadinessDeliverable` | function | Build the completion check a supervised KB update uses to stop only when the KB is ready. |
-| `runKnowledgeImprovementJob` | function | Run the full KB improvement job: candidate workspace, runtime supervisor update, readiness check, and promotion. |
+| `runKnowledgeImprovementJob` | function | Produce a frozen KB candidate, and promote it only when an exact signed review is supplied. |
 | `runSupervisedKnowledgeUpdate` | function | Run a runtime supervisor that updates one candidate knowledge base and stops on readiness. |
 | `RESEARCH_SUPERVISOR_SYSTEM_PROMPT` | const | Standing prompt for a supervisor that grows a shared knowledge base through spawned researchers. |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentKnowledgeReadinessCheckOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `RunKnowledgeImprovementJobOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `SupervisedKnowledgeUpdater`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentKnowledgeReadinessCheckOptions`, `ApprovedKnowledgeImprovementCandidate`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `RunKnowledgeImprovementJobOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `SupervisedKnowledgeUpdater`.
 
 ### Built-in agent profiles
 
@@ -870,7 +879,7 @@ Import from `@tangle-network/agent-runtime/platform` — 20 exports.
 
 ### Candidate execution — immutable prepare, run, grade, and receipt
 
-Import from `@tangle-network/agent-runtime/candidate-execution` — 95 exports.
+Import from `@tangle-network/agent-runtime/candidate-execution` — 93 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -890,6 +899,7 @@ Import from `@tangle-network/agent-runtime/candidate-execution` — 95 exports.
 | `recoverExpiredAgentCandidateExecution` | function | Close an expired crashed attempt from persisted non-secret handles, then record failure. |
 | `sealAgentCandidateBundle` | function | Validate and content-address a candidate bundle before it crosses an approval boundary. |
 | `verifyAgentCandidateBundle` | function | Verifies every digest, resource, workspace, and Git object in a candidate bundle. |
+| `AGENT_CANDIDATE_EXECUTION_SUPPORT` | const | Surfaces admitted by Runtime's verifier before an environment adapter is selected. |
 | `CANDIDATE_TRACE_ENV` | const | Environment keys used to propagate immutable candidate trace identity. |
 | `CANDIDATE_TRACE_TAGS` | const | Protected trace tags that bind a run to one prepared candidate execution. |
 | `FileAgentCandidateExecutionClaimStore` | class | Cross-process lifecycle implemented as fsynced, create-if-absent records. |
@@ -903,20 +913,17 @@ Import from `@tangle-network/agent-runtime/candidate-execution` — 95 exports.
 | `AgentCandidateExecutionCleanupHandles` | interface | Non-secret identities a trusted recovery worker needs to close an abandoned attempt. |
 | `AgentCandidateExecutionLease` | interface | Secret capability required to finish the acquired attempt. |
 | `AgentCandidateExecutionRecoveryEvidence` | interface | Trusted, independently observed closure facts for one expired winning lease. |
-| `AgentCandidateExecutionUsage` | interface | Exact fixed-point usage proven by the closed evaluator model ledger. |
-| `AgentCandidateExecutorFinalCapture` | interface | Idempotent executor result after process death and trace drain. |
+| `AgentCandidateExecutorFinalCapture` | interface | Replayable evaluator result captured only after process death and trace drain. |
 | `AgentCandidateExecutorMemoryCapture` | interface | Raw isolated-memory capture made only after access has been revoked. |
 | `AgentCandidateExecutorPort` | interface | Executes one prepared request inside an evaluator-owned isolation boundary. |
 | `AgentCandidateExecutorRequest` | interface | One detached request passed to the trusted environment-specific executor. |
 | `AgentCandidateExecutorStopRequest` | interface | Opaque process identity used for termination without re-exposing launch credentials. |
-| `AgentCandidateExecutorTaskOutcomeCapture` | interface | Raw evaluator capture made only after the candidate process is dead. |
 | `AgentCandidateModelGrantClient` | interface | Narrow transport contract for a service that owns scoped model credentials |
 | `AgentCandidateOutputArtifactPort` | interface | Durable content-addressed evidence store controlled only by the evaluator. |
-| `AgentCandidateProtectedModelCall` | interface | One evaluator-gateway call in the final, revoked model-access ledger. |
 | `AgentCandidateRepositoryPort` | interface | Resolves a declared GitHub repository to an already-present local Git object store. |
+| `AgentCandidateTaskExecution` | interface | One signed benchmark task and the exact result shape its executor must capture. |
 | `AgentCandidateWorkspacePort` | interface | Materializes an already-verified workspace archive. |
 | `BuildAgentCandidateBundleInput` | interface | Complete measured surfaces and execution policy compiled into one candidate bundle. |
-| `VerifiedAgentCandidateTaskOutcome` | interface | Branded task outcome that has survived independent patch and tree verification. |
 | `AgentCandidateBundleInput` | type | Exact candidate wire shape before the runtime computes its canonical digest. |
 | `AgentCandidateCodeSource` | type | Explicit control/no-op code or one finalized CodeSurface whose bytes must still verify. |
 | `AgentCandidateExecutionClaimResult` | type | Result of atomically claiming one execution attempt. |
@@ -927,11 +934,13 @@ Import from `@tangle-network/agent-runtime/candidate-execution` — 95 exports.
 | `AgentCandidateExecutionStageResult` | type | Result of durably staging the one immutable terminal outbox entry. |
 | `AgentCandidateExecutionTerminalRecord` | type | Durable terminal record for one acquired execution attempt. |
 | `AgentCandidateExecutionTerminalResult` | type | Evaluator-owned terminal facts staged durably before the terminal CAS. |
+| `AgentCandidateExecutorTaskOutcomeCapture` | type | Raw evaluator capture made only after the candidate process is dead. |
 | `AgentCandidateModelGrantReservation` | type | Secret-free response from the service's reservation endpoint. |
 | `AgentCandidateModelLimits` | type | Limits mechanically enforced by the evaluator-owned model gateway. |
 | `AgentCandidateProfileSource` | type | A complete profile that can be frozen without losing behavior. |
+| `VerifiedAgentCandidateTaskOutcome` | type | Branded task outcome that has survived independent evaluator verification. |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorProfileFile`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateTaskExecution`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `DisposePreparedAgentCandidateOptions`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `RecoverExpiredAgentCandidateOptions`, `ResolvedAgentCandidateContainer`, `VerifiedAgentCandidate`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `DisposePreparedAgentCandidateOptions`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `RecoverExpiredAgentCandidateOptions`, `ResolvedAgentCandidateContainer`, `VerifiedAgentCandidate`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`.
 
 ### MCP servers — delegate / coordination / detached-session
 
diff --git a/examples/coding-benchmark/dispatch.ts b/examples/coding-benchmark/dispatch.ts
index 62432978..30c7a258 100644
--- a/examples/coding-benchmark/dispatch.ts
+++ b/examples/coding-benchmark/dispatch.ts
@@ -40,13 +40,7 @@ import {
   sumSandboxUsage,
 } from '@tangle-network/agent-runtime/loops'
 import type { SandboxEvent } from '@tangle-network/sandbox'
-import {
-  type CheckBox,
-  gradeOnHiddenCriteria,
-  layerOutput,
-  type RunArtifact,
-  runChecks,
-} from './eval'
+import { gradeOnHiddenCriteria, layerOutput, type RunArtifact, runChecks } from './eval'
 import { harnessOf, type ToolPreset, withTools } from './profiles'
 import { type CodingScenario, checkCmds, routeCodingFields } from './scenarios'
 
@@ -152,7 +146,7 @@ export function codingDispatch(
         // these) steer the next round — the firewall keeps the held-out suite + rubric
         // out of the loop. `run.box` is a `SandboxInstance`; `CheckBox` is the minimal
         // `exec`(+optional `fs.write`) subset the checks use — a structural narrowing.
-        checks = await runChecks(run.box as CheckBox, scenario, cmds)
+        checks = await runChecks(run.box, scenario, cmds)
         if (checks.allPass) break // stop on worker-observable green only
       }
 
@@ -164,7 +158,7 @@ export function codingDispatch(
       // composite). A solution that hardcoded the visible examples fails the held-out inputs
       // it never saw — execution truth behind a substrate-enforced firewall, not a regex.
       const heldout = await gradeOnHiddenCriteria(
-        run.box as CheckBox,
+        run.box,
         scenario,
         cmds.heldout,
         { fields: routedFields, agentContext: agentContextParts.join('\n') },
diff --git a/examples/coding-benchmark/eval.ts b/examples/coding-benchmark/eval.ts
index 65c7866a..d5aa8e0c 100644
--- a/examples/coding-benchmark/eval.ts
+++ b/examples/coding-benchmark/eval.ts
@@ -94,7 +94,7 @@ export const heldoutPassRateOf = (artifact: RunArtifact): number => artifact.hel
  *  seeding never interpolates a path into a command string. */
 export interface CheckBox {
   exec(command: string): Promise<{ exitCode: number; stdout: string; stderr: string }>
-  fs?: { write(path: string, content: string): Promise<void> }
+  fs?: { write(path: string, content: string): Promise<unknown> }
 }
 
 /** Seed a test file into the box. Prefers the structured `fs.write` seam so the path/
diff --git a/package.json b/package.json
index bead0a17..b82ec432 100644
--- a/package.json
+++ b/package.json
@@ -105,8 +105,8 @@
   "devDependencies": {
     "@biomejs/biome": "^2.4.15",
     "@tangle-network/agent-eval": "^0.115.1",
-    "@tangle-network/agent-interface": "^0.25.0",
-    "@tangle-network/sandbox": "^0.9.7",
+    "@tangle-network/agent-interface": "file:/tmp/tangle-network-agent-interface-candidate.tgz",
+    "@tangle-network/sandbox": "^0.10.5",
     "@types/node": "^25.9.3",
     "@types/tar-stream": "3.1.4",
     "playwright": "^1.61.0",
@@ -137,7 +137,7 @@
   "peerDependencies": {
     "@tangle-network/agent-eval": ">=0.115.1 <1.0.0",
     "@tangle-network/agent-interface": ">=0.25.0 <1.0.0",
-    "@tangle-network/sandbox": ">=0.8.0 <1.0.0",
+    "@tangle-network/sandbox": ">=0.10.5 <1.0.0",
     "playwright": "^1.40.0"
   },
   "peerDependenciesMeta": {
@@ -149,7 +149,7 @@
     }
   },
   "dependencies": {
-    "@tangle-network/agent-knowledge": "^1.12.1",
+    "@tangle-network/agent-knowledge": "^2.0.0",
     "@tangle-network/agent-profile-materialize": "0.3.2",
     "tar-stream": "3.2.0"
   }
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 974b65ea..9244a569 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -9,8 +9,8 @@ importers:
   .:
     dependencies:
       '@tangle-network/agent-knowledge':
-        specifier: ^1.12.1
-        version: 1.12.1(typescript@5.9.3)
+        specifier: ^2.0.0
+        version: file:../tangle-network-agent-knowledge-candidate.tgz(typescript@5.9.3)
       '@tangle-network/agent-profile-materialize':
         specifier: 0.3.2
         version: 0.3.2
@@ -25,11 +25,11 @@ importers:
         specifier: ^0.115.1
         version: 0.115.1(typescript@5.9.3)
       '@tangle-network/agent-interface':
-        specifier: ^0.25.0
-        version: 0.25.0
+        specifier: file:/tmp/tangle-network-agent-interface-candidate.tgz
+        version: file:../tangle-network-agent-interface-candidate.tgz
       '@tangle-network/sandbox':
-        specifier: ^0.9.7
-        version: 0.9.7(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))
+        specifier: ^0.10.5
+        version: 0.10.5(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))
       '@types/node':
         specifier: ^25.9.3
         version: 25.9.3
@@ -649,31 +649,70 @@ packages:
   '@tangle-network/agent-core@0.3.8':
     resolution: {integrity: sha512-bZfVpdiFjXbcQwSxSQABdtXEmUuapWChCcXrHkP6fAnwqEy9hWEfeZLlMZUy+1XaOfpTaMLUOEAwYhXfN018wQ==}
 
+  '@tangle-network/agent-core@0.4.9':
+    resolution: {integrity: sha512-BFU4WV12Y6D28PX+o8LIVkIMD+cXwrL1uyV182l12Bn9zEu7VHt2oVMEEzbsN8L+X0xM9baEfMCHTFKFNJv7Aw==}
+
   '@tangle-network/agent-eval@0.115.1':
     resolution: {integrity: sha512-Yd487u8v0c8/+On9kaYGASS88oLhOJ5p0zwRMGdbWRSWC1UODbIgk1Fqdyq6EI60v+lZ+yDg/7LHkIrNi4kjPA==}
     engines: {node: '>=20'}
     hasBin: true
 
+  '@tangle-network/agent-eval@0.116.0':
+    resolution: {integrity: sha512-smOENca2M8s1JH/5vQc0vETR0LPhkZVC6tCtITDwwyof4toPPg4Rek9U5pRpykmvsGJrpiOzpVLPNBXzOVgYHQ==}
+    engines: {node: '>=20'}
+    hasBin: true
+
   '@tangle-network/agent-interface@0.13.0':
     resolution: {integrity: sha512-CeTPGRLoXqpt0h+BCyFgZPkfU1zyRpWmqfD+85i/uk+uvbqxkfI+JprfKVf3tBsQuCgJPSjPt5qjdW8n3h2BVg==}
 
   '@tangle-network/agent-interface@0.17.1':
     resolution: {integrity: sha512-B7dRJTo0HSUtgBCB1VMwkTFYkLUaRr/4BcRglrQuGhGUwOzKv1RYyMejOVh5M3a5AagY9N79f7GYbjcA3UmnIA==}
 
+  '@tangle-network/agent-interface@0.21.0':
+    resolution: {integrity: sha512-jDxhVJgxymrvU1RLWxWKueuaWQIpBAfrW8BuVTB5m2Y4eFMLo1SawDBEMDLdZN4/Gf34xrFrsRk+PAj9brGKMQ==}
+
   '@tangle-network/agent-interface@0.22.0':
     resolution: {integrity: sha512-7fsJhNdvTmOB1X9E2owl06jzyrqaN+jMkOPVKbK7dvNqQvf627PowCtL/edhUqEEv+K0FWtkwG3R3xqjX7VNZg==}
 
+  '@tangle-network/agent-interface@0.24.0':
+    resolution: {integrity: sha512-iaHWNTYne49cBYXwb72NjGDw5bjT2KVJlfawXygbvkqnfUsQG57BS++Yjf/XYvwJJdDeEaGs+Xvy8SWYkvu0qA==}
+
   '@tangle-network/agent-interface@0.25.0':
     resolution: {integrity: sha512-bMKjyjk8bk2651HleiTEUOTPgNE2bWxCEl9VBVUyub/UkP87Y09byXOxeDEsNGc4TIT19Y5Uo0b+k1kRWXkMVQ==}
 
-  '@tangle-network/agent-knowledge@1.12.1':
-    resolution: {integrity: sha512-j5riyIvz5C+ZBELTtNVtmUbcKAMpgTHMNmDQ12MhOIPJv5Y7AeX2oCWuF6c8aPqg1hOrsLqgqXSr1zkV0ePsrA==}
-    engines: {node: '>=20'}
+  '@tangle-network/agent-interface@file:../tangle-network-agent-interface-candidate.tgz':
+    resolution: {integrity: sha512-GG3hZc/bUPQ6ssGSzPkhJcmxq2lUPllSXt5VsCvFVl2fd6wf6kBtDEv38JyUyeJEF/utTFsxNg4dNDQfEYhFJQ==, tarball: file:../tangle-network-agent-interface-candidate.tgz}
+    version: 0.24.0
+
+  '@tangle-network/agent-knowledge@file:../tangle-network-agent-knowledge-candidate.tgz':
+    resolution: {integrity: sha512-8qOttTET4huTNz8t4NHLl+iMuiVlbzcxFiiGmUNzbEsfsGqE8vDc4Jiu7CVEl2yqwfo8YEw6tirWCKyB8zUAQg==, tarball: file:../tangle-network-agent-knowledge-candidate.tgz}
+    version: 2.0.0
+    engines: {node: '>=20.19.0'}
     hasBin: true
 
   '@tangle-network/agent-profile-materialize@0.3.2':
     resolution: {integrity: sha512-jCj1Hc/brPQ5eS7or9A+Klv0FAZcGtyFr7kx2cKfk+wCj0/4Di3k1pMjayrmRNgI/7Oc47gt6/t+qvdtZxBkAw==}
 
+  '@tangle-network/sandbox@0.10.5':
+    resolution: {integrity: sha512-u76Ht9f7HahLdIcZUz98FwomHbfWOsBusydbpd532nVF4UfcQ9tlntSXUJ4huiIvjplAp8HW/wCOeSt3aow7Qg==}
+    peerDependencies:
+      '@mastra/core': ^1.36.0
+      '@modelcontextprotocol/sdk': ^1.29.0
+      ai: ^6.0.175
+      openai: ^6.36.0
+      viem: ^2.0.0
+    peerDependenciesMeta:
+      '@mastra/core':
+        optional: true
+      '@modelcontextprotocol/sdk':
+        optional: true
+      ai:
+        optional: true
+      openai:
+        optional: true
+      viem:
+        optional: true
+
   '@tangle-network/sandbox@0.9.7':
     resolution: {integrity: sha512-9pCwJ5MlF7RUpp0AQKQDFyR0yu+E0udEhWkqhrlb/RuoJxlt72zVPuzO4FnMb1MZTkfjStmomC3k5xQyqi1YSA==}
     peerDependencies:
@@ -940,6 +979,9 @@ packages:
     engines: {node: ^8.16.0 || ^10.6.0 || >=11.0.0}
     os: [darwin]
 
+  graceful-fs@4.2.11:
+    resolution: {integrity: sha512-RbJ5/jmFcNNCcDV5o9eTnBLJ/HszWV0P73bc+Ff4nS/rJj+YaS6IGyiOL0VoBYX+l1Wrl3k63h/KrH+nhJ0XvQ==}
+
   hono@4.12.28:
     resolution: {integrity: sha512-YwUvVpSF7m1yOblFPrU3Hbo8XhPheBoiyfGuII6z19LnOr6JpDnyyp7LFNrfV56wS8tpvtBFGRISHN02pDdLOA==}
     engines: {node: '>=16.9.0'}
@@ -1072,6 +1114,9 @@ packages:
     resolution: {integrity: sha512-qif0+jGGZoLWdHey3UFHHWP0H7Gbmsk8T5VEqyYFbWqPr1XqvLGBbk/sl8V5exGmcYJklJOhOQq1pV9IcsiFag==}
     engines: {node: ^10 || ^12 || >=14}
 
+  proper-lockfile@4.1.2:
+    resolution: {integrity: sha512-TjNPblN4BwAWMXU8s9AEz4JmQxnD1NNL7bNOY/AKUzyamc379FWASUhc/K1pL2noVb+XmZKLL68cjzLsiOAMaA==}
+
   punycode.js@2.3.1:
     resolution: {integrity: sha512-uxFIHU0YlHYhDQtV4R9J6a52SLx28BCjT+4ieh7IGbgwVJWO+km431c4yRlREUAsAmt/uMjQUyQHNEPf0M39CA==}
     engines: {node: '>=6'}
@@ -1084,6 +1129,10 @@ packages:
     resolution: {integrity: sha512-qYg9KP24dD5qka9J47d0aVky0N+b4fTU89LN9iDnjB5waksiC49rvMB0PrUJQGoTmH50XPiqOvAjDfaijGxYZw==}
     engines: {node: '>=8'}
 
+  retry@0.12.0:
+    resolution: {integrity: sha512-9LkiTwjUh6rT555DtE9rTX+BKByPfrMzEAtnlEtdEwr3Nkffwiihqe2bWADg+OQRjt9gl6ICdmB/ZFDCGAtSow==}
+    engines: {node: '>= 4'}
+
   rollup@4.60.2:
     resolution: {integrity: sha512-J9qZyW++QK/09NyN/zeO0dG/1GdGfyp9lV8ajHnRVLfo/uFsbji5mHnDgn/qYdUHyCkM2N+8VyspgZclfAh0eQ==}
     engines: {node: '>=18.0.0', npm: '>=8.0.0'}
@@ -1092,6 +1141,9 @@ packages:
   siginfo@2.0.0:
     resolution: {integrity: sha512-ybx0WO1/8bSBLEWXZvEd7gMW3Sn3JFlW3TvX1nREbDLRNQNaeNN8WK0meBwPdAaOI7TtRRRJn/Es1zhrrCHu7g==}
 
+  signal-exit@3.0.7:
+    resolution: {integrity: sha512-wnD2ZE+l+SPC/uoS0vXeE9L1+0wuaMqKlfz9AMUo38JsyLSBWSFcHR1Rri62LZc12vLr1gb3jl7iwQhgwpAbGQ==}
+
   source-map-js@1.2.1:
     resolution: {integrity: sha512-UXWMKhLOwVKb728IUtQPXxfYU+usdybtUrK/8uGE8CQMvrhOpwvzDBwj0QhSL7MQc7vIsISBG8VQ8+IDQxpfQA==}
     engines: {node: '>=0.10.0'}
@@ -1694,6 +1746,11 @@ snapshots:
       '@tangle-network/agent-interface': 0.17.1
       zod: 4.4.3
 
+  '@tangle-network/agent-core@0.4.9':
+    dependencies:
+      '@tangle-network/agent-interface': 0.24.0
+      zod: 4.4.3
+
   '@tangle-network/agent-eval@0.115.1(typescript@5.9.3)':
     dependencies:
       '@asteasolutions/zod-to-openapi': 8.5.0(zod@4.4.3)
@@ -1712,6 +1769,24 @@ snapshots:
       - typescript
       - utf-8-validate
 
+  '@tangle-network/agent-eval@0.116.0(typescript@5.9.3)':
+    dependencies:
+      '@asteasolutions/zod-to-openapi': 8.5.0(zod@4.4.3)
+      '@ax-llm/ax': 19.0.45(zod@4.4.3)
+      '@hono/node-server': 2.0.8(hono@4.12.28)
+      '@tangle-network/agent-interface': 0.22.0
+      '@tangle-network/tcloud': 0.4.14(typescript@5.9.3)(zod@4.4.3)
+      hono: 4.12.28
+      zod: 4.4.3
+    transitivePeerDependencies:
+      - '@mastra/core'
+      - '@modelcontextprotocol/sdk'
+      - ai
+      - bufferutil
+      - openai
+      - typescript
+      - utf-8-validate
+
   '@tangle-network/agent-interface@0.13.0':
     dependencies:
       zod: 4.4.3
@@ -1720,17 +1795,31 @@ snapshots:
     dependencies:
       zod: 4.4.3
 
+  '@tangle-network/agent-interface@0.21.0':
+    dependencies:
+      zod: 4.4.3
+
   '@tangle-network/agent-interface@0.22.0':
     dependencies:
       zod: 4.4.3
 
+  '@tangle-network/agent-interface@0.24.0':
+    dependencies:
+      zod: 4.4.3
+
   '@tangle-network/agent-interface@0.25.0':
     dependencies:
       zod: 4.4.3
 
-  '@tangle-network/agent-knowledge@1.12.1(typescript@5.9.3)':
+  '@tangle-network/agent-interface@file:../tangle-network-agent-interface-candidate.tgz':
     dependencies:
-      '@tangle-network/agent-eval': 0.115.1(typescript@5.9.3)
+      '@noble/hashes': 1.8.0
+      zod: 4.4.3
+
+  '@tangle-network/agent-knowledge@file:../tangle-network-agent-knowledge-candidate.tgz(typescript@5.9.3)':
+    dependencies:
+      '@tangle-network/agent-eval': 0.116.0(typescript@5.9.3)
+      proper-lockfile: 4.1.2
       zod: 4.4.3
     transitivePeerDependencies:
       - '@mastra/core'
@@ -1745,6 +1834,13 @@ snapshots:
     dependencies:
       '@tangle-network/agent-interface': 0.25.0
 
+  '@tangle-network/sandbox@0.10.5(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))':
+    dependencies:
+      '@tangle-network/agent-core': 0.4.9
+      '@tangle-network/agent-interface': 0.21.0
+    optionalDependencies:
+      viem: 2.54.6(typescript@5.9.3)(zod@4.4.3)
+
   '@tangle-network/sandbox@0.9.7(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))':
     dependencies:
       '@tangle-network/agent-core': 0.3.8
@@ -2018,6 +2114,8 @@ snapshots:
   fsevents@2.3.3:
     optional: true
 
+  graceful-fs@4.2.11: {}
+
   hono@4.12.28: {}
 
   isows@1.0.7(ws@8.21.0):
@@ -2137,12 +2235,20 @@ snapshots:
       picocolors: 1.1.1
       source-map-js: 1.2.1
 
+  proper-lockfile@4.1.2:
+    dependencies:
+      graceful-fs: 4.2.11
+      retry: 0.12.0
+      signal-exit: 3.0.7
+
   punycode.js@2.3.1: {}
 
   readdirp@4.1.2: {}
 
   resolve-from@5.0.0: {}
 
+  retry@0.12.0: {}
+
   rollup@4.60.2:
     dependencies:
       '@types/estree': 1.0.8
@@ -2176,6 +2282,8 @@ snapshots:
 
   siginfo@2.0.0: {}
 
+  signal-exit@3.0.7: {}
+
   source-map-js@1.2.1: {}
 
   source-map@0.7.6: {}
diff --git a/scripts/verify-package-exports.mjs b/scripts/verify-package-exports.mjs
index 4503d598..c28e1f85 100644
--- a/scripts/verify-package-exports.mjs
+++ b/scripts/verify-package-exports.mjs
@@ -1,10 +1,11 @@
 import { spawnSync } from 'node:child_process'
-import { mkdirSync, mkdtempSync, readFileSync, rmSync, symlinkSync } from 'node:fs'
+import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
 import { dirname, join, resolve } from 'node:path'
 import { fileURLToPath } from 'node:url'
 
 const repoRoot = resolve(dirname(fileURLToPath(import.meta.url)), '..')
-const tempRoot = mkdtempSync(join(repoRoot, '.tmp-package-exports-'))
+const tempRoot = mkdtempSync(join(tmpdir(), 'agent-runtime-package-'))
 
 try {
   const packDir = join(tempRoot, 'pack')
@@ -12,7 +13,7 @@ try {
   const appDir = join(tempRoot, 'app')
   mkdirSync(packDir, { recursive: true })
   mkdirSync(unpackDir, { recursive: true })
-  mkdirSync(join(appDir, 'node_modules', '@tangle-network'), { recursive: true })
+  mkdirSync(appDir, { recursive: true })
 
   run('pnpm', ['pack', '--pack-destination', packDir], repoRoot)
   const tarballs = run('find', [packDir, '-maxdepth', '1', '-name', '*.tgz', '-print'], repoRoot)
@@ -52,7 +53,147 @@ try {
     }
   }
 
-  symlinkSync(packageDir, join(appDir, 'node_modules', '@tangle-network', 'agent-runtime'), 'dir')
+  const repoPackageJson = JSON.parse(readFileSync(join(repoRoot, 'package.json'), 'utf8'))
+  const knowledgePackageDir = join(
+    repoRoot,
+    'node_modules',
+    '@tangle-network',
+    'agent-knowledge',
+  )
+  if (!existsSync(join(knowledgePackageDir, 'package.json'))) {
+    throw new Error('packed consumer requires an installed @tangle-network/agent-knowledge package')
+  }
+  const knowledgePackDir = join(tempRoot, 'knowledge')
+  mkdirSync(knowledgePackDir, { recursive: true })
+  run('pnpm', ['pack', '--pack-destination', knowledgePackDir], knowledgePackageDir)
+  const knowledgeTarballs = run(
+    'find',
+    [knowledgePackDir, '-maxdepth', '1', '-name', '*.tgz', '-print'],
+    repoRoot,
+  )
+    .trim()
+    .split('\n')
+    .filter(Boolean)
+  if (knowledgeTarballs.length !== 1) {
+    throw new Error(`expected exactly one packed knowledge tarball, found ${knowledgeTarballs.length}`)
+  }
+  const peerPackages = [
+    '@tangle-network/agent-eval',
+    '@tangle-network/agent-interface',
+    '@tangle-network/sandbox',
+    'playwright',
+  ]
+  const peerDependencies = Object.fromEntries(
+    peerPackages.map((name) => {
+      const version = repoPackageJson.devDependencies?.[name]
+      if (typeof version !== 'string' || version.length === 0) {
+        throw new Error(`packed consumer requires a ${name} development dependency`)
+      }
+      return [name, version]
+    }),
+  )
+  writeFileSync(
+    join(appDir, 'package.json'),
+    `${JSON.stringify(
+      {
+        private: true,
+        type: 'module',
+        dependencies: {
+          '@tangle-network/agent-runtime': `file:${tarballs[0]}`,
+          ...peerDependencies,
+        },
+        devDependencies: { typescript: repoPackageJson.devDependencies.typescript },
+      },
+      null,
+      2,
+    )}\n`,
+  )
+  writeFileSync(
+    join(appDir, 'pnpm-workspace.yaml'),
+    `overrides:\n  '@tangle-network/agent-knowledge': file:${knowledgeTarballs[0]}\n`,
+  )
+  writeFileSync(
+    join(appDir, 'tsconfig.json'),
+    `${JSON.stringify(
+      {
+        compilerOptions: {
+          target: 'ES2022',
+          module: 'NodeNext',
+          moduleResolution: 'NodeNext',
+          strict: true,
+          noEmit: true,
+          skipLibCheck: false,
+        },
+        include: ['consumer.ts'],
+      },
+      null,
+      2,
+    )}\n`,
+  )
+  writeFileSync(
+    join(appDir, 'consumer.ts'),
+    `
+      import type {
+        AgentCandidateProfileActivation,
+        AgentImprovementProposal,
+        AgentImprovementReview,
+        CandidateExecutionEvidence,
+      } from '@tangle-network/agent-interface'
+      import { SandboxClient } from '@tangle-network/sandbox'
+      import type { AgentCandidateTaskExecution } from '@tangle-network/agent-runtime/candidate-execution'
+      import {
+        createSandboxApprovedCandidateExecutor,
+        parseAgentCandidateProfileActivation,
+        sandboxApprovedCandidateExecutionSupport,
+        verifyCandidateExecutionEvidence,
+        type CreateSandboxApprovedCandidateExecutorOptions,
+      } from '@tangle-network/agent-runtime/intelligence'
+
+      const client = new SandboxClient({
+        apiKey: 'sk_sandbox_compile_only',
+        baseUrl: 'https://sandbox.example.com',
+        trustLocalCliAuth: false,
+      })
+      declare const ports: CreateSandboxApprovedCandidateExecutorOptions['ports']
+      declare const grader: CreateSandboxApprovedCandidateExecutorOptions['grader']
+      declare const outputArtifacts: CreateSandboxApprovedCandidateExecutorOptions['outputArtifacts']
+      declare const traceStore: CreateSandboxApprovedCandidateExecutorOptions['traceStore']
+      declare const claimStore: CreateSandboxApprovedCandidateExecutorOptions['claimStore']
+      declare const proposal: AgentImprovementProposal
+      declare const review: AgentImprovementReview
+      declare const task: AgentCandidateTaskExecution
+      declare const storedEvidence: unknown
+
+      const executor = createSandboxApprovedCandidateExecutor({
+        client,
+        ports,
+        grader,
+        outputArtifacts,
+        traceStore,
+        claimStore,
+        authorizeReview: async () => true,
+      })
+      const execution: Promise<CandidateExecutionEvidence> = executor.execute({
+        proposal,
+        review,
+        task,
+      })
+      const evidence = verifyCandidateExecutionEvidence([storedEvidence], {
+        proposal,
+        review,
+        expectedCount: 1,
+      })[0]
+      const activation: AgentCandidateProfileActivation =
+        parseAgentCandidateProfileActivation(evidence?.profileActivation)
+      const outcome: 'output' = sandboxApprovedCandidateExecutionSupport.outcomes[0]
+      void execution
+      void activation
+      void outcome
+    `,
+  )
+  run('pnpm', ['install', '--ignore-scripts', '--config.auto-install-peers=false'], appDir)
+  run('pnpm', ['exec', 'tsc', '-p', 'tsconfig.json'], appDir)
+
   run(
     process.execPath,
     [
@@ -70,6 +211,10 @@ try {
           'composeCertifiedProfile',
           'manifestFromProfile',
           'CapabilityNotAdmittedError',
+          'createSandboxApprovedCandidateExecutor',
+          'parseAgentCandidateProfileActivation',
+          'sandboxApprovedCandidateExecutionSupport',
+          'verifyCandidateExecutionEvidence',
         ]
         for (const name of expectedIntelligence) {
           if (!(name in intelligence)) throw new Error('missing intelligence export ' + name)
diff --git a/src/candidate-execution/artifacts.ts b/src/candidate-execution/artifacts.ts
index bcca8faa..5c8ab84e 100644
--- a/src/candidate-execution/artifacts.ts
+++ b/src/candidate-execution/artifacts.ts
@@ -5,8 +5,8 @@ import { relative, resolve, sep } from 'node:path'
 import type {
   AgentCandidateArtifactRef,
   AgentCandidateCapturedArtifact,
-  AgentCandidateProfilePlanMaterialV1,
-  AgentCandidateWorkspaceManifestMaterialV1,
+  AgentCandidateProfilePlanMaterial,
+  AgentCandidateWorkspaceManifestMaterial,
   AgentCandidateWorkspaceSnapshotEvidence,
 } from '@tangle-network/agent-interface'
 
@@ -64,7 +64,7 @@ export async function verifyWorkspaceSnapshotArtifacts(
 
 export async function verifyMaterializedWorkspace(
   root: string,
-  expected: AgentCandidateWorkspaceManifestMaterialV1,
+  expected: AgentCandidateWorkspaceManifestMaterial,
   options: { ignoredProtectedRootEntries?: readonly ('.git' | '.sidecar')[] } = {},
 ): Promise<void> {
   const observed = await scanWorkspace(root, new Set(options.ignoredProtectedRootEntries ?? []))
@@ -82,8 +82,8 @@ export async function captureMaterializedWorkspace(
     }
   } = {},
 ): Promise<{
-  manifest: AgentCandidateWorkspaceManifestMaterialV1
-  files: ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>
+  manifest: AgentCandidateWorkspaceManifestMaterial
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>
 }> {
   const observed = await scanWorkspace(
     root,
@@ -103,9 +103,9 @@ export async function captureMaterializedWorkspace(
 /** Capture exact verified regular-file bytes for fresh isolated materialization. */
 export async function readMaterializedWorkspaceFiles(
   root: string,
-  expected: AgentCandidateWorkspaceManifestMaterialV1,
+  expected: AgentCandidateWorkspaceManifestMaterial,
   options: { ignoredProtectedRootEntries?: readonly ('.git' | '.sidecar')[] } = {},
-): Promise<ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>> {
+): Promise<ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>> {
   const observed = await captureMaterializedWorkspace(root, options)
   assertWorkspaceManifest(observed.manifest, expected)
   return observed.files.map((file) =>
@@ -114,8 +114,8 @@ export async function readMaterializedWorkspaceFiles(
 }
 
 function assertWorkspaceManifest(
-  observed: AgentCandidateWorkspaceManifestMaterialV1,
-  expected: AgentCandidateWorkspaceManifestMaterialV1,
+  observed: AgentCandidateWorkspaceManifestMaterial,
+  expected: AgentCandidateWorkspaceManifestMaterial,
 ): void {
   if (!Buffer.from(canonicalCandidateBytes(observed)).equals(canonicalCandidateBytes(expected))) {
     throw new Error(
@@ -126,7 +126,7 @@ function assertWorkspaceManifest(
 
 export async function verifyMaterializedProfileWorkspace(
   root: string,
-  expected: AgentCandidateProfilePlanMaterialV1,
+  expected: AgentCandidateProfilePlanMaterial,
 ): Promise<void> {
   const observed = await scanWorkspace(root, new Set())
   const observedProfile = observed.manifest.files.map(({ path, mode, sha256 }) => ({
@@ -152,8 +152,8 @@ async function scanWorkspace(
     maxTotalFileBytes: number
   },
 ): Promise<{
-  manifest: AgentCandidateWorkspaceManifestMaterialV1
-  files: Array<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>
+  manifest: AgentCandidateWorkspaceManifestMaterial
+  files: Array<{ path: string; mode: number; bytes: Uint8Array }>
 }> {
   const absoluteRoot = resolve(root)
   const rootStats = await lstat(absoluteRoot)
@@ -163,8 +163,8 @@ async function scanWorkspace(
   if ((await realpath(absoluteRoot)) !== absoluteRoot) {
     throw new Error('workspace root has a symlinked path component')
   }
-  const files: AgentCandidateWorkspaceManifestMaterialV1['files'] = []
-  const capturedFiles: Array<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }> = []
+  const files: AgentCandidateWorkspaceManifestMaterial['files'] = []
+  const capturedFiles: Array<{ path: string; mode: number; bytes: Uint8Array }> = []
   let totalBytes = 0
 
   async function visit(directory: string): Promise<void> {
@@ -207,9 +207,6 @@ async function scanWorkspace(
           throw new Error(`workspace contains a hard-linked file: ${relPath}`)
         }
         const mode = openedStats.mode & 0o777
-        if (mode !== 0o644 && mode !== 0o755) {
-          throw new Error(`workspace file has unsupported mode ${mode.toString(8)}: ${relPath}`)
-        }
         if (limits && openedStats.size > limits.maxFileBytes) {
           throw new Error(`workspace file exceeds maxFileBytes: ${relPath}`)
         }
@@ -223,14 +220,13 @@ async function scanWorkspace(
           relPath,
         )
         totalBytes += bytes.byteLength
-        const supportedMode = mode as 0o644 | 0o755
         files.push({
           path: relPath,
-          mode: supportedMode,
+          mode,
           sha256: sha256Bytes(bytes),
           byteLength: bytes.byteLength,
         })
-        capturedFiles.push({ path: relPath, mode: supportedMode, bytes: Uint8Array.from(bytes) })
+        capturedFiles.push({ path: relPath, mode, bytes: Uint8Array.from(bytes) })
       } finally {
         await descriptor.close()
       }
diff --git a/src/candidate-execution/claim-file-formats.ts b/src/candidate-execution/claim-file-formats.ts
index 54b99463..c6e4535c 100644
--- a/src/candidate-execution/claim-file-formats.ts
+++ b/src/candidate-execution/claim-file-formats.ts
@@ -1,12 +1,48 @@
-import type { Sha256Digest } from '@tangle-network/agent-interface'
+import {
+  type AgentCandidateArtifactRef,
+  agentCandidateArtifactRefSchema,
+  type Sha256Digest,
+} from '@tangle-network/agent-interface'
 
 import type { AgentCandidateExecutionClaim, AgentCandidateExecutionTerminalRecord } from './claim'
+import { immutableCandidateValue } from './digest'
+import { assertExactObjectKeys } from './exact-object'
 
-export const CLAIM_FORMAT_VERSION = 7
-export const PENDING_FORMAT_VERSION = 1
-export const TERMINAL_FORMAT_VERSION = 3
+export const CLAIM_FORMAT_VERSION = 8
+export const PENDING_FORMAT_VERSION = 2
+export const TERMINAL_FORMAT_VERSION = 4
 export const PHASE_FORMAT_VERSION = 1
 
+export interface AgentCandidatePreparationEvidence {
+  readonly executionPlan: AgentCandidateArtifactRef
+  readonly materializationReceipt: AgentCandidateArtifactRef
+}
+
+export function sealCandidatePreparationEvidence(
+  value: unknown,
+  executionPlanDigest: Sha256Digest,
+): AgentCandidatePreparationEvidence {
+  if (value === null || typeof value !== 'object' || Array.isArray(value)) {
+    throw new Error('candidate execution preparation evidence must be an object')
+  }
+  const record = value as Record<string, unknown>
+  assertExactObjectKeys(
+    record,
+    ['executionPlan', 'materializationReceipt'],
+    'candidate execution preparation evidence',
+  )
+  const executionPlan = immutableCandidateValue(
+    agentCandidateArtifactRefSchema.parse(record.executionPlan),
+  )
+  const materializationReceipt = immutableCandidateValue(
+    agentCandidateArtifactRefSchema.parse(record.materializationReceipt),
+  )
+  if (executionPlan.sha256 !== executionPlanDigest) {
+    throw new Error('candidate execution plan artifact digest does not match its claim')
+  }
+  return Object.freeze({ executionPlan, materializationReceipt })
+}
+
 export interface PersistedAgentCandidateExecutionClaim extends AgentCandidateExecutionClaim {
   version: typeof CLAIM_FORMAT_VERSION
   phase: 'claimed'
diff --git a/src/candidate-execution/claim-plan.ts b/src/candidate-execution/claim-plan.ts
index 8d8333a3..18283d83 100644
--- a/src/candidate-execution/claim-plan.ts
+++ b/src/candidate-execution/claim-plan.ts
@@ -1,4 +1,4 @@
-import type { Sha256Digest } from '@tangle-network/agent-interface'
+import type { AgentCandidateArtifactRef, Sha256Digest } from '@tangle-network/agent-interface'
 
 import { type AgentCandidateExecutionClaim, candidateClaimFileInternals } from './claim'
 import { canonicalCandidateDigest } from './digest'
@@ -9,6 +9,10 @@ import type { PreparedAgentCandidateExecution } from './types'
 /** Extract the complete durable claim from a prepared execution. */
 export function candidateExecutionClaim(
   prepared: PreparedAgentCandidateExecution,
+  preparationEvidence: {
+    executionPlan: AgentCandidateArtifactRef
+    materializationReceipt: AgentCandidateArtifactRef
+  },
 ): AgentCandidateExecutionClaim {
   const state = assertPreparedCandidateIntegrity(prepared)
   const material = prepared.executionPlan.value.material
@@ -32,6 +36,21 @@ export function candidateExecutionClaim(
       'candidate preparation expires before its full execution and cleanup owner window',
     )
   }
+  if (
+    preparationEvidence.executionPlan.sha256 !== prepared.executionPlan.value.digest ||
+    preparationEvidence.executionPlan.byteLength !== state.executionPlan.bytes.byteLength
+  ) {
+    throw new Error('persisted execution plan does not match its canonical preparation bytes')
+  }
+  if (
+    preparationEvidence.materializationReceipt.sha256 !== state.materializationReceipt.digest ||
+    preparationEvidence.materializationReceipt.byteLength !==
+      state.materializationReceipt.bytes.byteLength
+  ) {
+    throw new Error(
+      'persisted materialization receipt does not match its canonical preparation bytes',
+    )
+  }
   return candidateClaimFileInternals.sealClaim({
     executionId: prepared.executionId,
     attempt: attempt.number,
@@ -39,6 +58,7 @@ export function candidateExecutionClaim(
     retryPolicy: attempt.retryPolicy,
     bundleDigest: prepared.bundle.digest,
     executionPlanDigest: prepared.executionPlan.value.digest,
+    preparationEvidence,
     retryLineageDigest: retryLineageDigest(prepared, state.resultTimeoutMs),
     leaseExpiresAtMs,
     resultTimeoutMs: state.resultTimeoutMs,
diff --git a/src/candidate-execution/claim-terminal.ts b/src/candidate-execution/claim-terminal.ts
index d71b1a9c..684a3f08 100644
--- a/src/candidate-execution/claim-terminal.ts
+++ b/src/candidate-execution/claim-terminal.ts
@@ -1,4 +1,8 @@
-import type { AgentCandidateArtifactRef, Sha256Digest } from '@tangle-network/agent-interface'
+import type {
+  AgentCandidateArtifactRef,
+  AgentCandidateFixedSpend,
+  Sha256Digest,
+} from '@tangle-network/agent-interface'
 import { agentCandidateArtifactRefSchema } from '@tangle-network/agent-interface'
 
 import type {
@@ -10,8 +14,8 @@ import type {
   AgentCandidateExecutionStageResult,
   AgentCandidateExecutionTerminalRecord,
   AgentCandidateExecutionTerminalResult,
-  AgentCandidateExecutionUsage,
 } from './claim'
+import { sealCandidatePreparationEvidence } from './claim-file-formats'
 import { canonicalCandidateDigest, immutableCandidateValue } from './digest'
 import { assertExactObjectKeys as assertExactKeys } from './exact-object'
 
@@ -27,6 +31,7 @@ export function terminalRecord(
     attempt: claim.attempt,
     bundleDigest: claim.bundleDigest,
     executionPlanDigest: claim.executionPlanDigest,
+    preparationEvidence: claim.preparationEvidence,
     ...terminal,
   }
   return immutableCandidateValue({
@@ -140,6 +145,7 @@ export function sealTerminalRecordValue(
           'attempt',
           'bundleDigest',
           'executionPlanDigest',
+          'preparationEvidence',
           'terminalDigest',
           'schemaVersion',
           'status',
@@ -154,6 +160,7 @@ export function sealTerminalRecordValue(
           'attempt',
           'bundleDigest',
           'executionPlanDigest',
+          'preparationEvidence',
           'terminalDigest',
           'schemaVersion',
           'status',
@@ -164,15 +171,20 @@ export function sealTerminalRecordValue(
         ],
     label,
   )
+  const executionPlanDigest = requireString(
+    value.executionPlanDigest,
+    label,
+    'executionPlanDigest',
+  ) as Sha256Digest
   const identity = {
     executionId: requireString(value.executionId, label, 'executionId'),
     attempt: requireNumber(value.attempt, label, 'attempt'),
     bundleDigest: requireString(value.bundleDigest, label, 'bundleDigest') as Sha256Digest,
-    executionPlanDigest: requireString(
-      value.executionPlanDigest,
-      label,
-      'executionPlanDigest',
-    ) as Sha256Digest,
+    executionPlanDigest,
+    preparationEvidence: sealCandidatePreparationEvidence(
+      value.preparationEvidence,
+      executionPlanDigest,
+    ),
   }
   assertExecutionId(identity.executionId)
   if (!Number.isSafeInteger(identity.attempt) || identity.attempt < 1) {
@@ -180,16 +192,15 @@ export function sealTerminalRecordValue(
   }
   assertSha256Digest(identity.bundleDigest, 'bundleDigest')
   assertSha256Digest(identity.executionPlanDigest, 'executionPlanDigest')
+  if (identity.preparationEvidence.executionPlan.sha256 !== identity.executionPlanDigest) {
+    throw new Error(`${label} execution plan artifact does not match executionPlanDigest`)
+  }
   const result = sealTerminalResult(
     status === 'succeeded'
       ? {
           schemaVersion: requireNumber(value.schemaVersion, label, 'schemaVersion') as 1,
           status,
-          usage: requireObject(
-            value.usage,
-            label,
-            'usage',
-          ) as unknown as AgentCandidateExecutionUsage,
+          usage: requireObject(value.usage, label, 'usage') as unknown as AgentCandidateFixedSpend,
           modelSettlement: requireArtifactRef(value.modelSettlement, label, 'modelSettlement'),
           taskOutcome: requireArtifactRef(value.taskOutcome, label, 'taskOutcome'),
           benchmarkResult: requireArtifactRef(value.benchmarkResult, label, 'benchmarkResult'),
@@ -199,11 +210,7 @@ export function sealTerminalRecordValue(
           schemaVersion: requireNumber(value.schemaVersion, label, 'schemaVersion') as 1,
           status,
           failureClass: requireFailureClass(value.failureClass, label),
-          usage: requireObject(
-            value.usage,
-            label,
-            'usage',
-          ) as unknown as AgentCandidateExecutionUsage,
+          usage: requireObject(value.usage, label, 'usage') as unknown as AgentCandidateFixedSpend,
           modelSettlement: requireArtifactRef(value.modelSettlement, label, 'modelSettlement'),
           ...(value.failureEvidence
             ? {
@@ -241,7 +248,9 @@ export function assertTerminalMatchesClaim(
     terminal.executionId !== claim.executionId ||
     terminal.attempt !== claim.attempt ||
     terminal.bundleDigest !== claim.bundleDigest ||
-    terminal.executionPlanDigest !== claim.executionPlanDigest
+    terminal.executionPlanDigest !== claim.executionPlanDigest ||
+    canonicalCandidateDigest(terminal.preparationEvidence) !==
+      canonicalCandidateDigest(claim.preparationEvidence)
   ) {
     throw new Error(`candidate execution terminal record at ${path} does not match its claim`)
   }
@@ -402,7 +411,7 @@ function sealTerminalResult(
   })
 }
 
-function sealUsage(usage: AgentCandidateExecutionUsage): AgentCandidateExecutionUsage {
+function sealUsage(usage: AgentCandidateFixedSpend): AgentCandidateFixedSpend {
   assertExactKeys(
     usage,
     [
diff --git a/src/candidate-execution/claim.ts b/src/candidate-execution/claim.ts
index e31ac79d..5940620d 100644
--- a/src/candidate-execution/claim.ts
+++ b/src/candidate-execution/claim.ts
@@ -5,11 +5,13 @@ import { readFile } from 'node:fs/promises'
 import {
   type AgentCandidateArtifactRef,
   type AgentCandidateAttemptPolicy,
+  type AgentCandidateFixedSpend,
   type AgentCandidateResolvedModel,
   agentCandidateResolvedModelSchema,
   type Sha256Digest,
 } from '@tangle-network/agent-interface'
 import {
+  type AgentCandidatePreparationEvidence,
   CLAIM_FORMAT_VERSION,
   PENDING_FORMAT_VERSION,
   type PersistedAgentCandidateExecutionClaim,
@@ -17,6 +19,7 @@ import {
   type PersistedAgentCandidateExecutionPhase,
   type PersistedAgentCandidateExecutionTerminal,
   PHASE_FORMAT_VERSION,
+  sealCandidatePreparationEvidence,
   TERMINAL_FORMAT_VERSION,
 } from './claim-file-formats'
 import {
@@ -56,6 +59,8 @@ export interface AgentCandidateExecutionClaim {
   readonly retryPolicy: AgentCandidateAttemptPolicy['retryPolicy']
   readonly bundleDigest: Sha256Digest
   readonly executionPlanDigest: Sha256Digest
+  /** Durable canonical bytes needed to reconstruct the signed preparation. */
+  readonly preparationEvidence: AgentCandidatePreparationEvidence
   /** Frozen plan identity with only attempt number and per-attempt grant identity normalized. */
   readonly retryLineageDigest: Sha256Digest
   /** The winning lease stops authorizing a new terminal write at this instant. */
@@ -81,22 +86,12 @@ export type AgentCandidateExecutionFailureClass =
   | 'post-model-infrastructure'
   | 'unknown'
 
-/** Exact fixed-point usage proven by the closed evaluator model ledger. */
-export interface AgentCandidateExecutionUsage {
-  readonly costUsdNanos: number
-  readonly inputTokens: number
-  readonly outputTokens: number
-  readonly cachedInputTokens: number
-  readonly reasoningTokens: number
-  readonly modelCalls: number
-}
-
 /** Evaluator-owned terminal facts staged durably before the terminal CAS. */
 export type AgentCandidateExecutionTerminalResult =
   | {
       readonly schemaVersion: 1
       readonly status: 'succeeded'
-      readonly usage: AgentCandidateExecutionUsage
+      readonly usage: AgentCandidateFixedSpend
       readonly modelSettlement: AgentCandidateArtifactRef
       readonly taskOutcome: AgentCandidateArtifactRef
       readonly benchmarkResult: AgentCandidateArtifactRef
@@ -106,7 +101,7 @@ export type AgentCandidateExecutionTerminalResult =
       readonly schemaVersion: 1
       readonly status: 'failed'
       readonly failureClass: AgentCandidateExecutionFailureClass
-      readonly usage: AgentCandidateExecutionUsage
+      readonly usage: AgentCandidateFixedSpend
       readonly modelSettlement: AgentCandidateArtifactRef
       readonly failureEvidence?: AgentCandidateArtifactRef
     }
@@ -117,6 +112,7 @@ export type AgentCandidateExecutionTerminalRecord = AgentCandidateExecutionTermi
   readonly attempt: number
   readonly bundleDigest: Sha256Digest
   readonly executionPlanDigest: Sha256Digest
+  readonly preparationEvidence: AgentCandidateExecutionClaim['preparationEvidence']
   /** RFC 8785 SHA-256 of this record with `terminalDigest` omitted. */
   readonly terminalDigest: Sha256Digest
 }
@@ -127,7 +123,7 @@ export type AgentCandidateExecutionPhase = 'claimed' | 'candidate-may-run'
 /** Trusted, independently observed closure facts for one expired winning lease. */
 export interface AgentCandidateExecutionRecoveryEvidence {
   readonly failureClass: AgentCandidateExecutionFailureClass
-  readonly usage: AgentCandidateExecutionUsage
+  readonly usage: AgentCandidateFixedSpend
   readonly modelSettlement: AgentCandidateArtifactRef
   readonly failureEvidence?: AgentCandidateArtifactRef
   readonly process: {
@@ -422,6 +418,7 @@ function sealClaim(claim: AgentCandidateExecutionClaim): AgentCandidateExecution
       'retryPolicy',
       'bundleDigest',
       'executionPlanDigest',
+      'preparationEvidence',
       'retryLineageDigest',
       'leaseExpiresAtMs',
       'resultTimeoutMs',
@@ -447,6 +444,10 @@ function sealClaim(claim: AgentCandidateExecutionClaim): AgentCandidateExecution
   }
   assertSha256Digest(claim.bundleDigest, 'bundleDigest')
   assertSha256Digest(claim.executionPlanDigest, 'executionPlanDigest')
+  const preparationEvidence = sealCandidatePreparationEvidence(
+    claim.preparationEvidence,
+    claim.executionPlanDigest,
+  )
   assertSha256Digest(claim.retryLineageDigest, 'retryLineageDigest')
   assertPositiveTimestamp(claim.leaseExpiresAtMs, 'leaseExpiresAtMs')
   candidateResultTimeout(claim.resultTimeoutMs, claim.resultTimeoutMs)
@@ -458,6 +459,7 @@ function sealClaim(claim: AgentCandidateExecutionClaim): AgentCandidateExecution
     retryPolicy: claim.retryPolicy,
     bundleDigest: claim.bundleDigest,
     executionPlanDigest: claim.executionPlanDigest,
+    preparationEvidence,
     retryLineageDigest: claim.retryLineageDigest,
     leaseExpiresAtMs: claim.leaseExpiresAtMs,
     resultTimeoutMs: claim.resultTimeoutMs,
@@ -659,6 +661,7 @@ async function readClaim(path: string): Promise<StoredClaim> {
       'retryPolicy',
       'bundleDigest',
       'executionPlanDigest',
+      'preparationEvidence',
       'retryLineageDigest',
       'leaseExpiresAtMs',
       'resultTimeoutMs',
@@ -679,6 +682,11 @@ async function readClaim(path: string): Promise<StoredClaim> {
       path,
       'executionPlanDigest',
     ) as Sha256Digest,
+    preparationEvidence: requireObject(
+      record.preparationEvidence,
+      path,
+      'preparationEvidence',
+    ) as unknown as AgentCandidateExecutionClaim['preparationEvidence'],
     retryLineageDigest: requireString(
       record.retryLineageDigest,
       path,
diff --git a/src/candidate-execution/digest.ts b/src/candidate-execution/digest.ts
index 47fe0c7b..a1871626 100644
--- a/src/candidate-execution/digest.ts
+++ b/src/candidate-execution/digest.ts
@@ -43,6 +43,16 @@ export function canonicalCandidateDocument<T extends { digest: Sha256Digest }>(
   })
 }
 
+export function verifyCanonicalCandidateDocument<T extends { digest: Sha256Digest }>(
+  value: T,
+  label: string,
+): T {
+  if (canonicalCandidateDigest(omitTopLevelDigest(value)) !== value.digest) {
+    throw new Error(`${label} digest does not match`)
+  }
+  return immutableCandidateValue(value)
+}
+
 export function embeddedCandidateArtifact(bytes: Uint8Array): AgentCandidateEmbeddedArtifact {
   return {
     encoding: 'base64',
diff --git a/src/candidate-execution/execute.ts b/src/candidate-execution/execute.ts
index e9e8339e..336fd32f 100644
--- a/src/candidate-execution/execute.ts
+++ b/src/candidate-execution/execute.ts
@@ -2,6 +2,7 @@ import type { TraceStore } from '@tangle-network/agent-eval'
 import type { AgentCandidateTermination } from '@tangle-network/agent-interface'
 
 import type {
+  AgentCandidateExecutionClaim,
   AgentCandidateExecutionClaimStore,
   AgentCandidateExecutionFailureClass,
   AgentCandidateExecutionLease,
@@ -18,7 +19,9 @@ import {
 import { canonicalCandidateBytes } from './digest'
 import { candidatePostRunWindowMs, candidateTerminalWindowMs } from './execution-window'
 import {
+  type SealedAgentCandidateExecutorFinalCapture,
   sealAgentCandidateExecutorFinalCapture,
+  sealAgentCandidateExecutorStopAcknowledgement,
   sealAgentCandidateProtectedRunCapture,
 } from './executor-capture'
 import { failedAgentCandidateRun, finalizeAgentCandidateRun } from './finalize'
@@ -31,7 +34,7 @@ import {
   type PersistedAgentCandidateModelSettlement,
   persistCandidateBenchmarkResult,
   persistCandidateModelSettlement,
-  persistVerifiedCandidateTaskOutcome,
+  persistVerifiedAgentCandidateExecutorCapture,
 } from './outcome-evidence'
 import { persistCandidateOutputArtifact } from './output-artifacts'
 import {
@@ -48,7 +51,6 @@ import { redactProtectedReason } from './protected-redaction'
 import { ProtectedAgentCandidateTraceStore } from './protected-trace-store'
 import type {
   AgentCandidateBenchmarkGraderPort,
-  AgentCandidateExecutorFinalCapture,
   AgentCandidateExecutorPort,
   AgentCandidateExecutorRequest,
   AgentCandidateOutputArtifactPort,
@@ -105,9 +107,30 @@ export async function executePreparedAgentCandidate(
     return await failBeforeActivation(prepared, state, error, 'failed', cleanupTimeoutMs)
   }
 
+  let preparationEvidence: AgentCandidateExecutionClaim['preparationEvidence']
+  try {
+    const [executionPlan, materializationReceipt] = await Promise.all([
+      persistCandidateOutputArtifact(options.outputArtifacts, {
+        executionId: state.executionId,
+        purpose: 'execution-plan',
+        bytes: state.executionPlan.bytes,
+      }),
+      persistCandidateOutputArtifact(options.outputArtifacts, {
+        executionId: state.executionId,
+        purpose: 'materialization-receipt',
+        bytes: state.materializationReceipt.bytes,
+      }),
+    ])
+    preparationEvidence = Object.freeze({ executionPlan, materializationReceipt })
+  } catch (error) {
+    return await failBeforeActivation(prepared, state, error, 'failed', cleanupTimeoutMs)
+  }
+
   let acquired: Awaited<ReturnType<AgentCandidateExecutionClaimStore['tryClaim']>>
   try {
-    acquired = await options.claimStore.tryClaim(candidateExecutionClaim(prepared))
+    acquired = await options.claimStore.tryClaim(
+      candidateExecutionClaim(prepared, preparationEvidence),
+    )
   } catch (error) {
     return await failBeforeActivation(prepared, state, error, 'failed', cleanupTimeoutMs)
   }
@@ -381,13 +404,14 @@ export async function executePreparedAgentCandidate(
             state.trace.runId,
             settlementResult.settlement as SealedAgentCandidateModelSettlement,
           )
-          const taskOutcome = await persistVerifiedCandidateTaskOutcome(
-            state,
-            execution.finalCapture.taskOutcome!,
-            options.outputArtifacts,
-            protectedValues,
-            signal,
-          )
+          const { persistedCapture, taskOutcome } =
+            await persistVerifiedAgentCandidateExecutorCapture(
+              state,
+              execution.finalCapture,
+              options.outputArtifacts,
+              protectedValues,
+              signal,
+            )
           const benchmarkResult = await persistCandidateBenchmarkResult(
             state,
             capture.termination,
@@ -404,6 +428,7 @@ export async function executePreparedAgentCandidate(
             settlementResult.settlement as SealedAgentCandidateModelSettlement,
             {
               finalCapture: execution.finalCapture,
+              persistedCapture,
               modelSettlement,
               taskOutcome,
               benchmarkResult,
@@ -433,7 +458,7 @@ export async function executePreparedAgentCandidate(
       ? {
           schemaVersion: 1,
           status: 'succeeded',
-          usage: settlementResult.settlement.fixedUsage,
+          usage: settlementResult.settlement.usage,
           modelSettlement: result.artifacts.modelSettlement,
           taskOutcome: result.artifacts.taskOutcome,
           benchmarkResult: result.artifacts.benchmarkResult,
@@ -443,7 +468,7 @@ export async function executePreparedAgentCandidate(
           schemaVersion: 1,
           status: 'failed',
           failureClass,
-          usage: settlementResult.settlement.fixedUsage,
+          usage: settlementResult.settlement.usage,
           modelSettlement: modelSettlement.artifact,
           failureEvidence: await withinCandidateCleanupDeadline(
             () =>
@@ -499,14 +524,14 @@ type ExecutorOutcome =
       kind: 'capture'
       capture: AgentCandidateProtectedRunCapture
       termination: AgentCandidateTermination
-      finalCapture: AgentCandidateExecutorFinalCapture
+      finalCapture: SealedAgentCandidateExecutorFinalCapture
       processStopped: true
       error?: undefined
     }
   | {
       kind: 'timeout'
       termination: AgentCandidateTermination & { kind: 'timeout' }
-      finalCapture: AgentCandidateExecutorFinalCapture
+      finalCapture: SealedAgentCandidateExecutorFinalCapture
       processStopped: true
       error?: undefined
     }
@@ -514,7 +539,7 @@ type ExecutorOutcome =
       kind: 'error'
       error: unknown
       termination?: AgentCandidateTermination
-      finalCapture?: AgentCandidateExecutorFinalCapture
+      finalCapture?: SealedAgentCandidateExecutorFinalCapture
       processStopped: boolean
     }
 
@@ -585,11 +610,11 @@ async function runAndStopExecutor(
 
   if (!capture || timedOut) controller.abort(executionError)
   const stopReason = timedOut ? 'timeout' : capture ? 'completed' : 'failed'
-  let stopped: unknown
+  const cleanupDeadlineAtMs = Date.now() + cleanupTimeoutMs
   try {
-    stopped = await withinCandidateCleanupDeadline(
+    const stopped = await withinCandidateCleanupDeadline(
       () =>
-        executor.stopAndCapture(
+        executor.stop(
           {
             executionId: request.executionId,
             executionPlanDigest: request.executionPlan.value.digest,
@@ -601,16 +626,10 @@ async function runAndStopExecutor(
             deadlineAtMs,
           },
         ),
-      Date.now() + cleanupTimeoutMs,
+      cleanupDeadlineAtMs,
       'candidate process termination',
     )
-    if (
-      !stopped ||
-      typeof stopped !== 'object' ||
-      (stopped as { stopped?: unknown }).stopped !== true
-    ) {
-      throw new Error('candidate executor did not acknowledge exact process termination')
-    }
+    sealAgentCandidateExecutorStopAcknowledgement(stopped)
   } catch (stopError) {
     if (timer) clearTimeout(timer)
     controller.abort(stopError)
@@ -622,9 +641,26 @@ async function runAndStopExecutor(
     }
   }
 
-  let finalCapture: AgentCandidateExecutorFinalCapture
+  let finalCapture: SealedAgentCandidateExecutorFinalCapture
   try {
-    finalCapture = sealAgentCandidateExecutorFinalCapture(stopped)
+    finalCapture = sealAgentCandidateExecutorFinalCapture(
+      await withinCandidateCleanupDeadline(
+        () =>
+          executor.capture(
+            {
+              executionId: request.executionId,
+              executionPlanDigest: request.executionPlan.value.digest,
+            },
+            {
+              traceStore,
+              signal: new AbortController().signal,
+            },
+          ),
+        cleanupDeadlineAtMs,
+        'candidate final evidence capture',
+      ),
+      request.executionPlan.value.material.task.outcome,
+    )
   } catch (captureError) {
     if (timer) clearTimeout(timer)
     controller.abort(captureError)
@@ -747,7 +783,7 @@ async function failClaimedExecution(
           schemaVersion: 1,
           status: 'failed',
           failureClass,
-          usage: settled.settlement.fixedUsage,
+          usage: settled.settlement.usage,
           modelSettlement: modelSettlement.artifact,
           failureEvidence,
         },
diff --git a/src/candidate-execution/executor-capture-evidence.ts b/src/candidate-execution/executor-capture-evidence.ts
new file mode 100644
index 00000000..d674180b
--- /dev/null
+++ b/src/candidate-execution/executor-capture-evidence.ts
@@ -0,0 +1,168 @@
+import type {
+  AgentCandidateArtifactRef,
+  AgentCandidateTaskOutcomeSpec,
+  AgentCandidateTaskOutputSpec,
+  AgentCandidateWorkspaceManifestMaterial,
+  Sha256Digest,
+} from '@tangle-network/agent-interface'
+
+import { canonicalCandidateBytes } from './digest'
+import type { SealedAgentCandidateExecutorFinalCapture } from './executor-capture'
+import { persistCandidateOutputArtifact } from './output-artifacts'
+import type { AgentCandidateOutputArtifactPort } from './types'
+
+export type PersistedAgentCandidateTaskCapture =
+  | {
+      readonly kind: 'workspace'
+      readonly resultTree: string
+      readonly afterState: AgentCandidateWorkspaceManifestMaterial
+      readonly manifest: AgentCandidateArtifactRef
+      readonly archive: AgentCandidateArtifactRef
+      readonly gitDiff: AgentCandidateArtifactRef
+    }
+  | {
+      readonly kind: 'output'
+      readonly spec: AgentCandidateTaskOutputSpec
+      readonly artifact: AgentCandidateArtifactRef
+    }
+
+export interface PersistedAgentCandidateExecutorCapture {
+  readonly evidence: AgentCandidateArtifactRef
+  readonly taskOutcome?: PersistedAgentCandidateTaskCapture
+  readonly memoryAfter?: {
+    readonly afterState: AgentCandidateWorkspaceManifestMaterial
+    readonly manifest: AgentCandidateArtifactRef
+    readonly archive: AgentCandidateArtifactRef
+  }
+}
+
+/** Persist structurally valid post-stop bytes without claiming they passed outcome verification. */
+export async function persistAgentCandidateExecutorCapture(
+  identity: { executionId: string; executionPlanDigest: Sha256Digest },
+  expected: AgentCandidateTaskOutcomeSpec,
+  capture: SealedAgentCandidateExecutorFinalCapture,
+  outputArtifacts: AgentCandidateOutputArtifactPort,
+  signal?: AbortSignal,
+): Promise<PersistedAgentCandidateExecutorCapture> {
+  const executorEvidence = capture.evidence
+    ? await persistCandidateOutputArtifact(outputArtifacts, {
+        executionId: identity.executionId,
+        purpose: 'executor-capture',
+        bytes: capture.evidence,
+        signal,
+      })
+    : undefined
+  const taskOutcome = capture.taskOutcome
+    ? await persistTaskCapture(
+        identity.executionId,
+        expected,
+        capture.taskOutcome,
+        outputArtifacts,
+        signal,
+      )
+    : undefined
+  const memoryAfter = capture.memoryAfter
+    ? await persistWorkspaceCapture(
+        identity.executionId,
+        'memory-after-manifest',
+        'memory-after-archive',
+        capture.memoryAfter.afterState,
+        capture.memoryAfter.archive,
+        outputArtifacts,
+        signal,
+      )
+    : undefined
+  const bytes = canonicalCandidateBytes({
+    schemaVersion: 1,
+    kind: 'agent-candidate-executor-capture',
+    executionPlanDigest: identity.executionPlanDigest,
+    ...(executorEvidence ? { executorEvidence } : {}),
+    ...(taskOutcome ? { taskOutcome } : {}),
+    ...(memoryAfter ? { memoryAfter } : {}),
+  })
+  const evidence = await persistCandidateOutputArtifact(outputArtifacts, {
+    executionId: identity.executionId,
+    purpose: 'executor-capture',
+    bytes,
+    signal,
+  })
+  return Object.freeze({
+    evidence,
+    ...(taskOutcome ? { taskOutcome: Object.freeze(taskOutcome) } : {}),
+    ...(memoryAfter ? { memoryAfter: Object.freeze(memoryAfter) } : {}),
+  })
+}
+
+async function persistTaskCapture(
+  executionId: string,
+  expected: AgentCandidateTaskOutcomeSpec,
+  capture: NonNullable<SealedAgentCandidateExecutorFinalCapture['taskOutcome']>,
+  outputArtifacts: AgentCandidateOutputArtifactPort,
+  signal?: AbortSignal,
+): Promise<PersistedAgentCandidateTaskCapture> {
+  if (capture.kind === 'output') {
+    if (expected.kind !== 'output')
+      throw new Error('captured output does not match the signed task')
+    const artifact = await persistCandidateOutputArtifact(outputArtifacts, {
+      executionId,
+      purpose: 'task-output',
+      bytes: capture.bytes,
+      signal,
+    })
+    return {
+      kind: 'output',
+      spec: { mediaType: expected.mediaType, maxBytes: expected.maxBytes },
+      artifact,
+    }
+  }
+  if (expected.kind !== 'workspace') {
+    throw new Error('captured workspace does not match the signed task')
+  }
+  const workspace = await persistWorkspaceCapture(
+    executionId,
+    'task-manifest',
+    'task-archive',
+    capture.afterState,
+    capture.archive,
+    outputArtifacts,
+    signal,
+  )
+  const gitDiff = await persistCandidateOutputArtifact(outputArtifacts, {
+    executionId,
+    purpose: 'task-patch',
+    bytes: capture.gitDiff,
+    signal,
+  })
+  return { kind: 'workspace', resultTree: capture.resultTree, ...workspace, gitDiff }
+}
+
+async function persistWorkspaceCapture(
+  executionId: string,
+  manifestPurpose: 'task-manifest' | 'memory-after-manifest',
+  archivePurpose: 'task-archive' | 'memory-after-archive',
+  afterState: AgentCandidateWorkspaceManifestMaterial,
+  archiveBytes: Uint8Array,
+  outputArtifacts: AgentCandidateOutputArtifactPort,
+  signal?: AbortSignal,
+): Promise<{
+  afterState: AgentCandidateWorkspaceManifestMaterial
+  manifest: AgentCandidateArtifactRef
+  archive: AgentCandidateArtifactRef
+}> {
+  const manifestBytes = canonicalCandidateBytes(afterState)
+  const [manifest, archive] = await Promise.all([
+    persistCandidateOutputArtifact(outputArtifacts, {
+      executionId,
+      purpose: manifestPurpose,
+      bytes: manifestBytes,
+      signal,
+    }),
+    persistCandidateOutputArtifact(outputArtifacts, {
+      executionId,
+      purpose: archivePurpose,
+      bytes: archiveBytes,
+      signal,
+    }),
+  ])
+  return { afterState, manifest, archive }
+}
diff --git a/src/candidate-execution/executor-capture.ts b/src/candidate-execution/executor-capture.ts
index d8a68cee..1ee9e2d9 100644
--- a/src/candidate-execution/executor-capture.ts
+++ b/src/candidate-execution/executor-capture.ts
@@ -1,10 +1,17 @@
 import {
+  type AgentCandidateTaskOutcomeSpec,
   agentCandidateTerminationSchema,
   agentCandidateWorkspaceManifestMaterialSchema,
 } from '@tangle-network/agent-interface'
 import { assertExactObjectKeys as assertExactKeys } from './exact-object'
 import type { AgentCandidateExecutorFinalCapture, AgentCandidateProtectedRunCapture } from './types'
 
+const sealedFinalCaptureBrand: unique symbol = Symbol('sealedAgentCandidateExecutorFinalCapture')
+
+export type SealedAgentCandidateExecutorFinalCapture = AgentCandidateExecutorFinalCapture & {
+  readonly [sealedFinalCaptureBrand]: true
+}
+
 /** Validate and detach the only candidate-authored fields accepted from execution. */
 export function sealAgentCandidateProtectedRunCapture(
   value: unknown,
@@ -23,22 +30,40 @@ export function sealAgentCandidateProtectedRunCapture(
 /** Validate, detach, and freeze evaluator-owned evidence captured after process death. */
 export function sealAgentCandidateExecutorFinalCapture(
   value: unknown,
-): AgentCandidateExecutorFinalCapture {
+  expectedOutcome: AgentCandidateTaskOutcomeSpec,
+): SealedAgentCandidateExecutorFinalCapture {
   const capture = requireRecord(value, 'candidate final capture')
-  assertExactKeys(capture, ['stopped'], 'candidate final capture', ['taskOutcome', 'memoryAfter'])
-  if (capture.stopped !== true) {
-    throw new Error('candidate final capture does not prove process death')
-  }
+  assertExactKeys(capture, [], 'candidate final capture', [
+    'taskOutcome',
+    'memoryAfter',
+    'evidence',
+  ])
 
-  const taskOutcome = capture.taskOutcome ? sealTaskOutcomeCapture(capture.taskOutcome) : undefined
+  const taskOutcome = capture.taskOutcome
+    ? sealTaskOutcomeCapture(capture.taskOutcome, expectedOutcome)
+    : undefined
   const memoryAfter = capture.memoryAfter ? sealMemoryCapture(capture.memoryAfter) : undefined
+  if (capture.evidence !== undefined && !(capture.evidence instanceof Uint8Array)) {
+    throw new Error('candidate executor evidence must be a byte array')
+  }
+  const evidence = capture.evidence ? Uint8Array.from(capture.evidence) : undefined
   return Object.freeze({
-    stopped: true,
     ...(taskOutcome ? { taskOutcome } : {}),
     ...(memoryAfter ? { memoryAfter: Object.freeze(memoryAfter) } : {}),
+    ...(evidence ? { evidence } : {}),
+    [sealedFinalCaptureBrand]: true as const,
   })
 }
 
+/** Prove exact process death before any final evidence capture. */
+export function sealAgentCandidateExecutorStopAcknowledgement(value: unknown): void {
+  const capture = requireRecord(value, 'candidate stop acknowledgement')
+  assertExactKeys(capture, ['stopped'], 'candidate stop acknowledgement')
+  if (capture.stopped !== true) {
+    throw new Error('candidate stop acknowledgement does not prove process death')
+  }
+}
+
 function sealMemoryCapture(
   value: unknown,
 ): NonNullable<AgentCandidateExecutorFinalCapture['memoryAfter']> {
@@ -57,13 +82,33 @@ function sealMemoryCapture(
 
 function sealTaskOutcomeCapture(
   value: unknown,
+  expected: AgentCandidateTaskOutcomeSpec,
 ): NonNullable<AgentCandidateExecutorFinalCapture['taskOutcome']> {
   const capture = requireRecord(value, 'candidate task capture')
+  if (capture.kind === 'output') {
+    assertExactKeys(capture, ['kind', 'bytes'], 'candidate task output capture')
+    if (expected.kind !== 'output') {
+      throw new Error('candidate task output capture does not match the signed outcome kind')
+    }
+    if (!(capture.bytes instanceof Uint8Array)) {
+      throw new Error('candidate task output capture must contain a byte array')
+    }
+    if (capture.bytes.byteLength > expected.maxBytes) {
+      throw new Error(`candidate task output exceeds the frozen ${expected.maxBytes}-byte maximum`)
+    }
+    return Object.freeze({ kind: 'output', bytes: Uint8Array.from(capture.bytes) })
+  }
   assertExactKeys(
     capture,
-    ['resultTree', 'afterState', 'archive', 'gitDiff'],
+    ['kind', 'resultTree', 'afterState', 'archive', 'gitDiff'],
     'candidate task capture',
   )
+  if (capture.kind !== 'workspace') {
+    throw new Error('candidate task capture has an invalid outcome kind')
+  }
+  if (expected.kind !== 'workspace') {
+    throw new Error('candidate workspace capture does not match the signed outcome kind')
+  }
   if (
     typeof capture.resultTree !== 'string' ||
     !/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(capture.resultTree)
@@ -75,6 +120,7 @@ function sealTaskOutcomeCapture(
   }
   const afterState = agentCandidateWorkspaceManifestMaterialSchema.parse(capture.afterState)
   return Object.freeze({
+    kind: 'workspace',
     resultTree: capture.resultTree,
     afterState: Object.freeze(afterState),
     archive: Uint8Array.from(capture.archive),
diff --git a/src/candidate-execution/finalize.ts b/src/candidate-execution/finalize.ts
index 85eee32c..04142910 100644
--- a/src/candidate-execution/finalize.ts
+++ b/src/candidate-execution/finalize.ts
@@ -1,33 +1,21 @@
-import { mkdtemp, rm } from 'node:fs/promises'
-import { tmpdir } from 'node:os'
-import { join } from 'node:path'
-
 import type { Span, TraceStore } from '@tangle-network/agent-eval'
 import { isLlmSpan, REDACTION_VERSION } from '@tangle-network/agent-eval'
 import type {
   AgentCandidateBenchmarkResultEvidence,
+  AgentCandidateFixedSpend,
   AgentCandidateMemoryReceipt,
   AgentCandidateModelSettlementEvidence,
-  AgentCandidateRunReceiptV2,
-  AgentCandidateSpend,
+  AgentCandidateRunReceipt,
   AgentCandidateTermination,
 } from '@tangle-network/agent-interface'
 import {
-  agentCandidateRunReceiptV2Schema,
+  agentCandidateRunReceiptSchema,
   agentCandidateWorkspaceSnapshotEvidenceSchema,
 } from '@tangle-network/agent-interface'
 
-import { readMaterializedWorkspaceFiles } from './artifacts'
-import {
-  canonicalCandidateBytes,
-  canonicalCandidateDocument,
-  embeddedCandidateArtifact,
-  sha256Bytes,
-} from './digest'
-import {
-  sealAgentCandidateExecutorFinalCapture,
-  sealAgentCandidateProtectedRunCapture,
-} from './executor-capture'
+import { canonicalCandidateBytes, canonicalCandidateDocument, sha256Bytes } from './digest'
+import type { SealedAgentCandidateExecutorFinalCapture } from './executor-capture'
+import type { PersistedAgentCandidateExecutorCapture } from './executor-capture-evidence'
 import {
   assertTraceMatchesModelSettlement,
   type SealedAgentCandidateModelSettlement,
@@ -42,7 +30,6 @@ import {
   redactProtectedValue,
 } from './protected-redaction'
 import {
-  type AgentCandidateExecutorFinalCapture,
   type AgentCandidateOutputArtifactPort,
   type AgentCandidateProtectedRunCapture,
   type AgentCandidateRunFinalization,
@@ -51,7 +38,8 @@ import {
 } from './types'
 
 interface CandidateFinalizationEvidence {
-  finalCapture: AgentCandidateExecutorFinalCapture
+  finalCapture: SealedAgentCandidateExecutorFinalCapture
+  persistedCapture: PersistedAgentCandidateExecutorCapture
   modelSettlement: AgentCandidateModelSettlementEvidence & {
     artifact: import('@tangle-network/agent-interface').AgentCandidateArtifactRef
   }
@@ -76,11 +64,10 @@ export async function finalizeAgentCandidateRun(
   let termination: AgentCandidateTermination | undefined
   try {
     signal?.throwIfAborted()
-    const protectedCapture = sealAgentCandidateProtectedRunCapture(capture)
-    if (protectedCapture.executionId !== state.executionId) {
+    if (capture.executionId !== state.executionId) {
       throw new Error('protected capture execution id does not match the prepared execution')
     }
-    termination = protectedCapture.termination
+    termination = capture.termination
     if (
       termination.kind === 'timeout' &&
       termination.timeoutMs !== state.executionPlan.value.material.limits.timeoutMs
@@ -116,14 +103,11 @@ export async function finalizeAgentCandidateRun(
 
     const modelSpans = orderedSpans.filter(isLlmSpan)
     assertTraceMatchesModelSettlement(modelSpans, settlement)
-    const usage = settlement.usage
     enforceLimits(state, run.startedAt, run.endedAt, orderedSpans, settlement)
-    const finalCapture = sealAgentCandidateExecutorFinalCapture(evidence.finalCapture)
     const memory = await memoryReceipt(
       state,
-      finalCapture,
-      evidence.outputArtifacts,
-      protectedValues,
+      evidence.finalCapture,
+      evidence.persistedCapture.memoryAfter,
       signal,
     )
 
@@ -171,24 +155,22 @@ export async function finalizeAgentCandidateRun(
         orderedArtifacts.length,
       modelCallCount: modelSpans.length,
     }
-    const document = canonicalCandidateDocument<AgentCandidateRunReceiptV2>({
-      schemaVersion: 2,
+    const document = canonicalCandidateDocument<AgentCandidateRunReceipt>({
+      schemaVersion: 1,
       kind: 'agent-candidate-run',
       digestAlgorithm: 'rfc8785-sha256',
       bundleDigest: state.bundle.digest,
       materializationReceiptDigest: state.materializationReceipt.digest,
       executionPlanDigest: state.executionPlan.value.digest,
       memory,
-      usage,
-      modelUsage: { resolved: state.resolvedModel, usage },
       trace,
       termination,
-      fixedUsage: settlement.fixedUsage,
+      executorCapture: evidence.persistedCapture.evidence,
       modelSettlement: evidence.modelSettlement,
       taskOutcome: evidence.taskOutcome.evidence,
       benchmarkResult: evidence.benchmarkResult,
     })
-    agentCandidateRunReceiptV2Schema.parse(document.value)
+    agentCandidateRunReceiptSchema.parse(document.value)
     const runReceipt = await persistCandidateOutputArtifact(evidence.outputArtifacts, {
       executionId: state.executionId,
       purpose: 'run-receipt',
@@ -200,6 +182,7 @@ export async function finalizeAgentCandidateRun(
       receipt: document,
       artifacts: {
         modelSettlement: evidence.modelSettlement.artifact,
+        executorCapture: evidence.persistedCapture.evidence,
         taskOutcome: evidence.taskOutcome.evidence.artifact,
         benchmarkResult: evidence.benchmarkResult.artifact,
         runReceipt,
@@ -258,75 +241,37 @@ function enforceLimits(
   for (const [actual, limit, label] of checks) {
     if (actual > limit) throw new Error(`protected ${label} ${actual} exceeds ${limit}`)
   }
-  if (settlement.costUsdNanos > usdToNanos(limits.maxCostUsd, 'frozen maxCostUsd')) {
-    throw new Error(`protected cost USD ${usage.costUsd} exceeds ${limits.maxCostUsd}`)
+  if (usage.costUsdNanos > usdToNanos(limits.maxCostUsd, 'frozen maxCostUsd')) {
+    throw new Error(
+      `protected cost USD ${usage.costUsdNanos / 1_000_000_000} exceeds ${limits.maxCostUsd}`,
+    )
   }
 }
 
 async function memoryReceipt(
   state: PreparedCandidateState,
-  capture: AgentCandidateExecutorFinalCapture,
-  outputArtifacts: AgentCandidateOutputArtifactPort,
-  protectedValues: readonly string[],
+  capture: SealedAgentCandidateExecutorFinalCapture,
+  persisted: PersistedAgentCandidateExecutorCapture['memoryAfter'],
   signal?: AbortSignal,
 ): Promise<AgentCandidateMemoryReceipt> {
   signal?.throwIfAborted()
   if (state.memory.mode === 'disabled') {
-    if (capture.memoryAfter !== undefined) {
+    if (capture.memoryAfter !== undefined || persisted !== undefined) {
       throw new Error('disabled memory cannot return an after-state')
     }
     return { mode: 'disabled' }
   }
   if (!capture.memoryAfter) throw new Error('isolated memory is missing its protected after-state')
+  if (!persisted) throw new Error('isolated memory is missing persisted capture evidence')
   const afterState = capture.memoryAfter.afterState
-  const archive = Uint8Array.from(capture.memoryAfter.archive)
-  if (archive.byteLength === 0) throw new Error('isolated memory archive cannot be empty')
   const manifestBytes = canonicalCandidateBytes(afterState)
-  assertNoProtectedBytes(manifestBytes, protectedValues)
-  assertNoProtectedBytes(archive, protectedValues)
-  const provisionalSnapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 1,
-    kind: 'agent-candidate-workspace-snapshot',
-    digest: sha256Bytes(manifestBytes),
-    material: afterState,
-    manifest: embeddedCandidateArtifact(manifestBytes),
-    archive: embeddedCandidateArtifact(archive),
-  })
-  const root = await mkdtemp(join(tmpdir(), 'agent-candidate-memory-after-'))
-  try {
-    await state.ports.workspaces.materialize({
-      role: 'memory',
-      snapshot: provisionalSnapshot,
-      archive: Uint8Array.from(archive),
-      destination: root,
-    })
-    const files = await readMaterializedWorkspaceFiles(root, afterState)
-    for (const file of files) assertNoProtectedBytes(file.bytes, protectedValues)
-    signal?.throwIfAborted()
-  } finally {
-    await rm(root, { recursive: true, force: true })
-  }
-  const [manifest, archiveRef] = await Promise.all([
-    persistCandidateOutputArtifact(outputArtifacts, {
-      executionId: state.executionId,
-      purpose: 'memory-after-manifest',
-      bytes: manifestBytes,
-      signal,
-    }),
-    persistCandidateOutputArtifact(outputArtifacts, {
-      executionId: state.executionId,
-      purpose: 'memory-after-archive',
-      bytes: archive,
-      signal,
-    }),
-  ])
   const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
     schemaVersion: 1,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: afterState,
-    manifest,
-    archive: archiveRef,
+    manifest: persisted.manifest,
+    archive: persisted.archive,
   })
   return {
     mode: 'isolated',
@@ -342,7 +287,7 @@ export function failedAgentCandidateRun(
   state: PreparedCandidateState,
   reason: string,
   termination?: AgentCandidateTermination,
-  usage: AgentCandidateSpend | null = null,
+  usage: AgentCandidateFixedSpend | null = null,
 ): AgentCandidateRunFinalization & { succeeded: false } {
   return {
     succeeded: false,
diff --git a/src/candidate-execution/git-materialize.ts b/src/candidate-execution/git-materialize.ts
index 249b894a..6ce82eb5 100644
--- a/src/candidate-execution/git-materialize.ts
+++ b/src/candidate-execution/git-materialize.ts
@@ -7,7 +7,7 @@ import type {
   AgentCandidateCode,
   AgentCandidateGitHubRepository,
   AgentCandidateGitHubResource,
-  AgentCandidateWorkspaceManifestMaterialV1,
+  AgentCandidateWorkspaceManifestMaterial,
 } from '@tangle-network/agent-interface'
 
 import { verifyBytes } from './artifacts'
@@ -121,7 +121,7 @@ export async function verifyTaskOutcomePatch(input: {
   baseTree: string
   resultTree: string
   patch: Uint8Array
-  afterState: AgentCandidateWorkspaceManifestMaterialV1
+  afterState: AgentCandidateWorkspaceManifestMaterial
 }): Promise<{ resultTree: string; resultCommit: string }> {
   const repositoryRoot = resolve(input.repositoryRoot)
   await verifyTaskCheckout(repositoryRoot, input)
@@ -319,7 +319,7 @@ async function workspaceManifestFromGitTree(
   repositoryRoot: string,
   tree: string,
   environment: Record<string, string>,
-): Promise<AgentCandidateWorkspaceManifestMaterialV1> {
+): Promise<AgentCandidateWorkspaceManifestMaterial> {
   const files = (await readCandidateGitTreeFiles(repositoryRoot, tree, environment)).map(
     ({ path, mode, bytes }) => ({
       path,
diff --git a/src/candidate-execution/index.ts b/src/candidate-execution/index.ts
index 0b363fc0..d4fa1731 100644
--- a/src/candidate-execution/index.ts
+++ b/src/candidate-execution/index.ts
@@ -25,7 +25,6 @@ export {
   type AgentCandidateExecutionStageResult,
   type AgentCandidateExecutionTerminalRecord,
   type AgentCandidateExecutionTerminalResult,
-  type AgentCandidateExecutionUsage,
   type AgentCandidateRetryRejection,
   InMemoryAgentCandidateExecutionClaimStore,
 } from './claim'
@@ -74,7 +73,6 @@ export {
   type AgentCandidateExecutorFinalCapture,
   type AgentCandidateExecutorMemoryCapture,
   type AgentCandidateExecutorPort,
-  type AgentCandidateExecutorProfileFile,
   type AgentCandidateExecutorRequest,
   type AgentCandidateExecutorStopRequest,
   type AgentCandidateExecutorTaskOutcomeCapture,
@@ -87,7 +85,6 @@ export {
   type AgentCandidateOutputArtifactPort,
   type AgentCandidateOutputPurpose,
   type AgentCandidateProtectedModelActivation,
-  type AgentCandidateProtectedModelCall,
   type AgentCandidateProtectedModelReservation,
   type AgentCandidateProtectedModelSettlement,
   type AgentCandidateProtectedRunCapture,
@@ -107,7 +104,7 @@ export {
   type VerifiedAgentCandidate,
   type VerifiedAgentCandidateTaskOutcome,
 } from './types'
-export { verifyAgentCandidateBundle } from './verify'
+export { AGENT_CANDIDATE_EXECUTION_SUPPORT, verifyAgentCandidateBundle } from './verify'
 export {
   type AgentCandidateWorkspaceArchiveLimits,
   type CaptureAgentCandidateWorkspaceOptions,
diff --git a/src/candidate-execution/model-settlement.ts b/src/candidate-execution/model-settlement.ts
index e188f4d2..5ffba8ce 100644
--- a/src/candidate-execution/model-settlement.ts
+++ b/src/candidate-execution/model-settlement.ts
@@ -1,19 +1,16 @@
 import { isLlmSpan, type LlmSpan, type TraceStore } from '@tangle-network/agent-eval'
-import type { AgentCandidateSpend } from '@tangle-network/agent-interface'
-import type { AgentCandidateExecutionUsage } from './claim'
-import { assertExactObjectKeys } from './exact-object'
 import type {
-  AgentCandidateProtectedModelCall,
-  AgentCandidateProtectedModelSettlement,
-} from './types'
+  AgentCandidateFixedSpend,
+  AgentCandidateModelSettlementCall,
+} from '@tangle-network/agent-interface'
+import { assertExactObjectKeys } from './exact-object'
+import type { AgentCandidateProtectedModelSettlement } from './types'
 
 const USD_NANOS = 1_000_000_000
 
 export interface SealedAgentCandidateModelSettlement {
   readonly value: AgentCandidateProtectedModelSettlement
-  readonly usage: AgentCandidateSpend
-  readonly fixedUsage: AgentCandidateExecutionUsage
-  readonly costUsdNanos: number
+  readonly usage: AgentCandidateFixedSpend
 }
 
 /** Validate and detach the evaluator gateway's terminal, revoked call ledger. */
@@ -43,7 +40,6 @@ export function sealAgentCandidateModelSettlement(
   let outputTokens = 0
   let cachedInputTokens = 0
   let reasoningTokens = 0
-  let hasCachedInput = false
   let costUsdNanos = 0
   const calls = settlement.calls.map((source, index) => {
     assertExactObjectKeys(
@@ -96,7 +92,6 @@ export function sealAgentCandidateModelSettlement(
       source.cachedInputTokens,
       'cached input token total',
     )
-    hasCachedInput = true
     assertCount(source.reasoningTokens, `model settlement call ${index} reasoningTokens`)
     reasoningTokens = safeAdd(reasoningTokens, source.reasoningTokens, 'reasoning token total')
     assertCount(source.costUsdNanos, `model settlement call ${index} costUsdNanos`)
@@ -107,13 +102,6 @@ export function sealAgentCandidateModelSettlement(
   })
 
   const usage = Object.freeze({
-    costUsd: costUsdNanos / USD_NANOS,
-    inputTokens,
-    outputTokens,
-    ...(hasCachedInput ? { cachedInputTokens } : {}),
-    modelCalls: calls.length,
-  })
-  const fixedUsage = Object.freeze({
     costUsdNanos,
     inputTokens,
     outputTokens,
@@ -129,8 +117,6 @@ export function sealAgentCandidateModelSettlement(
       calls: Object.freeze(calls),
     }),
     usage,
-    fixedUsage,
-    costUsdNanos,
   })
 }
 
@@ -213,7 +199,7 @@ export function usdToNanos(value: number, label: string): number {
   return nanos
 }
 
-function assertTraceCall(span: LlmSpan, call: AgentCandidateProtectedModelCall): void {
+function assertTraceCall(span: LlmSpan, call: AgentCandidateModelSettlementCall): void {
   if (span.model !== call.model) {
     throw new Error(`protected trace span ${span.spanId} model does not match model ledger`)
   }
diff --git a/src/candidate-execution/outcome-evidence.ts b/src/candidate-execution/outcome-evidence.ts
index a4fd66e2..05c33973 100644
--- a/src/candidate-execution/outcome-evidence.ts
+++ b/src/candidate-execution/outcome-evidence.ts
@@ -9,6 +9,10 @@ import {
   type AgentCandidateModelSettlementEvidence,
   type AgentCandidateResolvedModel,
   type AgentCandidateTaskOutcomeEvidence,
+  type AgentCandidateTaskOutcomeMaterial,
+  type AgentCandidateTaskOutcomeSpec,
+  type AgentCandidateTaskOutputSpec,
+  type AgentCandidateTaskRepository,
   type AgentCandidateTermination,
   agentCandidateBenchmarkResultEvidenceSchema,
   agentCandidateModelSettlementEvidenceSchema,
@@ -25,6 +29,12 @@ import {
   immutableCandidateValue,
   sha256Bytes,
 } from './digest'
+import type { SealedAgentCandidateExecutorFinalCapture } from './executor-capture'
+import {
+  type PersistedAgentCandidateExecutorCapture,
+  type PersistedAgentCandidateTaskCapture,
+  persistAgentCandidateExecutorCapture,
+} from './executor-capture-evidence'
 import { verifyTaskOutcomePatch } from './git-materialize'
 import type { SealedAgentCandidateModelSettlement } from './model-settlement'
 import { persistCandidateOutputArtifact } from './output-artifacts'
@@ -46,7 +56,7 @@ export type PersistedAgentCandidateBenchmarkResult = AgentCandidateBenchmarkResu
   artifact: AgentCandidateArtifactRef
 }
 
-/** Persist the closed evaluator model ledger as canonical V2 receipt evidence. */
+/** Persist the closed evaluator model ledger as canonical receipt evidence. */
 export async function persistCandidateModelSettlement(
   state: PreparedCandidateState,
   settlement: SealedAgentCandidateModelSettlement,
@@ -74,7 +84,7 @@ export async function persistCandidateModelSettlementEvidence(
   outputArtifacts: AgentCandidateOutputArtifactPort,
 ): Promise<PersistedAgentCandidateModelSettlement> {
   const material = {
-    schemaVersion: 2 as const,
+    schemaVersion: 1 as const,
     kind: 'agent-candidate-model-settlement-material' as const,
     executionPlanDigest: identity.executionPlanDigest,
     preparationId: settlement.value.preparationId,
@@ -95,7 +105,7 @@ export async function persistCandidateModelSettlementEvidence(
       reasoningTokens: call.reasoningTokens ?? 0,
       costUsdNanos: call.costUsdNanos,
     })),
-    usage: settlement.fixedUsage,
+    usage: settlement.usage,
   }
   const bytes = canonicalCandidateBytes(material)
   const digest = sha256Bytes(bytes)
@@ -115,22 +125,93 @@ export async function persistCandidateModelSettlementEvidence(
   ) as PersistedAgentCandidateModelSettlement
 }
 
-/** Recompute the result tree from the patch, then persist its exact task evidence. */
-export async function persistVerifiedCandidateTaskOutcome(
+/** Verify once, persist the sealed capture once, and bind the resulting task evidence. */
+export async function persistVerifiedAgentCandidateExecutorCapture(
   state: PreparedCandidateState,
-  capture: AgentCandidateExecutorTaskOutcomeCapture,
+  capture: SealedAgentCandidateExecutorFinalCapture,
   outputArtifacts: AgentCandidateOutputArtifactPort,
   protectedValues: readonly string[],
   signal?: AbortSignal,
-): Promise<VerifiedAgentCandidateTaskOutcome> {
+): Promise<{
+  persistedCapture: PersistedAgentCandidateExecutorCapture
+  taskOutcome: VerifiedAgentCandidateTaskOutcome
+}> {
   signal?.throwIfAborted()
+  const taskCapture = capture.taskOutcome
+  if (!taskCapture) throw new Error('candidate final capture is missing the task outcome')
+  const expected = state.executionPlan.value.material.task.outcome
+  if (taskCapture.kind !== expected.kind) {
+    throw new Error(
+      `candidate captured ${taskCapture.kind} outcome does not match expected ${expected.kind} outcome`,
+    )
+  }
+  const verified =
+    taskCapture.kind === 'output' && expected.kind === 'output'
+      ? verifyOutputTaskCapture(taskCapture.bytes, expected, protectedValues)
+      : taskCapture.kind === 'workspace' && expected.kind === 'workspace'
+        ? await verifyWorkspaceTaskCapture(state, taskCapture, protectedValues, signal)
+        : undefined
+  if (!verified) throw new Error('candidate task outcome kind could not be narrowed')
+  await verifyMemoryCapture(state, capture, protectedValues, signal)
+
+  const persistedCapture = await persistAgentCandidateExecutorCapture(
+    {
+      executionId: state.executionId,
+      executionPlanDigest: state.executionPlan.value.digest,
+    },
+    expected,
+    capture,
+    outputArtifacts,
+    signal,
+  )
+  const persisted = persistedCapture.taskOutcome
+  if (!persisted || persisted.kind !== verified.kind) {
+    throw new Error('persisted task capture kind changed')
+  }
+  const taskOutcome =
+    verified.kind === 'output' && persisted.kind === 'output'
+      ? await persistVerifiedOutputTaskOutcome(state, verified, persisted, outputArtifacts, signal)
+      : verified.kind === 'workspace' && persisted.kind === 'workspace'
+        ? await persistVerifiedWorkspaceTaskOutcome(
+            state,
+            verified,
+            persisted,
+            outputArtifacts,
+            signal,
+          )
+        : undefined
+  if (!taskOutcome) throw new Error('persisted task outcome kind could not be narrowed')
+  return Object.freeze({ persistedCapture, taskOutcome })
+}
+
+interface VerifiedWorkspaceTaskCapture {
+  readonly kind: 'workspace'
+  readonly patch: Uint8Array
+  readonly archive: Uint8Array
+  readonly afterState: Extract<
+    AgentCandidateExecutorTaskOutcomeCapture,
+    { kind: 'workspace' }
+  >['afterState']
+  readonly manifestBytes: Uint8Array
+  readonly resultCommit: string
+  readonly resultTree: string
+  readonly repository: AgentCandidateTaskRepository
+}
+
+async function verifyWorkspaceTaskCapture(
+  state: PreparedCandidateState,
+  capture: Extract<AgentCandidateExecutorTaskOutcomeCapture, { kind: 'workspace' }>,
+  protectedValues: readonly string[],
+  signal?: AbortSignal,
+): Promise<VerifiedWorkspaceTaskCapture> {
+  const repository = state.executionPlan.value.material.task.repository
+  if (!repository) throw new Error('workspace task outcome is missing repository identity')
   const patch = Uint8Array.from(capture.gitDiff)
   const archive = Uint8Array.from(capture.archive)
   if (archive.byteLength === 0) throw new Error('candidate task archive cannot be empty')
   assertNoProtectedBytes(patch, protectedValues)
   assertNoProtectedBytes(archive, protectedValues)
   const afterState = immutableCandidateValue(capture.afterState)
-  const repository = state.executionPlan.value.material.task.repository
   const verified = await verifyTaskOutcomePatch({
     repositoryRoot: state.roots.staging.taskRoot,
     baseCommit: repository.baseCommit,
@@ -152,56 +233,142 @@ export async function persistVerifiedCandidateTaskOutcome(
   })
   await verifyTaskOutcomeArchive(state, provisionalSnapshot, archive, protectedValues)
   signal?.throwIfAborted()
-  const [manifest, archiveRef, gitDiff] = await Promise.all([
-    persistCandidateOutputArtifact(outputArtifacts, {
-      executionId: state.executionId,
-      purpose: 'task-manifest',
-      bytes: manifestBytes,
-      signal,
-    }),
-    persistCandidateOutputArtifact(outputArtifacts, {
-      executionId: state.executionId,
-      purpose: 'task-archive',
-      bytes: archive,
-      signal,
-    }),
-    persistCandidateOutputArtifact(outputArtifacts, {
-      executionId: state.executionId,
-      purpose: 'task-patch',
-      bytes: patch,
-      signal,
-    }),
-  ])
+  return {
+    kind: 'workspace',
+    patch,
+    archive,
+    afterState,
+    manifestBytes,
+    resultCommit: verified.resultCommit,
+    resultTree: verified.resultTree,
+    repository,
+  }
+}
+
+async function persistVerifiedWorkspaceTaskOutcome(
+  state: PreparedCandidateState,
+  verified: VerifiedWorkspaceTaskCapture,
+  persisted: Extract<PersistedAgentCandidateTaskCapture, { kind: 'workspace' }>,
+  outputArtifacts: AgentCandidateOutputArtifactPort,
+  signal?: AbortSignal,
+): Promise<VerifiedAgentCandidateTaskOutcome> {
+  const { afterState, manifestBytes, patch, repository } = verified
   const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
     schemaVersion: 1,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: afterState,
-    manifest,
-    archive: archiveRef,
+    manifest: persisted.manifest,
+    archive: persisted.archive,
   })
   const material = {
     schemaVersion: 1 as const,
     kind: 'agent-candidate-task-outcome-material' as const,
     executionPlanDigest: state.executionPlan.value.digest,
-    baseRepository: {
-      identity: repository.identity,
-      rootIdentity: repository.rootIdentity,
-      commit: repository.baseCommit,
-      tree: repository.baseTree,
+    outcome: {
+      kind: 'workspace' as const,
+      baseRepository: {
+        identity: repository.identity,
+        rootIdentity: repository.rootIdentity,
+        commit: repository.baseCommit,
+        tree: repository.baseTree,
+      },
+      resultRepository: {
+        identity: repository.identity,
+        rootIdentity: repository.rootIdentity,
+        commit: verified.resultCommit,
+        tree: verified.resultTree,
+      },
+      afterState: snapshot,
+      gitDiff: {
+        format: 'git-diff-binary' as const,
+        artifact: persisted.gitDiff,
+      },
     },
-    resultRepository: {
-      identity: repository.identity,
-      rootIdentity: repository.rootIdentity,
-      commit: verified.resultCommit,
-      tree: verified.resultTree,
+  } satisfies AgentCandidateTaskOutcomeMaterial
+  const evidence = await persistTaskOutcomeEvidence(state, material, outputArtifacts, signal)
+  const storedPatch = Uint8Array.from(patch)
+  return Object.freeze({
+    kind: 'workspace' as const,
+    evidence,
+    get patch(): Uint8Array {
+      return Uint8Array.from(storedPatch)
     },
-    afterState: snapshot,
-    gitDiff: {
-      format: 'git-diff-binary' as const,
-      artifact: gitDiff,
+    [verifiedTaskOutcomeBrand]: true as const,
+  })
+}
+
+interface VerifiedOutputTaskCapture {
+  readonly kind: 'output'
+  readonly bytes: Uint8Array
+  readonly spec: AgentCandidateTaskOutputSpec
+}
+
+function verifyOutputTaskCapture(
+  capturedBytes: Uint8Array,
+  expected: Extract<AgentCandidateTaskOutcomeSpec, { kind: 'output' }>,
+  protectedValues: readonly string[],
+): VerifiedOutputTaskCapture {
+  if (capturedBytes.byteLength === 0) throw new Error('candidate task output cannot be empty')
+  if (capturedBytes.byteLength > expected.maxBytes) {
+    throw new Error(`candidate task output exceeds the frozen ${expected.maxBytes}-byte maximum`)
+  }
+  const output = Uint8Array.from(capturedBytes)
+  assertNoProtectedBytes(output, protectedValues)
+  return {
+    kind: 'output',
+    bytes: output,
+    spec: { mediaType: expected.mediaType, maxBytes: expected.maxBytes },
+  }
+}
+
+async function persistVerifiedOutputTaskOutcome(
+  state: PreparedCandidateState,
+  verified: VerifiedOutputTaskCapture,
+  persisted: Extract<PersistedAgentCandidateTaskCapture, { kind: 'output' }>,
+  outputArtifacts: AgentCandidateOutputArtifactPort,
+  signal?: AbortSignal,
+): Promise<VerifiedAgentCandidateTaskOutcome> {
+  if (
+    persisted.spec.mediaType !== verified.spec.mediaType ||
+    persisted.spec.maxBytes !== verified.spec.maxBytes
+  ) {
+    throw new Error('persisted task output changed the signed media constraints')
+  }
+  const material = {
+    schemaVersion: 1 as const,
+    kind: 'agent-candidate-task-outcome-material' as const,
+    executionPlanDigest: state.executionPlan.value.digest,
+    outcome: {
+      kind: 'output' as const,
+      spec: verified.spec,
+      artifact: persisted.artifact,
     },
+  } satisfies AgentCandidateTaskOutcomeMaterial
+  const evidence = await persistTaskOutcomeEvidence(state, material, outputArtifacts, signal)
+  const storedOutput = Uint8Array.from(verified.bytes)
+  return Object.freeze({
+    kind: 'output' as const,
+    evidence,
+    spec: immutableCandidateValue(verified.spec),
+    get bytes(): Uint8Array {
+      return Uint8Array.from(storedOutput)
+    },
+    [verifiedTaskOutcomeBrand]: true as const,
+  })
+}
+
+async function persistTaskOutcomeEvidence<Material extends AgentCandidateTaskOutcomeMaterial>(
+  state: PreparedCandidateState,
+  material: Material,
+  outputArtifacts: AgentCandidateOutputArtifactPort,
+  signal?: AbortSignal,
+): Promise<
+  Omit<AgentCandidateTaskOutcomeEvidence, 'material'> & {
+    material: Material
+    artifact: AgentCandidateArtifactRef
   }
+> {
   const bytes = canonicalCandidateBytes(material)
   const digest = sha256Bytes(bytes)
   const artifact = await persistCandidateOutputArtifact(outputArtifacts, {
@@ -210,7 +377,7 @@ export async function persistVerifiedCandidateTaskOutcome(
     bytes,
     signal,
   })
-  const evidence = immutableCandidateValue(
+  return immutableCandidateValue(
     agentCandidateTaskOutcomeEvidenceSchema.parse({
       schemaVersion: 1,
       kind: 'agent-candidate-task-outcome',
@@ -218,15 +385,10 @@ export async function persistVerifiedCandidateTaskOutcome(
       material,
       artifact,
     }),
-  ) as AgentCandidateTaskOutcomeEvidence & { artifact: AgentCandidateArtifactRef }
-  const storedPatch = Uint8Array.from(patch)
-  return Object.freeze({
-    evidence,
-    get patch(): Uint8Array {
-      return Uint8Array.from(storedPatch)
-    },
-    [verifiedTaskOutcomeBrand]: true as const,
-  })
+  ) as Omit<AgentCandidateTaskOutcomeEvidence, 'material'> & {
+    material: Material
+    artifact: AgentCandidateArtifactRef
+  }
 }
 
 /** Grade only a runtime-verified outcome and persist both raw and normalized evidence. */
@@ -322,6 +484,46 @@ async function verifyTaskOutcomeArchive(
   }
 }
 
+async function verifyMemoryCapture(
+  state: PreparedCandidateState,
+  capture: SealedAgentCandidateExecutorFinalCapture,
+  protectedValues: readonly string[],
+  signal?: AbortSignal,
+): Promise<void> {
+  if (state.memory.mode === 'disabled') {
+    if (capture.memoryAfter) throw new Error('disabled memory cannot return an after-state')
+    return
+  }
+  const memory = capture.memoryAfter
+  if (!memory) throw new Error('isolated memory is missing its protected after-state')
+  if (memory.archive.byteLength === 0) throw new Error('isolated memory archive cannot be empty')
+  const manifestBytes = canonicalCandidateBytes(memory.afterState)
+  assertNoProtectedBytes(manifestBytes, protectedValues)
+  assertNoProtectedBytes(memory.archive, protectedValues)
+  const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
+    schemaVersion: 1,
+    kind: 'agent-candidate-workspace-snapshot',
+    digest: sha256Bytes(manifestBytes),
+    material: memory.afterState,
+    manifest: embeddedCandidateArtifact(manifestBytes),
+    archive: embeddedCandidateArtifact(memory.archive),
+  })
+  const root = await mkdtemp(join(tmpdir(), 'agent-candidate-memory-after-'))
+  try {
+    await state.ports.workspaces.materialize({
+      role: 'memory',
+      snapshot,
+      archive: Uint8Array.from(memory.archive),
+      destination: root,
+    })
+    const files = await readMaterializedWorkspaceFiles(root, memory.afterState)
+    for (const file of files) assertNoProtectedBytes(file.bytes, protectedValues)
+    signal?.throwIfAborted()
+  } finally {
+    await rm(root, { recursive: true, force: true })
+  }
+}
+
 function normalizeEvaluation(
   evaluation: BenchmarkEvaluation,
   termination: AgentCandidateTermination,
diff --git a/src/candidate-execution/prepare.ts b/src/candidate-execution/prepare.ts
index 3579e436..845ada76 100644
--- a/src/candidate-execution/prepare.ts
+++ b/src/candidate-execution/prepare.ts
@@ -7,11 +7,10 @@ import type {
   AgentCandidateEffectiveMemory,
   AgentCandidateExecutionLimits,
   AgentCandidateExecutionPlanEvidence,
-  AgentCandidateExecutionPlanMaterialV1,
+  AgentCandidateExecutionPlanMaterial,
   AgentCandidateMaterializationReceipt,
   AgentCandidateModelAccessNetwork,
   AgentCandidateResolvedModel,
-  HarnessType,
 } from '@tangle-network/agent-interface'
 import {
   agentCandidateContainerSchema,
@@ -20,12 +19,12 @@ import {
   agentCandidateExecutionPlanMaterialSchema,
   agentCandidateMaterializationReceiptSchema,
   agentCandidateModelAccessNetworkSchema,
+  agentCandidateTaskOutcomeSpecSchema,
   agentCandidateWorkspaceSnapshotEvidenceSchema,
   sha256DigestSchema,
 } from '@tangle-network/agent-interface'
 import {
   applyAgentCandidateWorkspacePlan,
-  type HarnessId,
   materializeCandidateProfile,
 } from '@tangle-network/agent-profile-materialize'
 
@@ -54,7 +53,7 @@ import { candidateExecutionOwnerWindowMs } from './execution-window'
 import { verifyTaskCheckout } from './git-materialize'
 import { sealAgentCandidateModelSettlement, usdToNanos } from './model-settlement'
 import { createPreparedCandidateExecution } from './prepared-state'
-import { assertCandidateProfileExecutionSupport } from './profile'
+import { candidateMaterializerHarness, createAgentCandidateProfileActivation } from './profile'
 import {
   type AgentCandidateExecutionPorts,
   type AgentCandidateTaskExecution,
@@ -70,21 +69,6 @@ import {
   verifiedResourceTextByDigest,
 } from './verify'
 
-const MATERIALIZER_HARNESSES = new Set<HarnessType>([
-  'claude-code',
-  'claude',
-  'claudish',
-  'nanoclaw',
-  'codex',
-  'opencode',
-  'kimi-code',
-  'kimi',
-  'pi',
-  'gemini',
-  'hermes',
-  'openclaw',
-])
-
 const MIN_RESERVATION_TTL_MS = 15 * 60_000
 const PREPARED_HOLD_MARGIN_MS = 5 * 60_000
 
@@ -105,8 +89,7 @@ export async function prepareAgentCandidateExecution(
   const verifiedState = getVerifiedCandidateState(candidate)
   assertSameVerificationPorts(verifiedState.ports, ports)
   const bundle = candidate.bundle
-  assertCandidateProfileExecutionSupport(bundle.profile)
-  const harness = materializerHarness(bundle.execution.harness)
+  const harness = candidateMaterializerHarness(bundle.execution.harness)
   assertTaskInput(task, bundle.execution.instructionDelivery)
   const resultTimeoutMs = candidateResultTimeout(options.resultTimeoutMs, task.limits.timeoutMs)
   const ownerWindowMs = candidateExecutionOwnerWindowMs(
@@ -133,7 +116,9 @@ export async function prepareAgentCandidateExecution(
   await verifyMaterializedWorkspace(task.stagingRoots.taskRoot, task.workspace.material, {
     ignoredProtectedRootEntries: ['.git', '.sidecar'],
   })
-  await verifyTaskCheckout(task.stagingRoots.taskRoot, task.repository)
+  if (task.repository) {
+    await verifyTaskCheckout(task.stagingRoots.taskRoot, task.repository)
+  }
   const taskExecutorFiles = await readMaterializedWorkspaceFiles(
     task.stagingRoots.taskRoot,
     task.workspace.material,
@@ -142,7 +127,7 @@ export async function prepareAgentCandidateExecution(
 
   let candidateArchive: Uint8Array | undefined
   let candidateExecutorFiles:
-    | ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>
+    | ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>
     | undefined
   if (bundle.execution.workspace) {
     if (!task.stagingRoots.candidateRoot || !task.executionRoots.candidateRoot) {
@@ -191,6 +176,10 @@ export async function prepareAgentCandidateExecution(
   ) {
     throw new Error('profile materializer did not capture exact canonical plan bytes')
   }
+  const profileActivation = createAgentCandidateProfileActivation(
+    profileWorkspacePlan,
+    profileApplication.profilePlan,
+  )
 
   const container = await resolveContainer(candidate, task, ports)
   const resolvedModel = await resolveModel(candidate, task, ports)
@@ -227,14 +216,6 @@ export async function prepareAgentCandidateExecution(
       cleanupTimeoutMs,
     )
     const memory = preparedMemory.value
-    const knowledge = bundle.knowledge
-      ? {
-          snapshotId: bundle.knowledge.snapshotId,
-          manifestDigest: bundle.knowledge.manifest.sha256,
-          manifest: await verifiedArtifactBytes(candidate, bundle.knowledge.manifest),
-        }
-      : undefined
-
     const baseLaunch = buildLaunch(candidate, task, profileApplication.flags)
     const publicEnv = mergePublicEnvironment(
       bundle.execution.env ?? {},
@@ -249,7 +230,7 @@ export async function prepareAgentCandidateExecution(
         : {},
     )
     const routes = modelRoutes(bundle.profile, task.model.requested)
-    const executionMaterial: AgentCandidateExecutionPlanMaterialV1 = {
+    const executionMaterial: AgentCandidateExecutionPlanMaterial = {
       schemaVersion: 1,
       kind: 'agent-candidate-execution-plan-material',
       bundleDigest: bundle.digest,
@@ -266,7 +247,8 @@ export async function prepareAgentCandidateExecution(
           byteLength: instructionBytes.byteLength,
           delivery: bundle.execution.instructionDelivery,
         },
-        repository: task.repository,
+        ...(task.repository ? { repository: task.repository } : {}),
+        outcome: task.outcome,
         workspace: task.workspace,
       },
       workspaces: {
@@ -298,7 +280,6 @@ export async function prepareAgentCandidateExecution(
         env: publicEnv,
         cwd: bundle.execution.cwd,
       },
-      ...(bundle.knowledge ? { knowledgeManifestDigest: bundle.knowledge.manifest.sha256 } : {}),
       memory,
       limits: task.limits,
       network: { mode: 'disabled' },
@@ -334,7 +315,6 @@ export async function prepareAgentCandidateExecution(
         harnessVersion: bundle.execution.harnessVersion,
         container,
         resolvedModel,
-        ...(bundle.knowledge ? { knowledgeManifestDigest: bundle.knowledge.manifest.sha256 } : {}),
         ...(entrypoint ? { entrypoint } : {}),
       },
     )
@@ -369,6 +349,7 @@ export async function prepareAgentCandidateExecution(
         bytes: profilePlanBytes,
         written: [...profileApplication.application.mountPaths],
       },
+      profileActivation,
       executionPlan: { value: executionPlan, bytes: executionBytes },
       materializationReceipt,
       launch: {
@@ -397,10 +378,6 @@ export async function prepareAgentCandidateExecution(
       executorInputs: {
         taskFiles: taskExecutorFiles,
         ...(candidateExecutorFiles ? { candidateFiles: candidateExecutorFiles } : {}),
-        profileFiles: exactProfileExecutorFiles(
-          profileWorkspacePlan.files,
-          profileApplication.profilePlan.material.files,
-        ),
       },
       ...(preparedMemory.accessDigest && preparedMemory.value.mode === 'isolated'
         ? {
@@ -412,7 +389,6 @@ export async function prepareAgentCandidateExecution(
             },
           }
         : {}),
-      ...(knowledge ? { knowledge } : {}),
       trace: { runId: traceRunId, tags: traceTags, env: traceEnv },
       memory,
     })
@@ -504,9 +480,16 @@ function assertTaskInput(
     ['benchmark', task.benchmark],
     ['benchmarkVersion', task.benchmarkVersion],
     ['taskId', task.taskId],
-    ['repository identity', task.repository.identity],
-    ['repository root identity', task.repository.rootIdentity],
   ]
+  if (task.outcome.kind === 'workspace' && !task.repository) {
+    throw new Error('workspace task outcome requires repository identity')
+  }
+  if (task.repository) {
+    requiredStrings.push(
+      ['repository identity', task.repository.identity],
+      ['repository root identity', task.repository.rootIdentity],
+    )
+  }
   for (const [name, value] of requiredStrings) {
     if (!value.trim()) throw new Error(`${name} must be non-empty`)
   }
@@ -517,16 +500,19 @@ function assertTaskInput(
     throw new Error('task instruction must be non-empty well-formed Unicode')
   }
   sha256DigestSchema.parse(task.splitDigest)
+  agentCandidateTaskOutcomeSpecSchema.parse(task.outcome)
   agentCandidateWorkspaceSnapshotEvidenceSchema.parse(task.workspace)
   agentCandidateExecutionLimitsSchema.parse(task.limits)
-  if (!/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(task.repository.baseCommit)) {
-    throw new Error('task repository base commit is not a full Git object id')
-  }
-  if (!/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(task.repository.baseTree)) {
-    throw new Error('task repository base tree is not a full Git object id')
-  }
-  if (task.repository.baseCommit.length !== task.repository.baseTree.length) {
-    throw new Error('task repository Git object formats disagree')
+  if (task.repository) {
+    if (!/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(task.repository.baseCommit)) {
+      throw new Error('task repository base commit is not a full Git object id')
+    }
+    if (!/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(task.repository.baseTree)) {
+      throw new Error('task repository base tree is not a full Git object id')
+    }
+    if (task.repository.baseCommit.length !== task.repository.baseTree.length) {
+      throw new Error('task repository Git object formats disagree')
+    }
   }
   for (const [name, root] of [
     ['execution task root', task.executionRoots.taskRoot],
@@ -646,15 +632,6 @@ async function assertEmptyDirectory(path: string): Promise<void> {
   }
 }
 
-function materializerHarness(harness: HarnessType): HarnessId {
-  if (!MATERIALIZER_HARNESSES.has(harness)) {
-    throw new Error(
-      `sealed candidate profile materialization is unsupported for harness ${harness}`,
-    )
-  }
-  return harness as HarnessId
-}
-
 async function resolveContainer(
   candidate: VerifiedAgentCandidate,
   task: AgentCandidateTaskExecution,
@@ -843,8 +820,8 @@ function unwrapPublicEnvironment(
 function modelRoutes(
   profile: VerifiedAgentCandidate['bundle']['profile'],
   requested: string,
-): AgentCandidateExecutionPlanMaterialV1['model']['routes'] {
-  const routes: AgentCandidateExecutionPlanMaterialV1['model']['routes'] = [
+): AgentCandidateExecutionPlanMaterial['model']['routes'] {
+  const routes: AgentCandidateExecutionPlanMaterial['model']['routes'] = [
     { kind: 'primary', requested },
   ]
   if (profile.model?.small) routes.push({ kind: 'small', requested })
@@ -940,30 +917,6 @@ function modelLimits(limits: AgentCandidateExecutionLimits): {
   }
 }
 
-function exactProfileExecutorFiles(
-  sourceFiles: ReadonlyArray<{ relPath: string; content: string; mode?: number }>,
-  expectedFiles: ReadonlyArray<{ relPath: string; mode: number; contentSha256: string }>,
-): Array<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }> {
-  const byPath = new Map(sourceFiles.map((file) => [file.relPath, file]))
-  if (byPath.size !== sourceFiles.length || sourceFiles.length !== expectedFiles.length) {
-    throw new Error('profile source files do not match the signed profile plan')
-  }
-  return expectedFiles.map((expected) => {
-    const source = byPath.get(expected.relPath)
-    const mode = source?.mode ?? 0o644
-    const bytes = Buffer.from(source?.content ?? '', 'utf8')
-    if (
-      !source ||
-      (mode !== 0o644 && mode !== 0o755) ||
-      mode !== expected.mode ||
-      sha256Bytes(bytes) !== expected.contentSha256
-    ) {
-      throw new Error('profile source files do not match the signed profile plan')
-    }
-    return { path: expected.relPath, mode, bytes: Uint8Array.from(bytes) }
-  })
-}
-
 function isWellFormedUnicode(value: string): boolean {
   for (let index = 0; index < value.length; index++) {
     const code = value.charCodeAt(index)
diff --git a/src/candidate-execution/prepared-state.ts b/src/candidate-execution/prepared-state.ts
index 87522785..b6ef6b22 100644
--- a/src/candidate-execution/prepared-state.ts
+++ b/src/candidate-execution/prepared-state.ts
@@ -14,9 +14,11 @@ import {
   sha256Bytes,
 } from './digest'
 import { verifyTaskCheckout } from './git-materialize'
+import { parseAgentCandidateProfileActivation } from './profile'
 import type {
   AgentCandidateExecutionPorts,
   AgentCandidateExecutorRequest,
+  AgentCandidateExecutorWorkspaceFile,
   AgentCandidateProtectedModelActivation,
   AgentCandidateProtectedModelReservation,
   PreparedAgentCandidateExecution,
@@ -29,6 +31,7 @@ export interface PreparedCandidateState {
   executionId: string
   roots: PreparedAgentCandidateExecution['roots']
   profilePlan: PreparedAgentCandidateExecution['profilePlan']
+  profileActivation: PreparedAgentCandidateExecution['profileActivation']
   executionPlan: PreparedAgentCandidateExecution['executionPlan']
   materializationReceipt: PreparedAgentCandidateExecution['materializationReceipt']
   launch: PreparedAgentCandidateExecution['launch']
@@ -42,17 +45,12 @@ export interface PreparedCandidateState {
   executorInputs: {
     taskFiles: ReadonlyArray<{
       path: string
-      mode: 0o644 | 0o755
+      mode: number
       bytes: Uint8Array
     }>
     candidateFiles?: ReadonlyArray<{
       path: string
-      mode: 0o644 | 0o755
-      bytes: Uint8Array
-    }>
-    profileFiles: ReadonlyArray<{
-      path: string
-      mode: 0o644 | 0o755
+      mode: number
       bytes: Uint8Array
     }>
   }
@@ -62,7 +60,6 @@ export interface PreparedCandidateState {
     expiresAtMs: number
     effectiveNamespace: string
   }
-  knowledge?: PreparedAgentCandidateExecution['knowledge']
   trace: PreparedAgentCandidateExecution['trace']
   memory: PreparedAgentCandidateExecution['memory']
 }
@@ -96,12 +93,12 @@ export function createPreparedCandidateExecution(
     executionId: state.executionId,
     roots: state.roots,
     profilePlan: evidenceView(state.profilePlan),
+    profileActivation: state.profileActivation,
     executionPlan: evidenceView(state.executionPlan),
     materializationReceipt: state.materializationReceipt,
     launch: state.launch,
     instruction: bytesView(state.instruction, 'bytes'),
     resolvedModel: state.resolvedModel,
-    ...(state.knowledge ? { knowledge: knowledgeView(state.knowledge) } : {}),
     trace: state.trace,
     memory: state.memory,
     [preparedCandidateBrand]: true as const,
@@ -180,14 +177,10 @@ export function beginPreparedCandidateRun(
             ),
           }
         : {}),
-      profile: Object.freeze({
-        files: Object.freeze(
-          state.executorInputs.profileFiles.map((file) => profileFileView(file)),
-        ),
-      }),
     }),
     roots: state.roots.execution,
     profilePlan: evidenceView(state.profilePlan),
+    profileActivation: state.profileActivation,
     executionPlan: evidenceView(state.executionPlan),
     materializationReceipt: state.materializationReceipt,
     launch: immutableCandidateValue({ ...state.launch, env: completeEnvironment }),
@@ -195,7 +188,6 @@ export function beginPreparedCandidateRun(
     resolvedModel: state.resolvedModel,
     hardLimits: Object.freeze({ timeoutMs: state.executionPlan.value.material.limits.timeoutMs }),
     observedLimits: Object.freeze({ maxSteps: state.executionPlan.value.material.limits.maxSteps }),
-    ...(state.knowledge ? { knowledge: knowledgeView(state.knowledge) } : {}),
     trace: state.trace,
     memory: state.memory,
   })
@@ -225,7 +217,9 @@ export async function assertPreparedCandidateWorkspaces(
   await verifyMaterializedWorkspace(state.roots.staging.taskRoot, plan.task.workspace.material, {
     ignoredProtectedRootEntries: ['.git', '.sidecar'],
   })
-  await verifyTaskCheckout(state.roots.staging.taskRoot, plan.task.repository)
+  if (plan.task.repository) {
+    await verifyTaskCheckout(state.roots.staging.taskRoot, plan.task.repository)
+  }
   await verifyMaterializedProfileWorkspace(
     state.roots.staging.profileRoot,
     state.profilePlan.value.material,
@@ -243,6 +237,10 @@ function detachPreparedCandidateState(input: PreparedCandidateState): PreparedCa
   const profilePlan = immutableCandidateValue(
     agentCandidateProfilePlanEvidenceSchema.parse(input.profilePlan.value),
   )
+  const profileActivation = parseAgentCandidateProfileActivation(
+    input.profileActivation,
+    profilePlan.digest,
+  )
   const executionPlan = immutableCandidateValue(
     agentCandidateExecutionPlanEvidenceSchema.parse(input.executionPlan.value),
   )
@@ -265,6 +263,7 @@ function detachPreparedCandidateState(input: PreparedCandidateState): PreparedCa
       bytes: Uint8Array.from(input.profilePlan.bytes),
       written: Object.freeze([...input.profilePlan.written]),
     }),
+    profileActivation,
     executionPlan: Object.freeze({
       value: executionPlan,
       bytes: Uint8Array.from(input.executionPlan.bytes),
@@ -286,24 +285,10 @@ function detachPreparedCandidateState(input: PreparedCandidateState): PreparedCa
       ...(input.executorInputs.candidateFiles
         ? { candidateFiles: immutableExecutorFiles(input.executorInputs.candidateFiles) }
         : {}),
-      profileFiles: Object.freeze(
-        input.executorInputs.profileFiles.map((file) =>
-          Object.freeze({ ...file, bytes: Uint8Array.from(file.bytes) }),
-        ),
-      ),
     }),
     ...(input.memoryReservation
       ? { memoryReservation: immutableCandidateValue(input.memoryReservation) }
       : {}),
-    ...(input.knowledge
-      ? {
-          knowledge: Object.freeze({
-            snapshotId: input.knowledge.snapshotId,
-            manifestDigest: input.knowledge.manifestDigest,
-            manifest: Uint8Array.from(input.knowledge.manifest),
-          }),
-        }
-      : {}),
     trace: immutableCandidateValue(input.trace),
     memory: immutableCandidateValue(input.memory),
   })
@@ -315,6 +300,7 @@ function assertPrivateCandidateIntegrity(state: PreparedCandidateState): void {
     throw new Error('prepared candidate bundle no longer matches its digest')
   }
   assertPlanEvidence(state.profilePlan.value, state.profilePlan.bytes, 'profile plan')
+  parseAgentCandidateProfileActivation(state.profileActivation, state.profilePlan.value.digest)
   assertPlanEvidence(state.executionPlan.value, state.executionPlan.bytes, 'execution plan')
   agentCandidateExecutionPlanEvidenceSchema.parse(state.executionPlan.value)
 
@@ -441,32 +427,19 @@ function bytesView<T extends { bytes: Uint8Array }>(value: T, key: 'bytes'): T {
   })
 }
 
-function knowledgeView(
-  knowledge: NonNullable<PreparedAgentCandidateExecution['knowledge']>,
-): NonNullable<PreparedAgentCandidateExecution['knowledge']> {
-  const manifest = Uint8Array.from(knowledge.manifest)
-  return Object.freeze({
-    snapshotId: knowledge.snapshotId,
-    manifestDigest: knowledge.manifestDigest,
-    get manifest(): Uint8Array {
-      return Uint8Array.from(manifest)
-    },
-  })
-}
-
 function workspaceInputView(
   snapshot: AgentCandidateExecutorRequest['inputs']['task']['snapshot'],
   sourceFiles: PreparedCandidateState['executorInputs']['taskFiles'],
 ): AgentCandidateExecutorRequest['inputs']['task'] {
   return Object.freeze({
     snapshot,
-    files: Object.freeze(sourceFiles.map((file) => profileFileView(file))),
+    files: Object.freeze(sourceFiles.map((file) => executorFileView(file))),
   })
 }
 
-function profileFileView(
-  source: PreparedCandidateState['executorInputs']['profileFiles'][number],
-): AgentCandidateExecutorRequest['inputs']['profile']['files'][number] {
+function executorFileView(
+  source: PreparedCandidateState['executorInputs']['taskFiles'][number],
+): AgentCandidateExecutorWorkspaceFile {
   const bytes = Uint8Array.from(source.bytes)
   return Object.freeze({
     path: source.path,
@@ -491,24 +464,6 @@ function assertExecutorInputs(state: PreparedCandidateState): void {
   } else if (state.executorInputs.candidateFiles) {
     throw new Error('disabled candidate has executor files')
   }
-
-  const expectedFiles = state.profilePlan.value.material.files
-  if (state.executorInputs.profileFiles.length !== expectedFiles.length) {
-    throw new Error('prepared profile executor files do not match the signed profile plan')
-  }
-  for (let index = 0; index < expectedFiles.length; index++) {
-    const expected = expectedFiles[index]
-    const actual = state.executorInputs.profileFiles[index]
-    if (
-      !expected ||
-      !actual ||
-      actual.path !== expected.relPath ||
-      actual.mode !== expected.mode ||
-      sha256Bytes(actual.bytes) !== expected.contentSha256
-    ) {
-      throw new Error('prepared profile executor files do not match the signed profile plan')
-    }
-  }
 }
 
 function assertWorkspaceExecutorFiles(
@@ -536,7 +491,7 @@ function assertWorkspaceExecutorFiles(
 
 function immutableExecutorFiles(
   files: PreparedCandidateState['executorInputs']['taskFiles'],
-): ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }> {
+): ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }> {
   return Object.freeze(
     files.map((file) => Object.freeze({ ...file, bytes: Uint8Array.from(file.bytes) })),
   )
diff --git a/src/candidate-execution/profile.ts b/src/candidate-execution/profile.ts
index 7f814121..b1ccb153 100644
--- a/src/candidate-execution/profile.ts
+++ b/src/candidate-execution/profile.ts
@@ -1,25 +1,114 @@
 import type {
   AgentCandidateConfigValue,
   AgentCandidateProfile,
+  AgentCandidateProfileActivation,
+  AgentCandidateProfilePlanEvidence,
   AgentCandidateResourceRef,
   AgentProfile,
   AgentProfileDiff,
   AgentProfileMcpServer,
   AgentProfileResourceRef,
+  HarnessType,
 } from '@tangle-network/agent-interface'
 import {
+  agentCandidateProfileActivationSchema,
   agentCandidateProfileSchema,
   agentProfileDiffSchema,
   agentProfileSchema,
   applyAgentProfileDiff,
 } from '@tangle-network/agent-interface'
+import type {
+  AgentCandidateWorkspacePlan,
+  HarnessId,
+} from '@tangle-network/agent-profile-materialize'
 
 import {
   canonicalCandidateBytes,
   canonicalCandidateDigest,
+  canonicalCandidateDocument,
   embeddedCandidateArtifact,
+  immutableCandidateValue,
+  omitTopLevelDigest,
+  sha256Bytes,
 } from './digest'
 
+const MATERIALIZER_HARNESSES = new Set<HarnessType>([
+  'claude-code',
+  'claude',
+  'claudish',
+  'nanoclaw',
+  'codex',
+  'opencode',
+  'kimi-code',
+  'kimi',
+  'pi',
+  'gemini',
+  'hermes',
+  'openclaw',
+])
+
+export function candidateMaterializerHarness(harness: HarnessType): HarnessId {
+  if (!MATERIALIZER_HARNESSES.has(harness)) {
+    throw new Error(
+      `sealed candidate profile materialization is unsupported for harness ${harness}`,
+    )
+  }
+  return harness as HarnessId
+}
+
+/** Bind exact native profile text to the canonical plan captured during preparation. */
+export function createAgentCandidateProfileActivation(
+  plan: AgentCandidateWorkspacePlan,
+  profilePlan: AgentCandidateProfilePlanEvidence,
+): AgentCandidateProfileActivation {
+  const sourceByPath = new Map(plan.files.map((file) => [file.relPath, file]))
+  if (sourceByPath.size !== plan.files.length) {
+    throw new Error('candidate profile activation contains duplicate native file paths')
+  }
+  const files = profilePlan.material.files.map((expected) => {
+    const source = sourceByPath.get(expected.relPath)
+    const mode = source?.mode ?? 0o644
+    if (!source || !Number.isSafeInteger(mode) || mode < 0 || mode > 0o777) {
+      throw new Error('candidate profile activation does not contain every planned native file')
+    }
+    return { path: expected.relPath, mode, content: source.content }
+  })
+  if (files.length !== plan.files.length) {
+    throw new Error('candidate profile activation contains unplanned native files')
+  }
+  return parseAgentCandidateProfileActivation(
+    canonicalCandidateDocument<AgentCandidateProfileActivation>({
+      schemaVersion: 1,
+      kind: 'agent-candidate-profile-activation',
+      profilePlan,
+      files,
+    }).value,
+    profilePlan.digest,
+  )
+}
+
+/** Parse and check every native file hash plus both canonical document digests. */
+export function parseAgentCandidateProfileActivation(
+  input: unknown,
+  expectedProfilePlanDigest?: AgentCandidateProfilePlanEvidence['digest'],
+): AgentCandidateProfileActivation {
+  const activation = agentCandidateProfileActivationSchema.parse(input)
+  const planBytes = canonicalCandidateBytes(activation.profilePlan.material)
+  if (
+    sha256Bytes(planBytes) !== activation.profilePlan.digest ||
+    activation.profilePlan.artifact.sha256 !== activation.profilePlan.digest ||
+    activation.profilePlan.artifact.byteLength !== planBytes.byteLength ||
+    (expectedProfilePlanDigest !== undefined &&
+      activation.profilePlan.digest !== expectedProfilePlanDigest)
+  ) {
+    throw new Error('candidate profile activation has an invalid canonical profile plan')
+  }
+  if (canonicalCandidateDigest(omitTopLevelDigest(activation)) !== activation.digest) {
+    throw new Error('candidate profile activation digest does not match')
+  }
+  return immutableCandidateValue(activation)
+}
+
 const CANDIDATE_PROFILE_DIRECT_FIELDS = [
   'name',
   'description',
@@ -86,7 +175,7 @@ export function assertCandidateProfileBinding(
   if (measured.connections || measured.metadata || measured.extensions) {
     throw new Error('proposal candidate profile contains fields unsupported by sealed candidates')
   }
-  const normalized = candidateProfileAsAgentProfile(bundled)
+  const normalized = agentCandidateProfileAsAgentProfile(bundled)
   if (canonicalCandidateDigest(measured) !== canonicalCandidateDigest(normalized)) {
     throw new Error('proposal candidateProfile does not match candidateBundle.profile')
   }
@@ -135,7 +224,9 @@ export function assertCandidateProfileExecutionSupport(profile: AgentCandidatePr
   }
 }
 
-function candidateProfileAsAgentProfile(candidate: AgentCandidateProfile): AgentProfile {
+export function agentCandidateProfileAsAgentProfile(
+  candidate: AgentCandidateProfile,
+): AgentProfile {
   const output: Record<string, unknown> = {}
   copyDirectProfileFields(output, candidate as unknown as Record<string, unknown>)
   if (candidate.model) output.model = { ...candidate.model }
diff --git a/src/candidate-execution/protected-model-port.ts b/src/candidate-execution/protected-model-port.ts
index e69194e1..5111a9c7 100644
--- a/src/candidate-execution/protected-model-port.ts
+++ b/src/candidate-execution/protected-model-port.ts
@@ -204,7 +204,7 @@ export function createProtectedAgentCandidateModelPort(
         grantDigest: request.grantDigest,
         model: request.resolved.model,
       })
-      if (state) assertWithinReservedLimits(sealed.fixedUsage, state)
+      if (state) assertWithinReservedLimits(sealed.usage, state)
 
       const settlementDigest = canonicalCandidateDigest(sealed.value)
       if (remembered?.settlementDigest && remembered.settlementDigest !== settlementDigest) {
diff --git a/src/candidate-execution/recover.ts b/src/candidate-execution/recover.ts
index cff11369..aeb47ff1 100644
--- a/src/candidate-execution/recover.ts
+++ b/src/candidate-execution/recover.ts
@@ -1,7 +1,17 @@
 import type { TraceStore } from '@tangle-network/agent-eval'
+import type {
+  AgentCandidateExecutionPlanMaterial,
+  AgentCandidateMaterializationReceipt,
+} from '@tangle-network/agent-interface'
+import {
+  agentCandidateExecutionPlanMaterialSchema,
+  agentCandidateMaterializationReceiptSchema,
+} from '@tangle-network/agent-interface'
 
+import { readVerifiedArtifact } from './artifacts'
 import type {
   AgentCandidateExecutionAttemptRef,
+  AgentCandidateExecutionClaim,
   AgentCandidateExecutionClaimStore,
   AgentCandidateExecutionFinishResult,
 } from './claim'
@@ -10,7 +20,12 @@ import {
   candidateCleanupTimeout,
   withinCandidateCleanupDeadline,
 } from './cleanup'
-import { sealAgentCandidateExecutorFinalCapture } from './executor-capture'
+import { canonicalCandidateBytes } from './digest'
+import {
+  sealAgentCandidateExecutorFinalCapture,
+  sealAgentCandidateExecutorStopAcknowledgement,
+} from './executor-capture'
+import { persistAgentCandidateExecutorCapture } from './executor-capture-evidence'
 import { sealAgentCandidateModelSettlement } from './model-settlement'
 import { persistCandidateModelSettlementEvidence } from './outcome-evidence'
 import { RecoveryAgentCandidateTraceStore } from './protected-trace-store'
@@ -57,9 +72,14 @@ export async function recoverExpiredAgentCandidateExecution(
   const controller = new AbortController()
   controller.abort(new Error('recovering an expired candidate execution'))
   const recoveryTraceStore = new RecoveryAgentCandidateTraceStore(options.traceStore)
+  const preparation = withinCandidateCleanupDeadline(
+    () => readRecoveryPreparation(record.claim, options.outputArtifacts),
+    cleanupDeadlineAtMs,
+    'expired candidate preparation evidence read',
+  )
   const processClosure = withinCandidateCleanupDeadline(
     async () => {
-      const stopped = await options.executor.stopAndCapture(
+      const stopped = await options.executor.stop(
         {
           executionId: record.claim.executionId,
           executionPlanDigest: record.claim.executionPlanDigest,
@@ -71,7 +91,7 @@ export async function recoverExpiredAgentCandidateExecution(
           deadlineAtMs: record.claim.leaseExpiresAtMs,
         },
       )
-      sealAgentCandidateExecutorFinalCapture(stopped)
+      sealAgentCandidateExecutorStopAcknowledgement(stopped)
       return { stopped: true as const }
     },
     cleanupDeadlineAtMs,
@@ -120,16 +140,72 @@ export async function recoverExpiredAgentCandidateExecution(
       )
     : undefined
 
-  const operations = [processClosure, modelClosure, ...(memoryClosure ? [memoryClosure] : [])]
-  const outcomes = await Promise.allSettled(operations)
-  const failures = outcomes
+  const closureOutcomes = Promise.allSettled([
+    modelClosure,
+    memoryClosure ?? Promise.resolve(undefined),
+  ] as const)
+  const [preparationOutcome, processOutcome] = await Promise.allSettled([
+    preparation,
+    processClosure,
+  ] as const)
+  const failures: unknown[] = [preparationOutcome, processOutcome]
     .filter((outcome): outcome is PromiseRejectedResult => outcome.status === 'rejected')
     .map((outcome) => outcome.reason)
+  const recoveredPreparation =
+    preparationOutcome.status === 'fulfilled' ? preparationOutcome.value : undefined
+  const processStopped = processOutcome.status === 'fulfilled'
+  let persistedCapture: Awaited<ReturnType<typeof persistAgentCandidateExecutorCapture>> | undefined
+  if (recoveredPreparation && processStopped) {
+    try {
+      const capture = sealAgentCandidateExecutorFinalCapture(
+        await withinCandidateCleanupDeadline(
+          () =>
+            options.executor.capture(
+              {
+                executionId: record.claim.executionId,
+                executionPlanDigest: record.claim.executionPlanDigest,
+              },
+              {
+                traceStore: recoveryTraceStore,
+                signal: new AbortController().signal,
+              },
+            ),
+          cleanupDeadlineAtMs,
+          'expired candidate evidence capture',
+        ),
+        recoveredPreparation.plan.task.outcome,
+      )
+      persistedCapture = await withinCandidateCleanupDeadline(
+        () =>
+          persistAgentCandidateExecutorCapture(
+            {
+              executionId: record.claim.executionId,
+              executionPlanDigest: record.claim.executionPlanDigest,
+            },
+            recoveredPreparation.plan.task.outcome,
+            capture,
+            options.outputArtifacts,
+          ),
+        cleanupDeadlineAtMs,
+        'expired candidate capture persistence',
+      )
+    } catch (error) {
+      failures.push(error)
+    }
+  }
+  const [modelOutcome, memoryOutcome] = await closureOutcomes
+  for (const outcome of [modelOutcome, memoryOutcome]) {
+    if (outcome.status === 'rejected') failures.push(outcome.reason)
+  }
   if (failures.length > 0) {
     throw new AggregateError(failures, 'expired candidate cleanup could not be proven')
   }
+  if (!persistedCapture) throw new Error('expired candidate capture was not persisted')
 
-  const model = await modelClosure
+  if (modelOutcome.status !== 'fulfilled') {
+    throw new Error('expired candidate model settlement was not recovered')
+  }
+  const model = modelOutcome.value
   const modelSettlement = await withinCandidateCleanupDeadline(
     () =>
       persistCandidateModelSettlementEvidence(
@@ -147,11 +223,12 @@ export async function recoverExpiredAgentCandidateExecution(
   const memory = record.claim.cleanup.memory
   return await options.claimStore.recoverExpired(options.attempt, {
     failureClass:
-      record.phase === 'claimed' && model.fixedUsage.modelCalls === 0
+      record.phase === 'claimed' && model.usage.modelCalls === 0
         ? 'pre-model-infrastructure'
         : 'unknown',
-    usage: model.fixedUsage,
+    usage: model.usage,
     modelSettlement: modelSettlement.artifact,
+    failureEvidence: persistedCapture.evidence,
     process: {
       stopped: true,
       executionPlanDigest: record.claim.executionPlanDigest,
@@ -173,3 +250,52 @@ export async function recoverExpiredAgentCandidateExecution(
       : {}),
   })
 }
+
+async function readRecoveryPreparation(
+  claim: AgentCandidateExecutionClaim,
+  artifacts: AgentCandidateOutputArtifactPort,
+): Promise<{
+  plan: AgentCandidateExecutionPlanMaterial
+  receipt: AgentCandidateMaterializationReceipt
+}> {
+  const [planBytes, receiptBytes] = await Promise.all([
+    readVerifiedArtifact(claim.preparationEvidence.executionPlan, artifacts),
+    readVerifiedArtifact(claim.preparationEvidence.materializationReceipt, artifacts),
+  ])
+  const plan = agentCandidateExecutionPlanMaterialSchema.parse(
+    parseCanonicalBytes(planBytes, 'plan'),
+  )
+  if (claim.preparationEvidence.executionPlan.sha256 !== claim.executionPlanDigest) {
+    throw new Error('recovery execution plan artifact does not match the claimed digest')
+  }
+  const receiptMaterial = parseCanonicalBytes(receiptBytes, 'materialization receipt')
+  const receipt = agentCandidateMaterializationReceiptSchema.parse({
+    ...receiptMaterial,
+    digest: claim.preparationEvidence.materializationReceipt.sha256,
+  })
+  if (
+    receipt.bundleDigest !== claim.bundleDigest ||
+    receipt.executionPlan.digest !== claim.executionPlanDigest ||
+    plan.bundleDigest !== claim.bundleDigest ||
+    plan.executionId !== claim.executionId
+  ) {
+    throw new Error('recovery preparation evidence does not match the durable claim')
+  }
+  return { plan, receipt }
+}
+
+function parseCanonicalBytes(bytes: Uint8Array, label: string): Record<string, unknown> {
+  let parsed: unknown
+  try {
+    parsed = JSON.parse(Buffer.from(bytes).toString('utf8'))
+  } catch (error) {
+    throw new Error(`candidate recovery ${label} is not JSON`, { cause: error })
+  }
+  if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed)) {
+    throw new Error(`candidate recovery ${label} must be an object`)
+  }
+  if (!Buffer.from(canonicalCandidateBytes(parsed)).equals(Buffer.from(bytes))) {
+    throw new Error(`candidate recovery ${label} bytes are not canonical`)
+  }
+  return parsed as Record<string, unknown>
+}
diff --git a/src/candidate-execution/types.ts b/src/candidate-execution/types.ts
index 097bc62b..57b63ad8 100644
--- a/src/candidate-execution/types.ts
+++ b/src/candidate-execution/types.ts
@@ -8,19 +8,25 @@ import type {
   AgentCandidateEffectiveMemory,
   AgentCandidateExecutionLimits,
   AgentCandidateExecutionPlanEvidence,
+  AgentCandidateFixedSpend,
   AgentCandidateGitHubRepository,
   AgentCandidateInstructionDelivery,
   AgentCandidateMaterializationReceipt,
   AgentCandidateMemoryReceipt,
   AgentCandidateModelAccessNetwork,
+  AgentCandidateModelSettlementCall,
   AgentCandidateOciPlatform,
+  AgentCandidateProfileActivation,
   AgentCandidateProfilePlanEvidence,
   AgentCandidateResolvedModel,
-  AgentCandidateRunReceiptV2,
-  AgentCandidateSpend,
+  AgentCandidateRunReceipt,
   AgentCandidateTaskOutcomeEvidence,
+  AgentCandidateTaskOutcomeMaterial,
+  AgentCandidateTaskOutcomeSpec,
+  AgentCandidateTaskOutputSpec,
+  AgentCandidateTaskRepository,
   AgentCandidateTermination,
-  AgentCandidateWorkspaceManifestMaterialV1,
+  AgentCandidateWorkspaceManifestMaterial,
   AgentCandidateWorkspaceSnapshotEvidence,
   ReasoningEffort,
   Sha256Digest,
@@ -36,11 +42,14 @@ export interface AgentCandidateArtifactPort {
 }
 
 export type AgentCandidateOutputPurpose =
+  | 'execution-plan'
+  | 'materialization-receipt'
   | 'candidate-workspace-manifest'
   | 'candidate-workspace-archive'
   | 'task-manifest'
   | 'task-archive'
   | 'task-patch'
+  | 'task-output'
   | 'task-outcome'
   | 'memory-after-manifest'
   | 'memory-after-archive'
@@ -49,6 +58,9 @@ export type AgentCandidateOutputPurpose =
   | 'model-settlement'
   | 'trace'
   | 'run-receipt'
+  | 'executor-capture'
+  | 'knowledge-retrieval-config'
+  | 'knowledge-evaluation'
   | 'failure-evidence'
 
 /** Durable content-addressed evidence store controlled only by the evaluator. */
@@ -175,30 +187,11 @@ export interface AgentCandidateProtectedModelActivation {
   env: Readonly<Record<string, string>>
 }
 
-/** One evaluator-gateway call in the final, revoked model-access ledger. */
-export interface AgentCandidateProtectedModelCall {
-  callId: string
-  /** Router-generated public response identity. */
-  generationId: string
-  /** Exact protected agent-eval LLM span produced from the router ledger. */
-  traceSpanId: string
-  status: 'succeeded' | 'failed'
-  model: string
-  startedAtMs: number
-  endedAtMs: number
-  inputTokens: number
-  outputTokens: number
-  cachedInputTokens: number
-  reasoningTokens: number
-  /** Integer billionths of one US dollar; avoids floating-point ledger drift. */
-  costUsdNanos: number
-}
-
 export interface AgentCandidateProtectedModelSettlement {
   preparationId: string
   grantDigest: Sha256Digest
   closed: true
-  calls: readonly AgentCandidateProtectedModelCall[]
+  calls: readonly AgentCandidateModelSettlementCall[]
 }
 
 export interface AgentCandidateMemoryResetResult {
@@ -256,6 +249,7 @@ export interface AgentCandidateExecutionPorts extends AgentCandidateVerification
   memory: AgentCandidateMemoryPort
 }
 
+/** One signed benchmark task and the exact result shape its executor must capture. */
 export interface AgentCandidateTaskExecution {
   executionId: string
   benchmark: string
@@ -264,12 +258,9 @@ export interface AgentCandidateTaskExecution {
   splitDigest: Sha256Digest
   /** Exact agent-visible task instruction. The runtime rejects malformed Unicode. */
   instruction: string
-  repository: {
-    identity: string
-    rootIdentity: string
-    baseCommit: string
-    baseTree: string
-  }
+  /** Optional source identity, required when the expected outcome is a workspace. */
+  repository?: AgentCandidateTaskRepository
+  outcome: AgentCandidateTaskOutcomeSpec
   attempt: AgentCandidateAttemptPolicy
   model: {
     requested: string
@@ -345,6 +336,7 @@ export interface PreparedAgentCandidateExecution {
     bytes: Uint8Array
     written: readonly string[]
   }
+  readonly profileActivation: AgentCandidateProfileActivation
   readonly executionPlan: {
     value: AgentCandidateExecutionPlanEvidence
     bytes: Uint8Array
@@ -353,11 +345,6 @@ export interface PreparedAgentCandidateExecution {
   readonly launch: PreparedAgentCandidateLaunch
   readonly instruction: PreparedAgentCandidateInstruction
   readonly resolvedModel: AgentCandidateResolvedModel
-  readonly knowledge?: {
-    snapshotId: string
-    manifestDigest: Sha256Digest
-    manifest: Uint8Array
-  }
   readonly trace: PreparedAgentCandidateTrace
   readonly memory: AgentCandidateEffectiveMemory
   readonly [preparedCandidateBrand]: true
@@ -369,40 +356,64 @@ export interface AgentCandidateProtectedRunCapture {
 }
 
 /** Raw evaluator capture made only after the candidate process is dead. */
-export interface AgentCandidateExecutorTaskOutcomeCapture {
-  /** Claimed final tree. The runtime recomputes it independently from `gitDiff`. */
-  resultTree: string
-  /** Complete evaluator-captured workspace description after candidate execution. */
-  afterState: AgentCandidateWorkspaceManifestMaterialV1
-  /** Reproducible workspace archive corresponding to `afterState`. */
-  archive: Uint8Array
-  /** Exact binary patch from the signed task base to `afterState`. */
-  gitDiff: Uint8Array
-}
+export type AgentCandidateExecutorTaskOutcomeCapture =
+  | {
+      readonly kind: 'workspace'
+      /** Claimed final tree. The runtime recomputes it independently from `gitDiff`. */
+      readonly resultTree: string
+      /** Complete evaluator-captured workspace description after candidate execution. */
+      readonly afterState: AgentCandidateWorkspaceManifestMaterial
+      /** Reproducible workspace archive corresponding to `afterState`. */
+      readonly archive: Uint8Array
+      /** Exact binary patch from the signed task base to `afterState`. */
+      readonly gitDiff: Uint8Array
+    }
+  | {
+      readonly kind: 'output'
+      /** Exact evaluator-captured final output bytes. */
+      readonly bytes: Uint8Array
+    }
 
 /** Raw isolated-memory capture made only after access has been revoked. */
 export interface AgentCandidateExecutorMemoryCapture {
-  readonly afterState: AgentCandidateWorkspaceManifestMaterialV1
+  readonly afterState: AgentCandidateWorkspaceManifestMaterial
   readonly archive: Uint8Array
 }
 
-/** Idempotent executor result after process death and trace drain. */
+/** Replayable evaluator result captured only after process death and trace drain. */
 export interface AgentCandidateExecutorFinalCapture {
-  readonly stopped: true
   readonly taskOutcome?: AgentCandidateExecutorTaskOutcomeCapture
   /** Required only when the prepared candidate uses isolated task memory. */
   readonly memoryAfter?: AgentCandidateExecutorMemoryCapture
+  /** Executor-native bytes preserved when a fresh worker cannot reconstruct a verified outcome. */
+  readonly evidence?: Uint8Array
 }
 
-/** Branded task outcome that has survived independent patch and tree verification. */
-export interface VerifiedAgentCandidateTaskOutcome {
-  readonly evidence: AgentCandidateTaskOutcomeEvidence & {
-    readonly artifact: AgentCandidateArtifactRef
+type PersistedTaskOutcomeEvidence<
+  Kind extends AgentCandidateTaskOutcomeMaterial['outcome']['kind'],
+> = Omit<AgentCandidateTaskOutcomeEvidence, 'material'> & {
+  readonly artifact: AgentCandidateArtifactRef
+  readonly material: Omit<AgentCandidateTaskOutcomeMaterial, 'outcome'> & {
+    readonly outcome: Extract<AgentCandidateTaskOutcomeMaterial['outcome'], { kind: Kind }>
   }
-  readonly patch: Uint8Array
-  readonly [verifiedTaskOutcomeBrand]: true
 }
 
+/** Branded task outcome that has survived independent evaluator verification. */
+export type VerifiedAgentCandidateTaskOutcome =
+  | {
+      readonly kind: 'workspace'
+      readonly evidence: PersistedTaskOutcomeEvidence<'workspace'>
+      readonly patch: Uint8Array
+      readonly [verifiedTaskOutcomeBrand]: true
+    }
+  | {
+      readonly kind: 'output'
+      readonly evidence: PersistedTaskOutcomeEvidence<'output'>
+      readonly spec: AgentCandidateTaskOutputSpec
+      readonly bytes: Uint8Array
+      readonly [verifiedTaskOutcomeBrand]: true
+    }
+
 /**
  * Evaluator-owned executable grader, pinned by immutable implementation bytes.
  *
@@ -450,12 +461,10 @@ export interface AgentCandidateExecutorRequest {
   readonly inputs: {
     readonly task: AgentCandidateExecutorWorkspaceInput
     readonly candidate?: AgentCandidateExecutorWorkspaceInput
-    readonly profile: {
-      readonly files: readonly AgentCandidateExecutorProfileFile[]
-    }
   }
   readonly roots: PreparedAgentCandidateExecution['roots']['execution']
   readonly profilePlan: PreparedAgentCandidateExecution['profilePlan']
+  readonly profileActivation: AgentCandidateProfileActivation
   readonly executionPlan: PreparedAgentCandidateExecution['executionPlan']
   readonly materializationReceipt: CanonicalCandidateDocument<AgentCandidateMaterializationReceipt>
   readonly launch: PreparedAgentCandidateLaunch
@@ -465,7 +474,6 @@ export interface AgentCandidateExecutorRequest {
   readonly hardLimits: Pick<AgentCandidateExecutionLimits, 'timeoutMs'>
   /** Validity bound checked against protected traces; generic black-box executors cannot preempt it. */
   readonly observedLimits: Pick<AgentCandidateExecutionLimits, 'maxSteps'>
-  readonly knowledge?: PreparedAgentCandidateExecution['knowledge']
   readonly trace: PreparedAgentCandidateTrace
   readonly memory: AgentCandidateEffectiveMemory
 }
@@ -489,14 +497,8 @@ export interface AgentCandidateExecutorPort {
       deadlineAtMs: number
     },
   ): Promise<AgentCandidateProtectedRunCapture>
-  /**
-   * Kill any process/container still associated with the request, drain trace
-   * writes, and capture the final task workspace before teardown.
-   * The runtime calls this on success, failure, and timeout before model settlement.
-   * Implementations must be idempotent and concurrency-safe for this exact
-   * execution/plan pair because a fresh worker may repeat crash recovery.
-   */
-  stopAndCapture(
+  /** Kill the exact process/container and drain trace writes. Must be idempotent. */
+  stop(
     request: AgentCandidateExecutorStopRequest,
     context: {
       traceStore: TraceStore
@@ -506,6 +508,15 @@ export interface AgentCandidateExecutorPort {
       /** Absolute execution deadline; a later stop acknowledgement cannot produce success. */
       deadlineAtMs: number
     },
+  ): Promise<{ readonly stopped: true }>
+  /** Capture immutable final evidence after stop. Must be replayable by a fresh worker. */
+  capture(
+    request: AgentCandidateExecutorStopRequest,
+    context: {
+      traceStore: TraceStore
+      /** Aborted at the frozen execution deadline or evaluator cleanup deadline. */
+      signal: AbortSignal
+    },
   ): Promise<AgentCandidateExecutorFinalCapture>
 }
 
@@ -522,22 +533,17 @@ export interface AgentCandidateExecutorWorkspaceInput {
 
 export interface AgentCandidateExecutorWorkspaceFile {
   readonly path: string
-  readonly mode: 0o644 | 0o755
-  readonly bytes: Uint8Array
-}
-
-export interface AgentCandidateExecutorProfileFile {
-  readonly path: string
-  readonly mode: 0o644 | 0o755
+  readonly mode: number
   readonly bytes: Uint8Array
 }
 
 export type AgentCandidateRunFinalization =
   | {
       succeeded: true
-      receipt: CanonicalCandidateDocument<AgentCandidateRunReceiptV2>
+      receipt: CanonicalCandidateDocument<AgentCandidateRunReceipt>
       artifacts: {
         modelSettlement: AgentCandidateArtifactRef
+        executorCapture: AgentCandidateArtifactRef
         taskOutcome: AgentCandidateArtifactRef
         benchmarkResult: AgentCandidateArtifactRef
         runReceipt: AgentCandidateArtifactRef
@@ -554,7 +560,7 @@ export type AgentCandidateRunFinalization =
         termination?: AgentCandidateTermination
       }
       /** Independent evaluator-gateway usage, even when execution or trace capture failed. */
-      usage: AgentCandidateSpend | null
+      usage: AgentCandidateFixedSpend | null
     }
 
 /** Protected trace tags that bind a run to one prepared candidate execution. */
diff --git a/src/candidate-execution/verify.ts b/src/candidate-execution/verify.ts
index befde8e2..290e7ee5 100644
--- a/src/candidate-execution/verify.ts
+++ b/src/candidate-execution/verify.ts
@@ -19,6 +19,7 @@ import {
   omitTopLevelDigest,
 } from './digest'
 import { readCandidateGitHubResource, verifyCandidateCode } from './git-materialize'
+import { assertCandidateProfileExecutionSupport } from './profile'
 import {
   type AgentCandidateVerificationPorts,
   type VerifiedAgentCandidate,
@@ -33,6 +34,22 @@ interface VerifiedCandidateState {
 
 const verifiedCandidateState = new WeakMap<VerifiedAgentCandidate, VerifiedCandidateState>()
 
+/** Surfaces admitted by Runtime's verifier before an environment adapter is selected. */
+export const AGENT_CANDIDATE_EXECUTION_SUPPORT = Object.freeze({
+  outcomes: Object.freeze(['workspace', 'output'] as const),
+  code: Object.freeze(['disabled', 'no-op', 'git-patch'] as const),
+  memory: Object.freeze(['disabled', 'isolated'] as const),
+  knowledge: false,
+  profile: Object.freeze({
+    mcpTransports: Object.freeze(['stdio'] as const),
+    remoteMcp: false,
+    tools: false,
+    permissions: false,
+    modes: false,
+    confidential: false,
+  }),
+})
+
 /** Verifies every digest, resource, workspace, and Git object in a candidate bundle. */
 export async function verifyAgentCandidateBundle(
   input: unknown,
@@ -46,6 +63,12 @@ export async function verifyAgentCandidateBundle(
   }
   const canonicalBytes = canonicalCandidateBytes(withoutDigest)
   verifyBytes(canonicalBytes, parsed.digest, canonicalBytes.byteLength, 'candidate bundle')
+  assertCandidateProfileExecutionSupport(parsed.profile)
+  if (parsed.knowledge) {
+    throw new Error(
+      'candidate knowledge execution is unsupported; promote the approved exact candidate through agent-runtime/knowledge',
+    )
+  }
 
   const artifactBytes = new Map<string, Uint8Array>()
   const readArtifact = async (artifact: AgentCandidateCapturedArtifact): Promise<Uint8Array> => {
@@ -87,7 +110,6 @@ export async function verifyAgentCandidateBundle(
     artifactBytes.set(artifactCacheKey(parsed.execution.workspace.manifest), workspace.manifest)
     artifactBytes.set(artifactCacheKey(parsed.execution.workspace.archive), workspace.archive)
   }
-  if (parsed.knowledge) await readArtifact(parsed.knowledge.manifest)
   if (parsed.memory.mode === 'isolated' && parsed.memory.seed)
     await readArtifact(parsed.memory.seed)
 
diff --git a/src/candidate-execution/workspace-archive.ts b/src/candidate-execution/workspace-archive.ts
index 3ca40a71..2975c6ac 100644
--- a/src/candidate-execution/workspace-archive.ts
+++ b/src/candidate-execution/workspace-archive.ts
@@ -18,7 +18,7 @@ import { isAnyArrayBuffer, isSharedArrayBuffer } from 'node:util/types'
 
 import type {
   AgentCandidateCapturedArtifact,
-  AgentCandidateWorkspaceManifestMaterialV1,
+  AgentCandidateWorkspaceManifestMaterial,
   AgentCandidateWorkspaceSnapshotEvidence,
 } from '@tangle-network/agent-interface'
 import { type Entry, extract, type Pack, pack } from 'tar-stream'
@@ -91,7 +91,7 @@ interface WorkspaceArchiveRepositoryMetadataV1 {
 }
 
 interface DecodedWorkspaceArchive {
-  files: Array<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>
+  files: Array<{ path: string; mode: number; bytes: Uint8Array }>
   repository?: WorkspaceArchiveRepositoryV1
 }
 
@@ -260,7 +260,7 @@ async function captureRepository(
   root: string,
   limits: AgentCandidateWorkspaceArchiveLimits,
 ): Promise<{
-  files: Array<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>
+  files: Array<{ path: string; mode: number; bytes: Uint8Array }>
   repository: WorkspaceArchiveRepositoryV1
 }> {
   const stats = await lstat(root)
@@ -333,7 +333,7 @@ async function captureRepository(
 }
 
 async function encodeWorkspaceArchive(
-  files: ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>,
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
   repository: WorkspaceArchiveRepositoryV1 | undefined,
   maxArchiveBytes: number,
 ): Promise<Uint8Array> {
@@ -351,7 +351,7 @@ async function encodeWorkspaceArchive(
 }
 
 async function verifyCanonicalWorkspaceArchive(
-  files: ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>,
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
   repository: WorkspaceArchiveRepositoryV1 | undefined,
   expected: Uint8Array,
   maxArchiveBytes: number,
@@ -376,7 +376,7 @@ async function verifyCanonicalWorkspaceArchive(
 
 async function writeWorkspaceArchiveEntries(
   archive: Pack,
-  files: ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>,
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
   repository: WorkspaceArchiveRepositoryV1 | undefined,
 ): Promise<void> {
   for (const file of files) {
@@ -429,7 +429,7 @@ async function parseWorkspaceArchive(
           throw new Error('candidate workspace archive has an invalid file count')
         }
         const path = safeArchivePath(name.slice(workspaceEntryPrefix.length), limits.maxPathBytes)
-        if (entry.header.mode !== 0o644 && entry.header.mode !== 0o755) {
+        if (!isWorkspaceFileMode(entry.header.mode)) {
           throw new Error(`candidate workspace archive has an unsupported mode: ${path}`)
         }
         assertRetainedArchiveSize(bytes.byteLength, retainedEntryBytes, entry.header.size, limits)
@@ -572,7 +572,7 @@ function repositoryFromTar(
 async function writeTarEntry(
   archive: Pack,
   name: string,
-  mode: 0o600 | 0o644 | 0o755,
+  mode: number,
   bytes: Uint8Array,
 ): Promise<void> {
   await new Promise<void>((resolveEntry, rejectEntry) => {
@@ -668,7 +668,7 @@ async function prepareEmptyDestination(destination: string): Promise<void> {
 
 async function writeWorkspaceFiles(
   destination: string,
-  files: ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>,
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
 ): Promise<void> {
   for (const file of files) {
     const path = workspacePath(destination, file.path)
@@ -738,8 +738,8 @@ async function materializeRepository(
 }
 
 function workspaceManifest(
-  files: ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>,
-): AgentCandidateWorkspaceManifestMaterialV1 {
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
+): AgentCandidateWorkspaceManifestMaterial {
   return {
     schemaVersion: 1,
     kind: 'agent-candidate-workspace-manifest',
@@ -753,7 +753,7 @@ function workspaceManifest(
 }
 
 function assertArchiveMatchesSnapshot(
-  files: ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>,
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
   snapshot: AgentCandidateWorkspaceSnapshotEvidence,
 ): void {
   const material = workspaceManifest(files)
@@ -773,7 +773,7 @@ function assertArchiveMatchesSnapshot(
 }
 
 function assertWorkspaceFilesWithinLimits(
-  files: ReadonlyArray<{ path: string; mode: 0o644 | 0o755; bytes: Uint8Array }>,
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
   limits: AgentCandidateWorkspaceArchiveLimits,
 ): void {
   if (files.length > limits.maxFiles) {
@@ -791,7 +791,7 @@ function assertWorkspaceFilesWithinLimits(
       'candidate workspace paths must be unique and sorted',
     )
     previousPath = path
-    if (file.mode !== 0o644 && file.mode !== 0o755) {
+    if (!isWorkspaceFileMode(file.mode)) {
       throw new Error(`candidate workspace file has unsupported mode: ${path}`)
     }
     if (file.bytes.byteLength > limits.maxFileBytes) {
@@ -833,7 +833,7 @@ function normalizeWorkspaceFiles(
     const path = safeArchivePath(descriptors.path?.value, limits.maxPathBytes)
     const mode = descriptors.mode?.value
     const inputBytes = descriptors.bytes?.value
-    if (mode !== 0o644 && mode !== 0o755) {
+    if (!isWorkspaceFileMode(mode)) {
       throw new Error(`candidate workspace file has unsupported mode: ${path}`)
     }
     const view = inspectWorkspaceBytes(inputBytes, path)
@@ -854,6 +854,10 @@ function normalizeWorkspaceFiles(
   return files
 }
 
+function isWorkspaceFileMode(value: unknown): value is number {
+  return Number.isSafeInteger(value) && (value as number) >= 0 && (value as number) <= 0o777
+}
+
 function workspaceLimits(
   overrides: Partial<AgentCandidateWorkspaceArchiveLimits> | undefined,
 ): AgentCandidateWorkspaceArchiveLimits {
diff --git a/src/intelligence/improvement-cycle.ts b/src/intelligence/improvement-cycle.ts
index 65ffd3d8..051f806d 100644
--- a/src/intelligence/improvement-cycle.ts
+++ b/src/intelligence/improvement-cycle.ts
@@ -1,11 +1,32 @@
+import { type PairedArmRow, pairArms } from '@tangle-network/agent-eval'
 import { type AnalystFinding, assertNoJudgeVerdict } from '@tangle-network/agent-eval/analyst'
-import type { Scenario, SelfImproveResult } from '@tangle-network/agent-eval/contract'
+import {
+  assertCodeSurfaceIdentity,
+  codeSurfaceIdentityMaterial,
+} from '@tangle-network/agent-eval/campaign'
+import type { CodeSurface, Scenario, SelfImproveResult } from '@tangle-network/agent-eval/contract'
+import { pairedBootstrap } from '@tangle-network/agent-eval/reporting'
 import type {
   AgentCandidateBundle,
+  AgentImprovementMeasuredComparison,
+  AgentImprovementProposal,
+  AgentImprovementReview,
+  AgentImprovementReviewDecision,
+  AgentImprovementSurface,
   AgentProfile,
+  CandidateExecutionEvidence,
   Sha256Digest,
 } from '@tangle-network/agent-interface'
-import { agentCandidateBundleSchema } from '@tangle-network/agent-interface'
+import {
+  agentCandidateBundleSchema,
+  agentCandidateMaterializationReceiptSchema,
+  agentCandidateRunReceiptSchema,
+  agentImprovementMeasuredComparisonSchema,
+  agentImprovementProposalSchema,
+  agentImprovementReviewSchema,
+  candidateExecutionEvidenceSchema,
+} from '@tangle-network/agent-interface'
+import { materializeCandidateProfile } from '@tangle-network/agent-profile-materialize'
 import { runAnalystLoop } from '../analyst-loop'
 import type { RunAnalystLoopOpts, RunAnalystLoopResult } from '../analyst-loop/types'
 import {
@@ -13,10 +34,12 @@ import {
   sealAgentCandidateBundle,
 } from '../candidate-execution/bundle'
 import {
+  canonicalCandidateBytes,
   canonicalCandidateDigest,
   canonicalCandidateDocument,
   immutableCandidateValue,
   omitTopLevelDigest,
+  verifyCanonicalCandidateDocument,
 } from '../candidate-execution/digest'
 import {
   type ExecutePreparedAgentCandidateOptions,
@@ -27,7 +50,10 @@ import {
   prepareAgentCandidateExecution,
 } from '../candidate-execution/prepare'
 import {
-  assertCandidateProfileBinding,
+  agentCandidateProfileAsAgentProfile,
+  candidateMaterializerHarness,
+  createAgentCandidateProfileActivation,
+  parseAgentCandidateProfileActivation,
   parseExactAgentProfile,
 } from '../candidate-execution/profile'
 import type {
@@ -35,7 +61,10 @@ import type {
   AgentCandidateRunFinalization,
   AgentCandidateTaskExecution,
 } from '../candidate-execution/types'
-import { verifyAgentCandidateBundle } from '../candidate-execution/verify'
+import {
+  verifiedResourceTextByDigest,
+  verifyAgentCandidateBundle,
+} from '../candidate-execution/verify'
 import {
   applyImprovementWinnerToProfile,
   type ImproveOptions,
@@ -44,77 +73,21 @@ import {
   improve,
 } from '../improvement/improve'
 
-export type AgentImprovementEvaluation<TScenario extends Scenario, TArtifact> = Pick<
-  SelfImproveResult<TScenario, TArtifact>,
-  | 'baseline'
-  | 'winner'
-  | 'lift'
-  | 'diff'
-  | 'provenance'
-  | 'gateDecision'
-  | 'generationsExplored'
-  | 'durationMs'
-  | 'totalCostUsd'
-  | 'insight'
-  | 'power'
->
-
-export interface AgentImprovementProposal<
-  TScenario extends Scenario = Scenario,
-  TArtifact = unknown,
-> {
-  schemaVersion: 1
-  kind: 'agent-improvement-proposal'
-  runId: string
-  surface: ImproveSurface
-  proposedAt: string
-  baselineProfile: AgentProfile
-  baselineProfileHash: string
-  candidateProfile: AgentProfile
-  candidateProfileHash: string
-  findings: AnalystFinding[]
-  evaluation: AgentImprovementEvaluation<TScenario, TArtifact>
-  candidateBundle?: AgentCandidateBundle
-  digest: Sha256Digest
-}
-
-export type AgentImprovementReviewDecision = 'approve' | 'reject' | 'request-changes'
-
-export interface AgentImprovementReview {
-  schemaVersion: 1
-  kind: 'agent-improvement-review'
-  proposalDigest: Sha256Digest
-  candidateBundleDigest?: Sha256Digest
-  decision: AgentImprovementReviewDecision
-  reviewedBy: string
-  reviewedAt: string
-  reason: string
-  feedback?: string
-  digest: Sha256Digest
-}
-
-export interface CandidateExecutionEvidence {
-  proposalDigest: Sha256Digest
-  reviewDigest: Sha256Digest
-  bundleDigest: Sha256Digest
-  executionId: string
-  executionPlanDigest: Sha256Digest
-  materializationReceiptDigest: Sha256Digest
-  succeeded: boolean
-  runReceiptDigest?: Sha256Digest
-}
+export type {
+  AgentImprovementMeasuredComparison,
+  AgentImprovementProposal,
+  AgentImprovementReview,
+  AgentImprovementReviewDecision,
+  CandidateExecutionEvidence,
+} from '@tangle-network/agent-interface'
 
 export interface ProposeAgentImprovementOptions<TScenario extends Scenario, TArtifact> {
   runId: string
   profile: AgentProfile
   analysis: Omit<RunAnalystLoopOpts, 'runId' | 'improvementAdapter' | 'autoApply'>
   improvement: ImproveOptions<TScenario, TArtifact>
-  /**
-   * Optional environment adapter that freezes an executable bundle after the
-   * measured comparison recommends the candidate. Return the sealed output of
-   * `buildAgentCandidateBundle` directly, or a low-level digest-free input.
-   */
-  buildCandidate?: (input: {
+  /** Freezes the measured winner into the exact bundle reviewed for execution. */
+  buildCandidate: (input: {
     analysis: RunAnalystLoopResult
     improvement: ImproveResult<TScenario, TArtifact>
   }) =>
@@ -127,20 +100,28 @@ export interface ProposeAgentImprovementOptions<TScenario extends Scenario, TArt
 export interface ProposeAgentImprovementResult<TScenario extends Scenario, TArtifact> {
   analysis: RunAnalystLoopResult
   improvement: ImproveResult<TScenario, TArtifact>
-  proposal: AgentImprovementProposal<TScenario, TArtifact>
+  proposal: AgentImprovementProposal
 }
 
-export interface CreateAgentImprovementProposalOptions<TScenario extends Scenario, TArtifact> {
+export interface CreateAgentImprovementProposalOptions {
   runId: string
-  surface: ImproveSurface
   baselineProfile: AgentProfile
-  candidateProfile: AgentProfile
   findings: readonly AnalystFinding[]
-  evaluation: SelfImproveResult<TScenario, TArtifact>
-  candidateBundle?: AgentCandidateBundleInput | AgentCandidateBundle
+  evaluation: AgentImprovementMeasuredComparison
+  candidateBundle: AgentCandidateBundleInput | AgentCandidateBundle
   now?: () => Date
 }
 
+export interface CreateAgentImprovementMeasuredComparisonOptions<
+  TScenario extends Scenario,
+  TArtifact,
+> {
+  result: SelfImproveResult<TScenario, TArtifact>
+  measuredSurface: ImproveSurface
+  baselineProfile: AgentProfile
+  candidateBundle: AgentCandidateBundle
+}
+
 export interface ReviewAgentImprovementInput {
   decision: AgentImprovementReviewDecision
   reviewedBy: string
@@ -163,12 +144,24 @@ export interface ExecuteApprovedAgentCandidateOptions {
   execution: ExecutePreparedAgentCandidateOptions
 }
 
-export interface ExecuteApprovedAgentCandidateResult {
-  finalization: AgentCandidateRunFinalization
-  evidence: CandidateExecutionEvidence
+export type ExecuteApprovedAgentCandidateResult =
+  | {
+      finalization: Extract<AgentCandidateRunFinalization, { succeeded: false }>
+      evidence?: never
+    }
+  | {
+      finalization: Extract<AgentCandidateRunFinalization, { succeeded: true }>
+      evidence: CandidateExecutionEvidence
+    }
+
+export interface VerifyCandidateExecutionEvidenceOptions {
+  proposal: AgentImprovementProposal
+  review: AgentImprovementReview
+  expectedCount: number
+  resolvedResources?: ReadonlyMap<Sha256Digest, string>
 }
 
-/** Analyze one run and produce one measured, review-only improvement proposal. */
+/** Analyze one run and freeze its measured winner into one exact proposal. */
 export async function proposeAgentImprovement<TScenario extends Scenario, TArtifact>(
   options: ProposeAgentImprovementOptions<TScenario, TArtifact>,
 ): Promise<ProposeAgentImprovementResult<TScenario, TArtifact>> {
@@ -182,27 +175,31 @@ export async function proposeAgentImprovement<TScenario extends Scenario, TArtif
       `proposeAgentImprovement cannot write ${writeBackSurface} before human approval`,
     )
   }
-  const analysis = await runAnalystLoop({
-    ...options.analysis,
-    runId: options.runId,
-  })
+  const analysis = await runAnalystLoop(
+    Object.assign({}, options.analysis, { runId: options.runId }),
+  )
   const findings = assertNoJudgeVerdict(
     analysis.analystResult.findings,
     'proposeAgentImprovement findings',
   )
   const improvement = await improve(options.profile, [...findings], options.improvement)
-  const candidateBundle =
-    improvement.shipped && options.buildCandidate
-      ? await options.buildCandidate({ analysis, improvement })
-      : undefined
+  if (!improvement.shipped) {
+    throw new Error('agent improvement proposal requires a passing measured comparison')
+  }
+  const candidateBundle = sealBuiltCandidate(
+    await options.buildCandidate({ analysis, improvement }),
+  )
   const proposal = createAgentImprovementProposal({
     runId: options.runId,
-    surface: options.improvement.surface ?? 'prompt',
     baselineProfile: options.profile,
-    candidateProfile: improvement.profile,
     findings,
-    evaluation: improvement.raw,
-    ...(candidateBundle ? { candidateBundle } : {}),
+    evaluation: createAgentImprovementMeasuredComparison({
+      result: improvement.raw,
+      measuredSurface: options.improvement.surface ?? 'prompt',
+      baselineProfile: options.profile,
+      candidateBundle,
+    }),
+    candidateBundle,
     ...(options.now ? { now: options.now } : {}),
   })
   return { analysis, improvement, proposal }
@@ -213,53 +210,42 @@ export async function proposeAgentImprovement<TScenario extends Scenario, TArtif
  * contract. Products that run analysis or evaluation in separate workers use
  * this constructor instead of rerunning either phase or rebuilding digests.
  */
-export function createAgentImprovementProposal<TScenario extends Scenario, TArtifact>(
-  options: CreateAgentImprovementProposalOptions<TScenario, TArtifact>,
-): AgentImprovementProposal<TScenario, TArtifact> {
+export function createAgentImprovementProposal(
+  options: CreateAgentImprovementProposalOptions,
+): AgentImprovementProposal {
   const findings = assertNoJudgeVerdict(
     [...options.findings],
     'createAgentImprovementProposal findings',
   )
-  const candidateBundle = options.candidateBundle
-    ? sealBuiltCandidate(options.candidateBundle)
-    : undefined
+  const candidateBundle = sealBuiltCandidate(options.candidateBundle)
   const baselineProfile = parseExactAgentProfile(
     options.baselineProfile,
     'proposal baseline profile',
   )
-  const candidateProfile = parseExactAgentProfile(
-    options.candidateProfile,
-    'proposal candidate profile',
-  )
-  assertMeasuredProfileBinding({
-    surface: options.surface,
-    baselineProfile,
-    candidateProfile,
-    evaluation: options.evaluation,
-    hasExecutableBundle: candidateBundle !== undefined,
-  })
-  if (candidateBundle) {
-    if (options.evaluation.gateDecision !== 'ship') {
-      throw new Error('executable candidate bundle requires a passing measured comparison')
-    }
-    assertCandidateProfileBinding(candidateProfile, candidateBundle.profile)
+  const candidateProfile = agentCandidateProfileAsAgentProfile(candidateBundle.profile)
+  const evaluation = agentImprovementMeasuredComparisonSchema.parse(options.evaluation)
+  if (evaluation.decision.outcome !== 'ship') {
+    throw new Error('agent improvement proposal requires a passing measured comparison')
   }
+  assertSupportedCandidateBundle(candidateBundle)
+  assertMeasuredCandidateBinding(evaluation, baselineProfile, candidateBundle)
+  const changedSurfaces = deriveChangedSurfaces(baselineProfile, candidateProfile, candidateBundle)
   const withoutDigest = {
     schemaVersion: 1 as const,
     kind: 'agent-improvement-proposal' as const,
     runId: options.runId,
-    surface: options.surface,
+    changedSurfaces,
     proposedAt: (options.now ?? (() => new Date()))().toISOString(),
     baselineProfile,
-    baselineProfileHash: canonicalCandidateDigest(baselineProfile),
-    candidateProfile,
-    candidateProfileHash: canonicalCandidateDigest(candidateProfile),
-    findings: canonicalJsonValue([...findings]),
-    evaluation: canonicalJsonValue(improvementEvaluation(options.evaluation)),
-    ...(candidateBundle ? { candidateBundle } : {}),
+    findings: immutableCandidateValue([
+      ...findings,
+    ]) as unknown as AgentImprovementProposal['findings'],
+    evaluation,
+    candidateBundle,
   }
-  return canonicalCandidateDocument<AgentImprovementProposal<TScenario, TArtifact>>(withoutDigest)
-    .value
+  return agentImprovementProposalSchema.parse(
+    canonicalCandidateDocument<AgentImprovementProposal>(withoutDigest).value,
+  )
 }
 
 /** Persist an approve/reject/change-request decision bound to one exact proposal. */
@@ -271,25 +257,24 @@ export function reviewAgentImprovementProposal(
   if (!input.reviewedBy.trim()) throw new Error('candidate review requires reviewedBy')
   if (!input.reason.trim()) throw new Error('candidate review requires a reason')
   if (input.decision === 'approve') {
-    if (proposal.evaluation.gateDecision !== 'ship') {
+    if (proposal.evaluation.decision.outcome !== 'ship') {
       throw new Error('candidate cannot be approved without a passing measured comparison')
     }
-    if (!proposal.candidateBundle) {
-      throw new Error('candidate cannot be approved until an executable bundle is sealed')
-    }
   }
   const withoutDigest = {
     schemaVersion: 1 as const,
     kind: 'agent-improvement-review' as const,
     proposalDigest: proposal.digest,
-    ...(proposal.candidateBundle ? { candidateBundleDigest: proposal.candidateBundle.digest } : {}),
+    candidateBundleDigest: proposal.candidateBundle.digest,
     decision: input.decision,
     reviewedBy: input.reviewedBy,
     reviewedAt: (input.now ?? (() => new Date()))().toISOString(),
     reason: input.reason,
     ...(input.feedback === undefined ? {} : { feedback: input.feedback }),
   }
-  return canonicalCandidateDocument<AgentImprovementReview>(withoutDigest).value
+  return agentImprovementReviewSchema.parse(
+    canonicalCandidateDocument<AgentImprovementReview>(withoutDigest).value,
+  )
 }
 
 /** Verify, materialize, run, grade, and receipt only the exact approved bundle. */
@@ -303,7 +288,6 @@ export async function executeApprovedAgentCandidate(
     throw new Error('candidate approval does not match the proposed improvement')
   }
   const bundle = proposal.candidateBundle
-  if (!bundle) throw new Error('approved proposal does not contain an executable candidate bundle')
   if (review.candidateBundleDigest !== bundle.digest) {
     throw new Error('candidate approval does not match the executable bundle')
   }
@@ -319,103 +303,311 @@ export async function executeApprovedAgentCandidate(
     options.preparation,
   )
   const finalization = await executePreparedAgentCandidate(prepared, options.execution)
+  if (!finalization.succeeded) return { finalization }
+  const evidence = canonicalCandidateDocument<CandidateExecutionEvidence>({
+    schemaVersion: 1,
+    kind: 'agent-candidate-execution-evidence',
+    proposalDigest: proposal.digest,
+    reviewDigest: review.digest,
+    executionId: prepared.executionId,
+    succeeded: true,
+    materializationReceipt: prepared.materializationReceipt.value,
+    profileActivation: prepared.profileActivation,
+    receipt: finalization.receipt.value,
+  }).value
+  const verifiedEvidence = verifyCandidateExecutionEvidence([evidence], {
+    proposal,
+    review,
+    expectedCount: 1,
+    resolvedResources: verifiedResourceTextByDigest(verified),
+  })[0]!
   return {
     finalization,
-    evidence: {
-      proposalDigest: proposal.digest,
-      reviewDigest: review.digest,
-      bundleDigest: bundle.digest,
-      executionId: prepared.executionId,
-      executionPlanDigest: prepared.executionPlan.value.digest,
-      materializationReceiptDigest: prepared.materializationReceipt.digest,
-      succeeded: finalization.succeeded,
-      ...(finalization.succeeded ? { runReceiptDigest: finalization.receipt.digest } : {}),
-    },
+    evidence: verifiedEvidence,
   }
 }
 
-/** Validate a proposal's schema, profile, sealed bundle, and canonical digest. */
-export function verifyAgentImprovementProposal(input: unknown): AgentImprovementProposal {
+/** Verify approval, receipts, uniqueness, and the exact native profile files executed. */
+export function verifyCandidateExecutionEvidence(
+  input: unknown,
+  options: VerifyCandidateExecutionEvidenceOptions,
+): CandidateExecutionEvidence[] {
+  const proposal = verifyAgentImprovementProposal(options.proposal)
+  const review = verifyAgentImprovementReview(options.review)
+  if (review.decision !== 'approve') throw new Error('candidate review is not an approval')
+  if (
+    review.proposalDigest !== proposal.digest ||
+    review.candidateBundleDigest !== proposal.candidateBundle.digest
+  ) {
+    throw new Error('candidate review does not bind the proposed bundle')
+  }
+  if (!Number.isSafeInteger(options.expectedCount) || options.expectedCount < 1) {
+    throw new Error('candidate execution evidence expectedCount must be a positive integer')
+  }
+  if (!Array.isArray(input) || input.length !== options.expectedCount) {
+    throw new Error(
+      `candidate execution evidence must contain exactly ${options.expectedCount} rows`,
+    )
+  }
+  const executionIds = new Set<string>()
+  const runReceiptDigests = new Set<Sha256Digest>()
+  const verified = input.map((entry) => {
+    const evidence = verifyCanonicalCandidateDocument(
+      candidateExecutionEvidenceSchema.parse(entry),
+      'candidate execution evidence',
+    )
+    const receipt = verifyCanonicalCandidateDocument(
+      agentCandidateRunReceiptSchema.parse(evidence.receipt),
+      'candidate execution run receipt',
+    )
+    const materializationReceipt = verifyCanonicalCandidateDocument(
+      agentCandidateMaterializationReceiptSchema.parse(evidence.materializationReceipt),
+      'candidate materialization receipt',
+    )
+    assertEvidenceMaterialDigest(materializationReceipt.profilePlan, 'candidate profile plan')
+    const profilePlan = materializeCandidateProfile(
+      proposal.candidateBundle.profile,
+      candidateMaterializerHarness(materializationReceipt.harness),
+      { resolvedResources: options.resolvedResources },
+    )
+    const profileActivation = parseAgentCandidateProfileActivation(
+      evidence.profileActivation,
+      materializationReceipt.profilePlan.digest,
+    )
+    const regenerated = createAgentCandidateProfileActivation(
+      profilePlan,
+      materializationReceipt.profilePlan,
+    )
+    if (regenerated.digest !== profileActivation.digest) {
+      throw new Error('candidate profile plan files do not match the executed materialization')
+    }
+    assertEvidenceMaterialDigest(materializationReceipt.executionPlan, 'candidate execution plan')
+    assertEvidenceMaterialDigest(receipt.modelSettlement, 'candidate model settlement')
+    assertEvidenceMaterialDigest(receipt.taskOutcome, 'candidate task outcome')
+    assertEvidenceMaterialDigest(receipt.benchmarkResult, 'candidate benchmark result')
+    if (
+      evidence.proposalDigest !== proposal.digest ||
+      evidence.reviewDigest !== review.digest ||
+      receipt.bundleDigest !== proposal.candidateBundle.digest ||
+      materializationReceipt.bundleDigest !== proposal.candidateBundle.digest ||
+      receipt.materializationReceiptDigest !== materializationReceipt.digest ||
+      receipt.executionPlanDigest !== materializationReceipt.executionPlan.digest ||
+      evidence.executionId !== materializationReceipt.executionPlan.material.executionId
+    ) {
+      throw new Error('candidate execution evidence does not bind the approved candidate')
+    }
+    if (receipt.termination.kind !== 'exit' || receipt.termination.exitCode !== 0) {
+      throw new Error('candidate execution evidence is not a successful process execution')
+    }
+    if (executionIds.has(evidence.executionId)) {
+      throw new Error('candidate execution evidence reuses an execution id')
+    }
+    if (runReceiptDigests.has(receipt.digest)) {
+      throw new Error('candidate execution evidence reuses a run receipt')
+    }
+    executionIds.add(evidence.executionId)
+    runReceiptDigests.add(receipt.digest)
+    return immutableCandidateValue(evidence)
+  })
+  return verified
+}
+
+function assertEvidenceMaterialDigest(
+  evidence: {
+    digest: Sha256Digest
+    material: unknown
+    artifact: { sha256: Sha256Digest; byteLength: number }
+  },
+  label: string,
+): void {
+  const bytes = canonicalCandidateBytes(evidence.material)
   if (
-    !isRecord(input) ||
-    input.kind !== 'agent-improvement-proposal' ||
-    input.schemaVersion !== 1
+    canonicalCandidateDigest(evidence.material) !== evidence.digest ||
+    evidence.artifact.sha256 !== evidence.digest ||
+    evidence.artifact.byteLength !== bytes.byteLength
   ) {
-    throw new Error('invalid agent improvement proposal')
+    throw new Error(`${label} digest does not match its canonical material`)
   }
-  const proposal = input as unknown as AgentImprovementProposal
+}
+
+/** Validate a proposal's schema, profile, sealed bundle, and canonical digest. */
+export function verifyAgentImprovementProposal(input: unknown): AgentImprovementProposal {
+  const proposal = agentImprovementProposalSchema.parse(input)
   const parsedBaselineProfile = parseExactAgentProfile(
     proposal.baselineProfile,
     'proposal baseline profile',
   )
-  if (proposal.baselineProfileHash !== canonicalCandidateDigest(parsedBaselineProfile)) {
-    throw new Error('proposal baselineProfileHash does not match baselineProfile')
+  const parsedBundle = agentCandidateBundleSchema.parse(proposal.candidateBundle)
+  const candidateBundle = sealAgentCandidateBundle(omitTopLevelDigest(parsedBundle))
+  if (candidateBundle.digest !== parsedBundle.digest)
+    throw new Error('proposal candidate bundle digest is invalid')
+  const candidateProfile = agentCandidateProfileAsAgentProfile(candidateBundle.profile)
+  if (proposal.evaluation.decision.outcome !== 'ship') {
+    throw new Error('agent improvement proposal requires a passing measured comparison')
   }
-  const parsedProfile = parseExactAgentProfile(
-    proposal.candidateProfile,
-    'proposal candidate profile',
+  assertSupportedCandidateBundle(candidateBundle)
+  assertMeasuredCandidateBinding(proposal.evaluation, parsedBaselineProfile, candidateBundle)
+  const changedSurfaces = deriveChangedSurfaces(
+    parsedBaselineProfile,
+    candidateProfile,
+    candidateBundle,
   )
-  if (proposal.candidateProfileHash !== canonicalCandidateDigest(parsedProfile)) {
-    throw new Error('proposal candidateProfileHash does not match candidateProfile')
-  }
-  assertMeasuredProfileBinding({
-    surface: proposal.surface,
-    baselineProfile: parsedBaselineProfile,
-    candidateProfile: parsedProfile,
-    evaluation: proposal.evaluation,
-    hasExecutableBundle: proposal.candidateBundle !== undefined,
-  })
-  if (!Array.isArray(proposal.findings)) throw new Error('proposal findings must be an array')
-  if (!isSha256Digest(proposal.digest)) throw new Error('proposal digest is invalid')
-  if (proposal.candidateBundle) {
-    const parsedBundle = agentCandidateBundleSchema.parse(proposal.candidateBundle)
-    const sealed = sealAgentCandidateBundle(omitTopLevelDigest(parsedBundle))
-    if (sealed.digest !== parsedBundle.digest)
-      throw new Error('proposal candidate bundle digest is invalid')
-    assertCandidateProfileBinding(parsedProfile, sealed.profile)
-  }
-  const actual = canonicalCandidateDigest(omitTopLevelDigest(proposal))
-  if (actual !== proposal.digest)
-    throw new Error('agent improvement proposal digest does not match')
-  return immutableCandidateValue(proposal)
-}
-
-function assertMeasuredProfileBinding(input: {
+  if (!sameOrderedValues(proposal.changedSurfaces, changedSurfaces)) {
+    throw new Error(
+      'proposal changed surfaces do not match the exact baseline and candidate bundle',
+    )
+  }
+  return verifyCanonicalCandidateDocument(proposal, 'agent improvement proposal')
+}
+
+function assertMeasuredCandidateBinding(
+  evaluation: AgentImprovementMeasuredComparison,
+  baselineProfile: AgentProfile,
+  candidateBundle: AgentCandidateBundle,
+): void {
+  if (evaluation.baselineProfileDigest !== canonicalCandidateDigest(baselineProfile)) {
+    throw new Error('measured comparison does not bind the included baseline profile')
+  }
+  if (evaluation.candidateBundleDigest !== candidateBundle.digest) {
+    throw new Error('measured comparison does not bind the exact candidate bundle')
+  }
+}
+
+function assertRawMeasuredWinnerBinding(input: {
   surface: ImproveSurface
   baselineProfile: AgentProfile
-  candidateProfile: AgentProfile
-  evaluation: Pick<AgentImprovementEvaluation<Scenario, unknown>, 'gateDecision' | 'winner'>
-  hasExecutableBundle: boolean
+  winner: unknown
+  candidateBundle: AgentCandidateBundle
 }): void {
+  const candidateProfile = agentCandidateProfileAsAgentProfile(input.candidateBundle.profile)
   const baselineDigest = canonicalCandidateDigest(input.baselineProfile)
-  if (input.evaluation.gateDecision !== 'ship') {
-    if (canonicalCandidateDigest(input.candidateProfile) !== baselineDigest) {
-      throw new Error('non-shipping evaluation must retain the baseline candidate profile')
+  if (input.surface === 'code') {
+    if (canonicalCandidateDigest(candidateProfile) !== baselineDigest) {
+      throw new Error('code improvement must retain the measured profile')
+    }
+    assertMeasuredCodeBundle(input.winner, input.candidateBundle)
+  } else if (input.surface === 'memory') {
+    throw new Error(
+      'memory improvement proposals require a content-addressed bundle binding, which is unsupported',
+    )
+  } else {
+    const measured = measuredWinnerProfile(input.baselineProfile, input.surface, input.winner)
+    if (!measured) {
+      throw new Error(
+        'skill-document improvement proposals require a content-addressed bundle binding, which is unsupported',
+      )
+    }
+    if (canonicalCandidateDigest(candidateProfile) !== canonicalCandidateDigest(measured)) {
+      throw new Error('proposal candidate profile does not match the measured winner surface')
     }
-    return
   }
 
-  const measured = measuredWinnerProfile(
+  const changedSurfaces = deriveChangedSurfaces(
     input.baselineProfile,
-    input.surface,
-    input.evaluation.winner.surface,
+    candidateProfile,
+    input.candidateBundle,
   )
-  if (!measured) {
-    if (input.hasExecutableBundle) {
-      throw new Error(
-        `executable '${input.surface}' candidate cannot be bound to its measured winner surface`,
-      )
+  if (input.surface === 'agent-profile') {
+    if (changedSurfaces.includes('code') || changedSurfaces.includes('knowledge')) {
+      throw new Error('agent-profile measurement does not cover code or knowledge bundle changes')
     }
-    if (canonicalCandidateDigest(input.candidateProfile) !== baselineDigest) {
-      throw new Error(
-        `unbound '${input.surface}' improvement must retain the baseline candidate profile`,
-      )
-    }
-    return
+  } else if (changedSurfaces.length !== 1 || changedSurfaces[0] !== input.surface) {
+    throw new Error(
+      `measured '${input.surface}' winner does not cover candidate changes: ${changedSurfaces.join(', ')}`,
+    )
+  }
+}
+
+function assertSupportedCandidateBundle(candidateBundle: AgentCandidateBundle): void {
+  if (candidateBundle.memory.mode !== 'disabled') {
+    throw new Error(
+      'memory improvement proposals require a content-addressed policy binding, which is unsupported',
+    )
+  }
+}
+
+const CHANGED_SURFACE_ORDER: readonly AgentImprovementSurface[] = [
+  'prompt',
+  'skills',
+  'tools',
+  'mcp',
+  'hooks',
+  'subagents',
+  'agent-profile',
+  'memory',
+  'code',
+  'knowledge',
+]
+
+function deriveChangedSurfaces(
+  baseline: AgentProfile,
+  candidate: AgentProfile,
+  bundle: AgentCandidateBundle,
+): [AgentImprovementSurface, ...AgentImprovementSurface[]] {
+  const changed = new Set<AgentImprovementSurface>()
+  const pairs: Array<[AgentImprovementSurface, unknown, unknown]> = [
+    [
+      'prompt',
+      { prompt: baseline.prompt ?? null, instructions: baseline.resources?.instructions ?? null },
+      { prompt: candidate.prompt ?? null, instructions: candidate.resources?.instructions ?? null },
+    ],
+    ['skills', baseline.resources?.skills ?? null, candidate.resources?.skills ?? null],
+    [
+      'tools',
+      { tools: baseline.tools ?? null, resources: baseline.resources?.tools ?? null },
+      { tools: candidate.tools ?? null, resources: candidate.resources?.tools ?? null },
+    ],
+    ['mcp', baseline.mcp ?? null, candidate.mcp ?? null],
+    ['hooks', baseline.hooks ?? null, candidate.hooks ?? null],
+    [
+      'subagents',
+      { subagents: baseline.subagents ?? null, resources: baseline.resources?.agents ?? null },
+      { subagents: candidate.subagents ?? null, resources: candidate.resources?.agents ?? null },
+    ],
+    ['agent-profile', opaqueProfileSlice(baseline), opaqueProfileSlice(candidate)],
+  ]
+  for (const [surface, before, after] of pairs) {
+    if (!sameCanonicalValue(before, after)) changed.add(surface)
   }
-  if (canonicalCandidateDigest(input.candidateProfile) !== canonicalCandidateDigest(measured)) {
-    throw new Error('proposal candidate profile does not match the measured winner surface')
+  if (bundle.memory.mode !== 'disabled') changed.add('memory')
+  if (bundle.code.kind !== 'disabled') changed.add('code')
+  if (bundle.knowledge) changed.add('knowledge')
+  const ordered = CHANGED_SURFACE_ORDER.filter((surface) => changed.has(surface))
+  if (ordered.length === 0) {
+    throw new Error('candidate bundle does not contain a change from the included baseline profile')
   }
+  return ordered as [AgentImprovementSurface, ...AgentImprovementSurface[]]
+}
+
+function opaqueProfileSlice(profile: AgentProfile): unknown {
+  const {
+    prompt: _prompt,
+    tools: _tools,
+    mcp: _mcp,
+    hooks: _hooks,
+    subagents: _subagents,
+    resources,
+    ...opaqueProfile
+  } = profile
+  const {
+    instructions: _instructions,
+    skills: _skills,
+    tools: _resourceTools,
+    agents: _agents,
+    ...opaqueResources
+  } = resources ?? {}
+  return immutableCandidateValue({
+    ...opaqueProfile,
+    ...(Object.keys(opaqueResources).length > 0 ? { resources: opaqueResources } : {}),
+  })
+}
+
+function sameCanonicalValue(left: unknown, right: unknown): boolean {
+  return canonicalCandidateDigest(left) === canonicalCandidateDigest(right)
+}
+
+function sameOrderedValues<T>(left: readonly T[], right: readonly T[]): boolean {
+  return left.length === right.length && left.every((value, index) => value === right[index])
 }
 
 function measuredWinnerProfile(
@@ -423,7 +615,6 @@ function measuredWinnerProfile(
   surface: ImproveSurface,
   winner: unknown,
 ): AgentProfile | null {
-  if (surface === 'code' || surface === 'memory') return null
   if (typeof winner !== 'string') {
     throw new Error(`measured '${surface}' winner is not a serializable profile surface`)
   }
@@ -439,6 +630,28 @@ function measuredWinnerProfile(
   }
 }
 
+function assertMeasuredCodeBundle(winner: unknown, candidateBundle: AgentCandidateBundle): void {
+  assertCodeSurfaceIdentity(winner)
+  if (candidateBundle.code.kind !== 'git-patch') {
+    throw new Error('code improvement bundle must contain the measured git patch')
+  }
+  const bundledSurface: CodeSurface = {
+    ...winner,
+    baseCommit: candidateBundle.code.baseCommit,
+    baseTree: candidateBundle.code.baseTree,
+    candidateTree: candidateBundle.code.candidateTree,
+    patch: {
+      format: candidateBundle.code.patch.format,
+      sha256: candidateBundle.code.patch.artifact.sha256,
+      byteLength: candidateBundle.code.patch.artifact.byteLength,
+    },
+  }
+  assertCodeSurfaceIdentity(bundledSurface)
+  if (codeSurfaceIdentityMaterial(winner) !== codeSurfaceIdentityMaterial(bundledSurface)) {
+    throw new Error('proposal candidate bundle does not match the measured code winner')
+  }
+}
+
 function sealBuiltCandidate(
   input: AgentCandidateBundleInput | AgentCandidateBundle,
 ): AgentCandidateBundle {
@@ -453,67 +666,377 @@ function sealBuiltCandidate(
 
 /** Validate a review's decision fields and canonical digest. */
 export function verifyAgentImprovementReview(input: unknown): AgentImprovementReview {
-  if (!isRecord(input) || input.kind !== 'agent-improvement-review' || input.schemaVersion !== 1) {
-    throw new Error('invalid agent improvement review')
+  const review = agentImprovementReviewSchema.parse(input)
+  return verifyCanonicalCandidateDocument(review, 'agent improvement review')
+}
+
+/** Convert agent-eval's paired result into the portable Interface comparison. */
+export function createAgentImprovementMeasuredComparison<TScenario extends Scenario, TArtifact>(
+  options: CreateAgentImprovementMeasuredComparisonOptions<TScenario, TArtifact>,
+): AgentImprovementMeasuredComparison {
+  const { result } = options
+  const baselineProfile = parseExactAgentProfile(
+    options.baselineProfile,
+    'measured baseline profile',
+  )
+  const candidateBundle = sealBuiltCandidate(options.candidateBundle)
+  assertSupportedCandidateBundle(candidateBundle)
+  assertRawMeasuredWinnerBinding({
+    surface: options.measuredSurface,
+    baselineProfile,
+    winner: result.winner.surface,
+    candidateBundle,
+  })
+  const benchmark = candidateBundle.lineage.benchmark
+  if (!benchmark) {
+    throw new Error('agent improvement proposal requires a benchmark and heldout split identity')
   }
-  const review = input as unknown as AgentImprovementReview
-  if (!isSha256Digest(review.digest) || !isSha256Digest(review.proposalDigest)) {
-    throw new Error('candidate review digest is invalid')
+  const power = result.power
+  if (!power) throw new Error('agent improvement proposal requires heldout power analysis')
+  if (result.provenance.gate.reasons.length === 0) {
+    throw new Error('agent improvement proposal requires measured decision reasons')
   }
-  if (review.candidateBundleDigest !== undefined && !isSha256Digest(review.candidateBundleDigest)) {
-    throw new Error('candidate review bundle digest is invalid')
+  const pairs = pairMeasuredCells(
+    result.raw.baselineOnHoldout.cells,
+    result.raw.winnerOnHoldout.cells,
+  )
+  const composite = measuredObjective(
+    {
+      kind: 'objective',
+      name: 'composite',
+      direction: 'higher-is-better',
+      unit: 'score',
+    },
+    pairs,
+    measuredComposite,
+  )
+  assertMeasuredNumber(result.lift, composite.delta, 'heldout lift')
+  assertMeasuredNumber(result.baseline.compositeMean, composite.baseline, 'heldout baseline')
+  assertMeasuredNumber(result.winner.compositeMean, composite.candidate, 'heldout candidate')
+  assertMeasuredNumber(result.provenance.heldOutLift, composite.delta, 'provenance heldout lift')
+  if (
+    result.gateDecision !== result.provenance.gate.decision ||
+    power.n !== composite.n ||
+    power.confidence !== 0.95
+  ) {
+    throw new Error('agent improvement measurement sources do not agree')
   }
-  if (!['approve', 'reject', 'request-changes'].includes(review.decision)) {
-    throw new Error('candidate review decision is invalid')
+
+  const comparison = immutableCandidateValue({
+    schemaVersion: 1 as const,
+    kind: 'agent-improvement-measured-comparison' as const,
+    benchmark,
+    baselineProfileDigest: canonicalCandidateDigest(baselineProfile),
+    candidateBundleDigest: candidateBundle.digest,
+    overall: {
+      name: 'composite' as const,
+      baseline: composite.baseline,
+      candidate: composite.candidate,
+      delta: composite.delta,
+      confidenceInterval: composite.confidenceInterval,
+      n: composite.n,
+      direction: 'higher-is-better' as const,
+      unit: 'score' as const,
+    },
+    objectives: measuredObjectives(pairs),
+    ...(result.winner.label || result.winner.rationale
+      ? {
+          candidate: {
+            ...(result.winner.label ? { label: result.winner.label } : {}),
+            ...(result.winner.rationale ? { rationale: result.winner.rationale } : {}),
+          },
+        }
+      : {}),
+    decision: {
+      outcome: result.gateDecision,
+      reasons: result.provenance.gate.reasons,
+      contributingChecks: result.provenance.gate.contributingGates.map((check) => ({
+        name: check.name,
+        passed: check.passed,
+      })),
+    },
+    power: {
+      sufficient: power.scaleAssumed && !power.underpowered,
+      n: power.n,
+      minimumDetectableDelta: power.mde,
+      confidenceLevel: power.confidence,
+      scaleAssumed: power.scaleAssumed,
+      sharedScorerChannel: power.sharedChannelCaveat !== undefined,
+      reason: power.recommendation,
+    },
+    provenance: {
+      kind: 'agent-eval-loop' as const,
+      schema: result.provenance.schema,
+      runId: result.provenance.runId,
+      recordDigest: canonicalCandidateDigest(result.provenance),
+      baselineContentHash: result.provenance.baselineContentHash,
+      candidateContentHash: result.provenance.winnerContentHash,
+    },
+    diff: result.diff,
+    evaluation: {
+      generationsExplored: result.generationsExplored,
+      durationMs: result.durationMs,
+      totalCostUsd: result.totalCostUsd,
+    },
+  })
+  return agentImprovementMeasuredComparisonSchema.parse(comparison)
+}
+
+interface MeasuredEvaluationCell {
+  scenarioId: string
+  rep: number
+  judgeScores: Record<
+    string,
+    { composite: number; dimensions: Record<string, number>; failed?: true }
+  >
+  costUsd: number
+  tokenUsage?: { input: number; output: number }
+  durationMs: number
+  error?: string
+}
+
+function measuredObjectives(
+  pairs: ReadonlyArray<readonly [MeasuredEvaluationCell, MeasuredEvaluationCell]>,
+): AgentImprovementMeasuredComparison['objectives'] {
+  const qualityColumns = new Map<string, MeasuredQualityColumn>()
+  for (const [baseline, candidate] of pairs) {
+    for (const cell of [baseline, candidate]) {
+      for (const [objective, score] of Object.entries(cell.judgeScores)) {
+        if (score.failed) continue
+        qualityColumns.set(`objective:${objective}`, {
+          kind: 'objective',
+          name: objective,
+          direction: 'higher-is-better',
+          unit: 'score',
+        })
+        for (const name of Object.keys(score.dimensions)) {
+          qualityColumns.set(`dimension:${objective}:${name}`, {
+            kind: 'dimension',
+            objective,
+            name,
+            direction: 'higher-is-better',
+            unit: 'score',
+          })
+        }
+      }
+    }
   }
-  const actual = canonicalCandidateDigest(omitTopLevelDigest(review))
-  if (actual !== review.digest) throw new Error('agent improvement review digest does not match')
-  return immutableCandidateValue(review)
+  const objectives: AgentImprovementMeasuredComparison['objectives'] = [
+    ...[...qualityColumns.entries()]
+      .sort(([left], [right]) => left.localeCompare(right))
+      .map(([, column]) =>
+        measuredObjective(column, pairs, (cell) => measuredQuality(cell, column)),
+      ),
+    measuredCostObjective(pairs),
+    measuredObjective(
+      {
+        kind: 'latency',
+        name: 'latency',
+        direction: 'lower-is-better',
+        unit: 'milliseconds',
+      },
+      pairs,
+      (cell) => cell.durationMs,
+    ),
+  ]
+  return objectives
 }
 
-function improvementEvaluation<TScenario extends Scenario, TArtifact>(
-  result: SelfImproveResult<TScenario, TArtifact>,
-): AgentImprovementEvaluation<TScenario, TArtifact> {
-  return {
-    baseline: result.baseline,
-    winner: result.winner,
-    lift: result.lift,
-    diff: result.diff,
-    provenance: result.provenance,
-    gateDecision: result.gateDecision,
-    generationsExplored: result.generationsExplored,
-    durationMs: result.durationMs,
-    totalCostUsd: result.totalCostUsd,
-    insight: result.insight,
-    ...(result.power === undefined ? {} : { power: result.power }),
+function measuredCostObjective(
+  pairs: ReadonlyArray<readonly [MeasuredEvaluationCell, MeasuredEvaluationCell]>,
+): AgentImprovementMeasuredComparison['objectives'][number] {
+  const cells = pairs.flat()
+  for (const cell of cells) finiteMeasuredValue(cell.costUsd, 'cost:cost')
+  const reported = cells.some(
+    (cell) =>
+      cell.costUsd !== 0 ||
+      (cell.tokenUsage !== undefined && (cell.tokenUsage.input > 0 || cell.tokenUsage.output > 0)),
+  )
+  if (!reported) {
+    return {
+      kind: 'cost',
+      name: 'cost',
+      availability: 'unavailable',
+      reason: 'heldout cells did not report model usage or cost',
+      direction: 'lower-is-better',
+      unit: 'usd',
+    }
   }
+  return measuredObjective(
+    {
+      kind: 'cost',
+      name: 'cost',
+      direction: 'lower-is-better',
+      unit: 'usd',
+    },
+    pairs,
+    (cell) => cell.costUsd,
+  )
 }
 
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return value !== null && typeof value === 'object' && !Array.isArray(value)
+function measuredComposite(cell: MeasuredEvaluationCell): number {
+  const values = Object.values(cell.judgeScores)
+    .filter((score) => !score.failed)
+    .map((score) => score.composite)
+    .filter(Number.isFinite)
+  if (values.length === 0) {
+    throw new Error(`heldout cell ${measuredCellKey(cell)} has no successful composite score`)
+  }
+  return measuredMean(values)
 }
 
-function isSha256Digest(value: unknown): value is Sha256Digest {
-  return typeof value === 'string' && /^sha256:[a-f0-9]{64}$/.test(value)
+function pairMeasuredCells(
+  baselineCells: readonly MeasuredEvaluationCell[],
+  candidateCells: readonly MeasuredEvaluationCell[],
+): Array<readonly [MeasuredEvaluationCell, MeasuredEvaluationCell]> {
+  type MeasuredArmRow = PairedArmRow & { cell: MeasuredEvaluationCell }
+  const rows: MeasuredArmRow[] = [
+    ...baselineCells
+      .filter((cell) => !cell.error)
+      .map((cell) => ({
+        pairKey: cell.scenarioId,
+        repKey: String(cell.rep),
+        arm: 'baseline',
+        cell,
+      })),
+    ...candidateCells
+      .filter((cell) => !cell.error)
+      .map((cell) => ({
+        pairKey: cell.scenarioId,
+        repKey: String(cell.rep),
+        arm: 'candidate',
+        cell,
+      })),
+  ]
+  const paired = pairArms(rows, { baselineArm: 'baseline', treatmentArm: 'candidate' })
+  if (
+    paired.pairs.length === 0 ||
+    paired.unpairedBaseline.length > 0 ||
+    paired.unpairedTreatment.length > 0
+  ) {
+    throw new Error('measured objectives require the same non-empty paired heldout cells')
+  }
+  return paired.pairs.map(
+    (pair) =>
+      [(pair.baseline as MeasuredArmRow).cell, (pair.treatment as MeasuredArmRow).cell] as const,
+  )
+}
+
+type MeasuredObjectiveIdentity =
+  | {
+      kind: 'objective'
+      name: string
+      direction: 'higher-is-better'
+      unit: 'score'
+    }
+  | {
+      kind: 'dimension'
+      objective: string
+      name: string
+      direction: 'higher-is-better'
+      unit: 'score'
+    }
+  | {
+      kind: 'cost'
+      name: 'cost'
+      direction: 'lower-is-better'
+      unit: 'usd'
+    }
+  | {
+      kind: 'latency'
+      name: 'latency'
+      direction: 'lower-is-better'
+      unit: 'milliseconds'
+    }
+
+type MeasuredQualityColumn = Extract<MeasuredObjectiveIdentity, { kind: 'objective' | 'dimension' }>
+
+function measuredObjective(
+  identity: MeasuredObjectiveIdentity,
+  pairs: ReadonlyArray<readonly [MeasuredEvaluationCell, MeasuredEvaluationCell]>,
+  value: (cell: MeasuredEvaluationCell) => number,
+): Extract<AgentImprovementMeasuredComparison['objectives'][number], { availability: 'measured' }> {
+  const key = measuredObjectiveKey(identity)
+  const baseline = pairs.map(([cell]) => finiteMeasuredValue(value(cell), key))
+  const candidate = pairs.map(([, cell]) => finiteMeasuredValue(value(cell), key))
+  const interval = pairedBootstrap(baseline, candidate, {
+    confidence: 0.95,
+    resamples: 2_000,
+    statistic: 'mean',
+    seed: measuredSeed(key),
+  })
+  const baselineMean = measuredMean(baseline)
+  const candidateMean = measuredMean(candidate)
+  const estimate = {
+    availability: 'measured',
+    baseline: baselineMean,
+    candidate: candidateMean,
+    delta: candidateMean - baselineMean,
+    confidenceInterval: {
+      level: interval.confidence,
+      lower: interval.low,
+      upper: interval.high,
+      method: 'paired-bootstrap',
+      statistic: 'mean',
+      resamples: interval.resamples,
+    },
+    n: interval.n,
+  } as const
+  return { ...identity, ...estimate }
 }
 
-function canonicalJsonValue<T>(value: T, path = '$'): T {
-  if (value === null || typeof value === 'string' || typeof value === 'boolean') return value
-  if (typeof value === 'number') {
-    if (!Number.isFinite(value)) throw new Error(`non-finite number at ${path}`)
-    return value
+function measuredQuality(cell: MeasuredEvaluationCell, column: MeasuredQualityColumn): number {
+  const objective = column.kind === 'objective' ? column.name : column.objective
+  const score = cell.judgeScores[objective]
+  if (!score || score.failed) {
+    throw new Error(
+      `heldout cell ${measuredCellKey(cell)} is missing measured objective '${objective}'`,
+    )
   }
-  if (Array.isArray(value)) {
-    return value.map((entry, index) => {
-      if (entry === undefined) throw new Error(`undefined array entry at ${path}[${index}]`)
-      return canonicalJsonValue(entry, `${path}[${index}]`)
-    }) as T
+  const value = column.kind === 'objective' ? score.composite : score.dimensions[column.name]
+  if (value === undefined) {
+    throw new Error(
+      `heldout cell ${measuredCellKey(cell)} is missing '${objective}' dimension '${column.name}'`,
+    )
   }
-  if (typeof value === 'object') {
-    const entries = Object.entries(value as Record<string, unknown>)
-      .filter(([, entry]) => entry !== undefined)
-      .map(([key, entry]) => [key, canonicalJsonValue(entry, `${path}.${key}`)])
-    return Object.fromEntries(entries) as T
+  return finiteMeasuredValue(value, measuredObjectiveKey(column))
+}
+
+function measuredObjectiveKey(identity: MeasuredObjectiveIdentity): string {
+  return identity.kind === 'dimension'
+    ? `${identity.kind}:${identity.objective}:${identity.name}`
+    : `${identity.kind}:${identity.name}`
+}
+
+function measuredCellKey(cell: Pick<MeasuredEvaluationCell, 'scenarioId' | 'rep'>): string {
+  return `${cell.scenarioId}:${cell.rep}`
+}
+
+function finiteMeasuredValue(value: number, name: string): number {
+  if (!Number.isFinite(value)) throw new Error(`measured objective '${name}' is not finite`)
+  return value
+}
+
+function measuredMean(values: readonly number[]): number {
+  if (values.length === 0) throw new Error('measured objective has no paired values')
+  return values.reduce((sum, value) => sum + value, 0) / values.length
+}
+
+function measuredSeed(value: string): number {
+  let seed = 0x811c9dc5
+  for (const byte of Buffer.from(value, 'utf8')) {
+    seed = Math.imul(seed ^ byte, 0x01000193) >>> 0
+  }
+  return seed
+}
+
+function assertMeasuredNumber(actual: number, expected: number, name: string): void {
+  const tolerance = Number.EPSILON * Math.max(1, Math.abs(actual), Math.abs(expected)) * 8
+  if (
+    !Number.isFinite(actual) ||
+    !Number.isFinite(expected) ||
+    Math.abs(actual - expected) > tolerance
+  ) {
+    throw new Error(`${name} does not agree across the measured comparison`)
   }
-  throw new Error(`unsupported proposal value at ${path}`)
 }
diff --git a/src/intelligence/index.ts b/src/intelligence/index.ts
index f81bf8f2..9d4122f8 100644
--- a/src/intelligence/index.ts
+++ b/src/intelligence/index.ts
@@ -19,7 +19,7 @@
  */
 
 import { contentHash } from '@tangle-network/agent-eval'
-import type { AgentProfile } from '@tangle-network/agent-interface'
+import type { AgentProfile, CandidateExecutionEvidence } from '@tangle-network/agent-interface'
 import {
   buildLoopOtelSpans,
   buildRuntimeEventOtelSpans,
@@ -39,9 +39,17 @@ import {
   isIntelligenceOff,
   resolveEffort,
 } from './effort'
-import type { CandidateExecutionEvidence } from './improvement-cycle'
 import { type Redactor, resolveRedactor } from './redact'
 
+export type {
+  AgentCandidateProfileActivation,
+  AgentImprovementMeasuredComparison,
+  AgentImprovementProposal,
+  AgentImprovementReview,
+  AgentImprovementReviewDecision,
+  CandidateExecutionEvidence,
+} from '@tangle-network/agent-interface'
+export { parseAgentCandidateProfileActivation } from '../candidate-execution/profile'
 export type {
   CapabilityAuth,
   CapabilityInterface,
@@ -94,25 +102,23 @@ export {
   resolveEffort,
 } from './effort'
 export type {
-  AgentImprovementEvaluation,
-  AgentImprovementProposal,
-  AgentImprovementReview,
-  AgentImprovementReviewDecision,
-  CandidateExecutionEvidence,
   CreateAgentImprovementProposalOptions,
   ExecuteApprovedAgentCandidateOptions,
   ExecuteApprovedAgentCandidateResult,
   ProposeAgentImprovementOptions,
   ProposeAgentImprovementResult,
   ReviewAgentImprovementInput,
+  VerifyCandidateExecutionEvidenceOptions,
 } from './improvement-cycle'
 export {
+  createAgentImprovementMeasuredComparison,
   createAgentImprovementProposal,
   executeApprovedAgentCandidate,
   proposeAgentImprovement,
   reviewAgentImprovementProposal,
   verifyAgentImprovementProposal,
   verifyAgentImprovementReview,
+  verifyCandidateExecutionEvidence,
 } from './improvement-cycle'
 export type { Redactor } from './redact'
 export { defaultRedactor, resolveRedactor } from './redact'
@@ -121,6 +127,15 @@ export {
   composeCertifiedProfile,
   composeCertifiedProfileFromWire,
 } from './resolver'
+export type {
+  CreateSandboxApprovedCandidateExecutorOptions,
+  SandboxApprovedCandidateExecution,
+  SandboxApprovedCandidateExecutor,
+} from './sandbox-approved-candidate'
+export {
+  createSandboxApprovedCandidateExecutor,
+  sandboxApprovedCandidateExecutionSupport,
+} from './sandbox-approved-candidate'
 export type {
   AppliedIntelligence,
   IntelligenceAgent,
@@ -586,19 +601,14 @@ export function createIntelligenceClient(config: IntelligenceConfig): Intelligen
           ? {
               'tangle.candidate.proposal_digest': record.candidateExecution.proposalDigest,
               'tangle.candidate.review_digest': record.candidateExecution.reviewDigest,
-              'tangle.candidate.bundle_digest': record.candidateExecution.bundleDigest,
+              'tangle.candidate.bundle_digest': record.candidateExecution.receipt.bundleDigest,
               'tangle.candidate.execution_id': record.candidateExecution.executionId,
               'tangle.candidate.execution_plan_digest':
-                record.candidateExecution.executionPlanDigest,
+                record.candidateExecution.receipt.executionPlanDigest,
               'tangle.candidate.materialization_receipt_digest':
-                record.candidateExecution.materializationReceiptDigest,
+                record.candidateExecution.receipt.materializationReceiptDigest,
               'tangle.candidate.succeeded': record.candidateExecution.succeeded,
-              ...(record.candidateExecution.runReceiptDigest
-                ? {
-                    'tangle.candidate.run_receipt_digest':
-                      record.candidateExecution.runReceiptDigest,
-                  }
-                : {}),
+              'tangle.candidate.run_receipt_digest': record.candidateExecution.receipt.digest,
             }
           : {}),
       }
diff --git a/src/intelligence/sandbox-approved-candidate.ts b/src/intelligence/sandbox-approved-candidate.ts
new file mode 100644
index 00000000..97d06916
--- /dev/null
+++ b/src/intelligence/sandbox-approved-candidate.ts
@@ -0,0 +1,516 @@
+import { posix } from 'node:path'
+
+import type { TraceStore } from '@tangle-network/agent-eval'
+import type {
+  AgentCandidateTermination,
+  AgentImprovementProposal,
+  AgentImprovementReview,
+  CandidateExecutionEvidence,
+  Sha256Digest,
+} from '@tangle-network/agent-interface'
+import type {
+  CreateSandboxOptions,
+  ListSandboxOptions,
+  Process,
+  ProcessStatus,
+  SandboxClient,
+  SandboxInstance,
+  SandboxResources,
+} from '@tangle-network/sandbox'
+
+import type { AgentCandidateExecutionClaimStore } from '../candidate-execution/claim'
+import { canonicalCandidateBytes } from '../candidate-execution/digest'
+import type { PrepareAgentCandidateExecutionOptions } from '../candidate-execution/prepare'
+import type {
+  AgentCandidateBenchmarkGraderPort,
+  AgentCandidateExecutionPorts,
+  AgentCandidateExecutorPort,
+  AgentCandidateExecutorRequest,
+  AgentCandidateExecutorStopRequest,
+  AgentCandidateOutputArtifactPort,
+} from '../candidate-execution/types'
+import { AGENT_CANDIDATE_EXECUTION_SUPPORT } from '../candidate-execution/verify'
+import { executeApprovedAgentCandidate } from './improvement-cycle'
+
+type SandboxClientPort = Pick<SandboxClient, 'create' | 'get' | 'list'>
+
+type SandboxExecutorOptions = NonNullable<CreateSandboxApprovedCandidateExecutorOptions['sandbox']>
+
+interface SandboxRunState {
+  sandbox: SandboxInstance
+  output?: Uint8Array
+  termination?: AgentCandidateTermination
+}
+
+/** Declares the exact candidate surfaces the sandbox executor can run. */
+export const sandboxApprovedCandidateExecutionSupport = Object.freeze({
+  outcomes: Object.freeze(['output'] as const),
+  outputMediaTypes: Object.freeze(['text/*', 'application/json', '*+json'] as const),
+  code: Object.freeze(['disabled'] as const),
+  memory: Object.freeze(['disabled'] as const),
+  knowledge: false,
+  profile: AGENT_CANDIDATE_EXECUTION_SUPPORT.profile,
+  isolation: Object.freeze({
+    freshSandbox: true,
+    exactProcess: true,
+    egress: Object.freeze(['blocked', 'strict'] as const),
+  }),
+})
+
+export interface CreateSandboxApprovedCandidateExecutorOptions {
+  client: SandboxClientPort
+  ports: AgentCandidateExecutionPorts
+  grader: AgentCandidateBenchmarkGraderPort
+  outputArtifacts: AgentCandidateOutputArtifactPort
+  traceStore: TraceStore
+  claimStore: AgentCandidateExecutionClaimStore
+  authorizeReview: (
+    review: AgentImprovementReview,
+    proposal: AgentImprovementProposal,
+  ) => boolean | Promise<boolean>
+  sandbox?: {
+    teamId?: string
+    resources?: SandboxResources
+    createTimeoutMs?: number
+    evidenceRetentionSeconds?: number
+  }
+  cleanupTimeoutMs?: number
+  resultTimeoutMs?: number
+}
+
+export interface SandboxApprovedCandidateExecution {
+  proposal: AgentImprovementProposal
+  review: AgentImprovementReview
+  task: Parameters<typeof executeApprovedAgentCandidate>[0]['task']
+  preparation?: PrepareAgentCandidateExecutionOptions
+}
+
+export interface SandboxApprovedCandidateExecutor {
+  /** The same port is usable by Runtime's expired-claim recovery path. */
+  readonly executor: AgentCandidateExecutorPort
+  execute(input: SandboxApprovedCandidateExecution): Promise<CandidateExecutionEvidence>
+}
+
+/** Compose approved-candidate execution directly onto fresh Tangle sandboxes. */
+export function createSandboxApprovedCandidateExecutor(
+  options: CreateSandboxApprovedCandidateExecutorOptions,
+): SandboxApprovedCandidateExecutor {
+  const executor = new SandboxAgentCandidateExecutor(options.client, options.sandbox)
+  return Object.freeze({
+    executor,
+    async execute(input: SandboxApprovedCandidateExecution): Promise<CandidateExecutionEvidence> {
+      const result = await executeApprovedAgentCandidate({
+        proposal: input.proposal,
+        review: input.review,
+        authorizeReview: options.authorizeReview,
+        task: input.task,
+        ports: options.ports,
+        ...(input.preparation ? { preparation: input.preparation } : {}),
+        execution: {
+          executor,
+          grader: options.grader,
+          outputArtifacts: options.outputArtifacts,
+          traceStore: options.traceStore,
+          claimStore: options.claimStore,
+          ...(options.cleanupTimeoutMs === undefined
+            ? {}
+            : { cleanupTimeoutMs: options.cleanupTimeoutMs }),
+          ...(options.resultTimeoutMs === undefined
+            ? {}
+            : { resultTimeoutMs: options.resultTimeoutMs }),
+        },
+      })
+      if (!result.finalization.succeeded) {
+        throw new Error(`approved candidate execution failed: ${result.finalization.reason}`)
+      }
+      if (!result.evidence) throw new Error('approved candidate execution returned no evidence')
+      const evidence = result.evidence
+      await executor.delete({
+        executionId: evidence.executionId,
+        executionPlanDigest: evidence.receipt.executionPlanDigest,
+      })
+      return evidence
+    },
+  })
+}
+
+class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
+  private readonly states = new Map<string, SandboxRunState>()
+
+  constructor(
+    private readonly client: SandboxClientPort,
+    private readonly options: SandboxExecutorOptions = {},
+  ) {}
+
+  async execute(
+    request: AgentCandidateExecutorRequest,
+    context: Parameters<AgentCandidateExecutorPort['execute']>[1],
+  ) {
+    assertSupportedRequest(request)
+    const outcome = request.executionPlan.value.material.task.outcome
+    if (outcome.kind !== 'output') throw new Error('sandbox executor requires an output task')
+    const key = executionKey(request.executionId, request.executionPlan.value.digest)
+    const sandbox = await this.client.create(sandboxCreateOptions(request, this.options), {
+      signal: context.signal,
+      timeoutMs: this.options.createTimeoutMs ?? 120_000,
+    })
+    const state: SandboxRunState = { sandbox }
+    this.states.set(key, state)
+    if ((await sandbox.process.list()).length !== 0) {
+      throw new Error('fresh candidate sandbox already contains a process')
+    }
+    await materializeRequest(sandbox, request)
+    const launch = exactLaunch(request)
+    const startedAt = Date.now()
+    const process = await sandbox.process.spawnExact(launch.executable, launch.args, {
+      cwd: request.launch.cwd,
+      env: { ...request.launch.env },
+      ...(launch.stdin === undefined ? {} : { stdin: launch.stdin }),
+      timeoutMs: request.hardLimits.timeoutMs,
+    })
+    const outputPromise = collectOutput(process, outcome.maxBytes)
+    let cancellation: Promise<void> | undefined
+    let cancellationTermination: AgentCandidateTermination | undefined
+    const cancel = () => {
+      cancellationTermination =
+        Date.now() >= context.deadlineAtMs
+          ? { kind: 'timeout', timeoutMs: request.hardLimits.timeoutMs }
+          : { kind: 'cancelled' }
+      cancellation ??= killProcess(process)
+    }
+    context.signal.addEventListener('abort', cancel, { once: true })
+    if (context.signal.aborted) cancel()
+    try {
+      const waitedExitCode = await process.wait()
+      if (cancellation) await cancellation
+      const status = await process.status()
+      if (status.running || status.exitCode !== waitedExitCode) {
+        throw new Error('sandbox process returned an inconsistent terminal status')
+      }
+      state.output = await outputPromise
+      state.termination = cancellationTermination ?? processTermination(status)
+      await context.traceStore.appendRun({
+        runId: request.trace.runId,
+        scenarioId: request.executionPlan.value.material.task.taskId,
+        startedAt,
+        endedAt: Date.now(),
+        status: status.exitCode === 0 ? 'completed' : 'failed',
+        tags: { ...request.trace.tags, sandboxId: sandbox.id },
+      })
+      if (context.signal.aborted) {
+        throw context.signal.reason ?? new Error('candidate execution was cancelled')
+      }
+      return { executionId: request.executionId, termination: state.termination }
+    } finally {
+      context.signal.removeEventListener('abort', cancel)
+      void outputPromise.catch(() => undefined)
+    }
+  }
+
+  async stop(request: AgentCandidateExecutorStopRequest): Promise<{ readonly stopped: true }> {
+    const sandbox = await this.resolve(request, true)
+    if (!sandbox) return { stopped: true }
+    const statuses = await sandbox.process.list()
+    if (statuses.length > 1) throw new Error('candidate sandbox contains more than one process')
+    const status = statuses[0]
+    if (status?.running) {
+      const process = await sandbox.process.get(status.pid)
+      if (!process) throw new Error('candidate sandbox lost its running process handle')
+      await killProcess(process)
+    }
+    const remaining = await sandbox.process.list()
+    if (remaining.some((entry) => entry.running)) {
+      throw new Error('candidate sandbox process termination is not proven')
+    }
+    return { stopped: true }
+  }
+
+  async capture(request: AgentCandidateExecutorStopRequest) {
+    const sandbox = await this.resolve(request, false)
+    const statuses = await sandbox.process.list()
+    if (statuses.length > 1 || statuses.some((entry) => entry.running)) {
+      throw new Error('candidate sandbox must contain one stopped process before capture')
+    }
+    const state = this.states.get(executionKey(request.executionId, request.executionPlanDigest))
+    let output = state?.output
+    let termination = state?.termination
+    const status = statuses[0]
+    if (status && (!output || !termination)) {
+      const process = await sandbox.process.get(status.pid)
+      if (!process) throw new Error('candidate sandbox lost its captured process handle')
+      const expected = sandboxOutputSpec(sandbox)
+      output = await collectOutput(process, expected.maxBytes)
+      termination = processTermination(await process.status())
+    }
+    const evidence = canonicalCandidateBytes({
+      schemaVersion: 1,
+      kind: 'sandbox-agent-candidate-capture',
+      sandboxId: sandbox.id,
+      executionId: request.executionId,
+      executionPlanDigest: request.executionPlanDigest,
+      ...(status
+        ? {
+            process: {
+              pid: status.pid,
+              exitCode: status.exitCode,
+              ...(status.exitSignal ? { exitSignal: status.exitSignal } : {}),
+            },
+          }
+        : {}),
+      ...(termination ? { termination } : {}),
+    })
+    return {
+      ...(output ? { taskOutcome: { kind: 'output' as const, bytes: output } } : {}),
+      evidence,
+    }
+  }
+
+  async delete(request: AgentCandidateExecutorStopRequest): Promise<void> {
+    const sandbox = await this.resolve(request, true)
+    if (!sandbox) return
+    try {
+      await sandbox.delete()
+    } catch (error) {
+      if (await this.client.get(sandbox.id)) throw error
+    }
+    this.states.delete(executionKey(request.executionId, request.executionPlanDigest))
+  }
+
+  private async resolve(
+    request: AgentCandidateExecutorStopRequest,
+    missingIsDeleted: false,
+  ): Promise<SandboxInstance>
+  private async resolve(
+    request: AgentCandidateExecutorStopRequest,
+    missingIsDeleted: true,
+  ): Promise<SandboxInstance | undefined>
+  private async resolve(
+    request: AgentCandidateExecutorStopRequest,
+    missingIsDeleted: boolean,
+  ): Promise<SandboxInstance | undefined> {
+    const key = executionKey(request.executionId, request.executionPlanDigest)
+    const active = this.states.get(key)?.sandbox
+    if (active) return active
+    const matches: SandboxInstance[] = []
+    const scope: ListSandboxOptions['scope'] = this.options.teamId
+      ? `team:${this.options.teamId}`
+      : 'personal'
+    for (let offset = 0; ; offset += 100) {
+      const page = await this.client.list({ scope, limit: 100, offset })
+      for (const candidate of page) {
+        if (matchesExecution(candidate, request)) matches.push(candidate)
+      }
+      if (page.length < 100) break
+    }
+    if (matches.length === 1) return matches[0]
+    if (matches.length > 1) throw new Error('multiple sandboxes match one candidate execution')
+    if (missingIsDeleted) return undefined
+    throw new Error('candidate sandbox evidence is unavailable')
+  }
+}
+
+function assertSupportedRequest(request: AgentCandidateExecutorRequest): void {
+  if (request.executionPlan.value.material.task.outcome.kind !== 'output') {
+    throw new Error('sandbox candidate executor does not support workspace outcomes; use Pier')
+  }
+  if (request.executionPlan.value.material.codeKind !== 'disabled' || request.inputs.candidate) {
+    throw new Error('sandbox candidate executor does not support code workspaces; use Pier')
+  }
+  if (request.memory.mode !== 'disabled') {
+    throw new Error('sandbox candidate executor does not support isolated memory')
+  }
+  const mediaType = request.executionPlan.value.material.task.outcome.mediaType.toLowerCase()
+  if (
+    !(
+      mediaType.startsWith('text/') ||
+      mediaType === 'application/json' ||
+      mediaType.endsWith('+json')
+    )
+  ) {
+    throw new Error('sandbox candidate executor supports only UTF-8 text and JSON outputs')
+  }
+}
+
+function sandboxCreateOptions(
+  request: AgentCandidateExecutorRequest,
+  options: SandboxExecutorOptions = {},
+): CreateSandboxOptions {
+  const material = request.executionPlan.value.material
+  const retentionSeconds = options.evidenceRetentionSeconds ?? 900
+  if (!Number.isSafeInteger(retentionSeconds) || retentionSeconds < 60) {
+    throw new Error('sandbox evidence retention must be at least 60 seconds')
+  }
+  const maxLifetimeSeconds = Math.ceil(request.hardLimits.timeoutMs / 1_000) + retentionSeconds
+  const network = material.model.access.network
+  const egressPolicy =
+    network.mode === 'disabled'
+      ? { mode: 'blocked' as const }
+      : {
+          mode: 'strict' as const,
+          allowDomains: [...network.domains],
+          includeImplicitDomains: false,
+        }
+  const value = {
+    image: exactImage(material.container.image, material.container.manifestDigest),
+    bare: true,
+    publicEdge: false,
+    ephemeral: true,
+    sshEnabled: false,
+    webTerminalEnabled: false,
+    secrets: [],
+    capabilities: [],
+    egressPolicy,
+    maxLifetimeSeconds,
+    idempotencyKey: `candidate-${request.executionPlan.value.digest.slice('sha256:'.length)}`,
+    metadata: {
+      kind: 'agent-candidate-execution',
+      executionId: request.executionId,
+      executionPlanDigest: request.executionPlan.value.digest,
+      outputMediaType:
+        material.task.outcome.kind === 'output' ? material.task.outcome.mediaType : '',
+      outputMaxBytes: material.task.outcome.kind === 'output' ? material.task.outcome.maxBytes : 0,
+    },
+    ...(options.teamId ? { teamId: options.teamId } : {}),
+    ...(options.resources ? { resources: options.resources } : {}),
+  }
+  return value
+}
+
+async function materializeRequest(
+  sandbox: SandboxInstance,
+  request: AgentCandidateExecutorRequest,
+): Promise<void> {
+  for (const file of request.inputs.task.files) {
+    await writeExactFile(sandbox, beneath(request.roots.taskRoot, file.path), file.bytes, file.mode)
+  }
+  const profileRoot =
+    request.executionPlan.value.material.profile.targetWorkspace === 'task'
+      ? request.roots.taskRoot
+      : request.roots.candidateRoot
+  if (!profileRoot) throw new Error('candidate profile targets a missing workspace')
+  for (const file of request.profileActivation.files) {
+    await writeExactFile(
+      sandbox,
+      beneath(profileRoot, file.path),
+      Buffer.from(file.content, 'utf8'),
+      file.mode,
+    )
+  }
+  if (request.instruction.delivery.kind === 'utf8-file') {
+    await writeExactFile(
+      sandbox,
+      request.instruction.delivery.path,
+      request.instruction.bytes,
+      0o644,
+    )
+  }
+}
+
+function exactLaunch(request: AgentCandidateExecutorRequest): {
+  executable: string
+  args: readonly string[]
+  stdin?: string
+} {
+  const instruction = new TextDecoder('utf-8', { fatal: true }).decode(request.instruction.bytes)
+  switch (request.instruction.delivery.kind) {
+    case 'argv-append':
+      return { executable: request.launch.executable, args: [...request.launch.args, instruction] }
+    case 'stdin-utf8':
+      return {
+        executable: request.launch.executable,
+        args: request.launch.args,
+        stdin: instruction,
+      }
+    case 'utf8-file':
+      return { executable: request.launch.executable, args: request.launch.args }
+  }
+}
+
+async function writeExactFile(
+  sandbox: SandboxInstance,
+  path: string,
+  bytes: Uint8Array,
+  mode: number,
+): Promise<void> {
+  await sandbox.fs.write(path, Buffer.from(bytes).toString('base64'), {
+    encoding: 'base64',
+    mode,
+  })
+}
+
+async function collectOutput(process: Process, maxBytes: number): Promise<Uint8Array> {
+  const chunks: Buffer[] = []
+  let byteLength = 0
+  for await (const chunk of process.stdout()) {
+    const bytes = Buffer.from(chunk, 'utf8')
+    byteLength += bytes.byteLength
+    if (byteLength > maxBytes) {
+      await killProcess(process)
+      throw new Error(`sandbox candidate output exceeds its ${maxBytes}-byte maximum`)
+    }
+    chunks.push(bytes)
+  }
+  return Uint8Array.from(Buffer.concat(chunks, byteLength))
+}
+
+async function killProcess(process: Process): Promise<void> {
+  const before = await process.status()
+  if (!before.running) return
+  try {
+    await process.kill('SIGKILL', { tree: true })
+  } catch (error) {
+    if ((await process.status()).running) throw error
+  }
+}
+
+function processTermination(status: ProcessStatus): AgentCandidateTermination {
+  if (status.running) throw new Error('candidate process is still running')
+  if (status.exitSignal) {
+    if (!/^SIG[A-Z0-9]+$/.test(status.exitSignal)) {
+      throw new Error('sandbox returned an invalid process signal')
+    }
+    return { kind: 'signal', signal: status.exitSignal }
+  }
+  return { kind: 'exit', exitCode: status.exitCode }
+}
+
+function matchesExecution(
+  sandbox: { metadata?: Record<string, unknown> },
+  request: AgentCandidateExecutorStopRequest,
+): boolean {
+  return (
+    sandbox.metadata?.kind === 'agent-candidate-execution' &&
+    sandbox.metadata.executionId === request.executionId &&
+    sandbox.metadata.executionPlanDigest === request.executionPlanDigest
+  )
+}
+
+function sandboxOutputSpec(sandbox: { metadata?: Record<string, unknown> }): { maxBytes: number } {
+  const maxBytes = sandbox.metadata?.outputMaxBytes
+  if (!Number.isSafeInteger(maxBytes) || Number(maxBytes) < 1) {
+    throw new Error('candidate sandbox output bound is unavailable')
+  }
+  return { maxBytes: Number(maxBytes) }
+}
+
+function executionKey(executionId: string, executionPlanDigest: Sha256Digest): string {
+  return `${executionId}\0${executionPlanDigest}`
+}
+
+function exactImage(image: string, manifestDigest: Sha256Digest): string {
+  const marker = image.lastIndexOf('@')
+  if (marker < 0) return `${image}@${manifestDigest}`
+  if (image.slice(marker + 1) !== manifestDigest) {
+    throw new Error('candidate container image conflicts with its resolved manifest')
+  }
+  return image
+}
+
+function beneath(root: string, relativePath: string): string {
+  if (posix.isAbsolute(relativePath)) throw new Error('candidate input path must be relative')
+  const path = posix.normalize(relativePath)
+  if (path === '..' || path.startsWith('../')) {
+    throw new Error('candidate input path escapes its execution root')
+  }
+  return posix.join(root, path)
+}
diff --git a/src/intelligence/with-intelligence.test.ts b/src/intelligence/with-intelligence.test.ts
index e38e4478..2a9f89f7 100644
--- a/src/intelligence/with-intelligence.test.ts
+++ b/src/intelligence/with-intelligence.test.ts
@@ -1,4 +1,8 @@
-import type { AgentProfile, AgentProfileDiff } from '@tangle-network/agent-interface'
+import type {
+  AgentProfile,
+  AgentProfileDiff,
+  CandidateExecutionEvidence,
+} from '@tangle-network/agent-interface'
 import { applyAgentProfileDiff } from '@tangle-network/agent-interface'
 import { afterEach, describe, expect, it, vi } from 'vitest'
 import type { ProposedProfileDiff } from './delivery'
@@ -243,14 +247,21 @@ describe('withIntelligence — SEND (a typed RunRecord to /v1/otlp)', () => {
               },
             ],
             candidateExecution: {
+              schemaVersion: 1,
+              kind: 'agent-candidate-execution-evidence',
               proposalDigest: `sha256:${'1'.repeat(64)}`,
               reviewDigest: `sha256:${'2'.repeat(64)}`,
-              bundleDigest: `sha256:${'3'.repeat(64)}`,
               executionId: 'candidate-execution-1',
-              executionPlanDigest: `sha256:${'4'.repeat(64)}`,
-              materializationReceiptDigest: `sha256:${'5'.repeat(64)}`,
               succeeded: true,
-              runReceiptDigest: `sha256:${'6'.repeat(64)}`,
+              materializationReceipt: {} as CandidateExecutionEvidence['materializationReceipt'],
+              profileActivation: {} as CandidateExecutionEvidence['profileActivation'],
+              receipt: {
+                bundleDigest: `sha256:${'3'.repeat(64)}`,
+                executionPlanDigest: `sha256:${'4'.repeat(64)}`,
+                materializationReceiptDigest: `sha256:${'5'.repeat(64)}`,
+                digest: `sha256:${'6'.repeat(64)}`,
+              } as CandidateExecutionEvidence['receipt'],
+              digest: `sha256:${'7'.repeat(64)}`,
             },
           })
           return 'answer'
diff --git a/src/knowledge/improvement-job.ts b/src/knowledge/improvement-job.ts
index c1f7e8da..660efbe8 100644
--- a/src/knowledge/improvement-job.ts
+++ b/src/knowledge/improvement-job.ts
@@ -1,12 +1,33 @@
+import { realpath } from 'node:fs/promises'
+import {
+  type AgentCandidateCapturedArtifact,
+  type AgentCandidateKnowledge,
+  type AgentImprovementProposal,
+  type AgentImprovementReview,
+  agentCandidateKnowledgeSchema,
+} from '@tangle-network/agent-interface'
 import {
   type BuildEvalKnowledgeBundleOptions,
   evaluateKnowledgeBaseReadiness,
+  hashKnowledgeBase,
   improveKnowledgeBase,
   type KnowledgeBaseQualityOptions,
+  type KnowledgeImprovementCandidateRef,
   type KnowledgeImprovementOptions,
   type KnowledgeImprovementResult,
   type KnowledgeReadinessSpec,
+  knowledgeImprovementCandidateRef,
+  promoteKnowledgeCandidate,
+  withKnowledgeImprovementCandidate,
 } from '@tangle-network/agent-knowledge'
+import { canonicalCandidateBytes, embeddedCandidateArtifact } from '../candidate-execution/digest'
+import { persistCandidateOutputArtifact } from '../candidate-execution/output-artifacts'
+import type { AgentCandidateOutputArtifactPort } from '../candidate-execution/types'
+import { captureAgentCandidateWorkspace } from '../candidate-execution/workspace-archive'
+import {
+  verifyAgentImprovementProposal,
+  verifyAgentImprovementReview,
+} from '../intelligence/improvement-cycle'
 import type { ExecutorConfig } from '../runtime/supervise/runtime'
 import type { SuperviseOptions } from '../runtime/supervise/supervise'
 import type { SupervisorProfile } from '../runtime/supervise/supervisor-agent'
@@ -39,9 +60,20 @@ export interface RunKnowledgeImprovementJobOptions
     task: unknown,
     opts: SuperviseOptions,
   ) => Promise<SupervisedResult<unknown>>
+  candidateArtifacts?: AgentCandidateOutputArtifactPort
+  approval?: ApprovedKnowledgeImprovementCandidate
   onMeasurement?: (measurement: KnowledgeImprovementJobMeasurement) => Promise<void> | void
 }
 
+export interface ApprovedKnowledgeImprovementCandidate {
+  proposal: AgentImprovementProposal
+  review: AgentImprovementReview
+  authorizeReview: (
+    review: AgentImprovementReview,
+    proposal: AgentImprovementProposal,
+  ) => boolean | Promise<boolean>
+}
+
 export interface KnowledgeImprovementJobMeasurement {
   startedAt: string
   finishedAt: string
@@ -59,6 +91,7 @@ export interface KnowledgeImprovementJobMeasurement {
 
 export interface KnowledgeImprovementJobResult {
   improvement: KnowledgeImprovementResult
+  candidateKnowledge?: AgentCandidateKnowledge
   measurement: KnowledgeImprovementJobMeasurement
   promoted: boolean
   blocked: boolean
@@ -104,7 +137,7 @@ export function createAgentKnowledgeReadinessCheck(
   }
 }
 
-/** Run the full KB improvement job: candidate workspace, runtime supervisor update, readiness check, and promotion. */
+/** Produce a frozen KB candidate, and promote it only when an exact signed review is supplied. */
 export async function runKnowledgeImprovementJob(
   options: RunKnowledgeImprovementJobOptions,
 ): Promise<KnowledgeImprovementJobResult> {
@@ -112,9 +145,11 @@ export async function runKnowledgeImprovementJob(
     allowedModels,
     backend,
     budget,
+    candidateArtifacts,
     harness,
     makeWorkerAgent,
     onMeasurement,
+    approval,
     readinessCheck,
     runSupervised,
     supervisorModel,
@@ -147,17 +182,57 @@ export async function runKnowledgeImprovementJob(
     runSupervised,
   } satisfies SupervisedKnowledgeUpdateOptions)
 
-  const resolvedImprovement = await improveKnowledgeBase({
-    ...knowledgeOptions,
-    updateKnowledge: async (input) => {
-      const updateStartedAt = Date.now()
-      updateCalls += 1
-      const result = await updateKnowledge(input)
-      updateDurationMs += Date.now() - updateStartedAt
-      addSpent(supervisedSpent, result.supervised)
-      return result
-    },
-  } as KnowledgeImprovementOptions)
+  const instrumentedUpdateKnowledge: KnowledgeImprovementOptions['updateKnowledge'] = async (
+    input,
+  ) => {
+    const updateStartedAt = Date.now()
+    updateCalls += 1
+    const result = await updateKnowledge(input)
+    updateDurationMs += Date.now() - updateStartedAt
+    addSpent(supervisedSpent, result.supervised)
+    return result
+  }
+  let resolvedImprovement: KnowledgeImprovementResult
+  let candidateKnowledge: AgentCandidateKnowledge | undefined
+  if (approval) {
+    const approvedKnowledge = await approvedKnowledgeCandidate(approval)
+    const candidate = agentKnowledgeCandidateRef(approvedKnowledge)
+    knowledgeOptions.signal?.throwIfAborted()
+    await withKnowledgeImprovementCandidate({ root: options.root, candidate }, () => undefined)
+    resolvedImprovement = await promoteKnowledgeCandidate({
+      root: options.root,
+      candidate,
+      ...(knowledgeOptions.ownerId ? { ownerId: knowledgeOptions.ownerId } : {}),
+      ...(knowledgeOptions.leaseTtlMs ? { leaseTtlMs: knowledgeOptions.leaseTtlMs } : {}),
+      ...(knowledgeOptions.now ? { now: knowledgeOptions.now } : {}),
+      ...(knowledgeOptions.onState ? { onState: knowledgeOptions.onState } : {}),
+    })
+    knowledgeOptions.signal?.throwIfAborted()
+    const promotedHash = knowledgeDigest(
+      await hashKnowledgeBase(options.root),
+      'promoted knowledge base',
+    )
+    if (
+      !resolvedImprovement.promoted ||
+      promotedHash !== approvedKnowledge.candidate.candidateHash
+    ) {
+      throw new Error('knowledge promotion did not activate the approved snapshot bytes')
+    }
+    candidateKnowledge = approvedKnowledge
+  } else {
+    resolvedImprovement = await improveKnowledgeBase({
+      ...knowledgeOptions,
+      updateKnowledge: instrumentedUpdateKnowledge,
+    } as KnowledgeImprovementOptions)
+    if (resolvedImprovement.candidate) {
+      candidateKnowledge = await freezeKnowledgeCandidate(
+        options.root,
+        resolvedImprovement,
+        candidateArtifacts,
+        knowledgeOptions.signal,
+      )
+    }
+  }
   const finishedAtMs = Date.now()
   const measurement: KnowledgeImprovementJobMeasurement = {
     startedAt,
@@ -170,12 +245,146 @@ export async function runKnowledgeImprovementJob(
   await onMeasurement?.(measurement)
   return {
     improvement: resolvedImprovement,
+    ...(candidateKnowledge ? { candidateKnowledge } : {}),
     measurement,
     promoted: resolvedImprovement.promoted,
     blocked: resolvedImprovement.blocked,
   }
 }
 
+async function freezeKnowledgeCandidate(
+  root: string,
+  improvement: KnowledgeImprovementResult,
+  artifacts: AgentCandidateOutputArtifactPort | undefined,
+  signal: AbortSignal | undefined,
+): Promise<AgentCandidateKnowledge> {
+  const candidate = knowledgeImprovementCandidateRef(improvement)
+  return withKnowledgeImprovementCandidate({ root, candidate }, async (resolved) => {
+    const executionId = `knowledge-${candidate.candidateId}`
+    const captured = await captureAgentCandidateWorkspace(await realpath(resolved.root), {
+      ...(artifacts
+        ? {
+            artifactPersistence: {
+              executionId,
+              outputArtifacts: artifacts,
+              ...(signal ? { signal } : {}),
+            },
+          }
+        : {}),
+    })
+    const evaluation = await captureKnowledgeEvidence(
+      canonicalCandidateBytes({
+        schemaVersion: 1,
+        kind: 'agent-knowledge-candidate-evaluation',
+        candidate: interfaceKnowledgeCandidateRef(candidate),
+        metric: resolved.evaluation,
+      }),
+      'knowledge-evaluation',
+      executionId,
+      artifacts,
+      signal,
+    )
+    return agentCandidateKnowledgeSchema.parse({
+      candidate: interfaceKnowledgeCandidateRef(candidate),
+      snapshot: captured.snapshot,
+      evaluation,
+    })
+  })
+}
+
+async function captureKnowledgeEvidence(
+  bytes: Uint8Array,
+  purpose: 'knowledge-retrieval-config' | 'knowledge-evaluation',
+  executionId: string,
+  artifacts: AgentCandidateOutputArtifactPort | undefined,
+  signal: AbortSignal | undefined,
+): Promise<AgentCandidateCapturedArtifact> {
+  if (!artifacts) return embeddedCandidateArtifact(bytes)
+  return persistCandidateOutputArtifact(artifacts, {
+    executionId,
+    purpose,
+    bytes,
+    ...(signal ? { signal } : {}),
+  })
+}
+
+async function approvedKnowledgeCandidate(
+  approval: ApprovedKnowledgeImprovementCandidate,
+): Promise<AgentCandidateKnowledge> {
+  const proposal = verifyAgentImprovementProposal(approval.proposal)
+  const review = verifyAgentImprovementReview(approval.review)
+  if (
+    proposal.changedSurfaces.length !== 1 ||
+    proposal.changedSurfaces[0] !== 'knowledge' ||
+    !proposal.candidateBundle.knowledge ||
+    review.decision !== 'approve' ||
+    review.proposalDigest !== proposal.digest ||
+    review.candidateBundleDigest !== proposal.candidateBundle.digest
+  ) {
+    throw new Error('knowledge promotion requires an approved exact knowledge proposal')
+  }
+  if (!(await approval.authorizeReview(review, proposal))) {
+    throw new Error('knowledge candidate approval was not authorized')
+  }
+  return proposal.candidateBundle.knowledge
+}
+
+function interfaceKnowledgeCandidateRef(
+  candidate: KnowledgeImprovementCandidateRef,
+): AgentCandidateKnowledge['candidate'] {
+  return {
+    schemaVersion: 1,
+    kind: 'knowledge-improvement-candidate',
+    runId: candidate.runId,
+    candidateId: candidate.candidateId,
+    goalHash: knowledgeDigest(candidate.goalHash, 'knowledge candidate goal'),
+    baseHash: knowledgeDigest(candidate.baseHash, 'knowledge candidate base'),
+    candidateHash: knowledgeDigest(candidate.candidateHash, 'knowledge candidate snapshot'),
+    evidenceHash: knowledgeDigest(candidate.evidenceHash, 'knowledge candidate evidence'),
+    promotionPlanHash: knowledgeDigest(
+      candidate.promotionPlanHash,
+      'knowledge candidate promotion plan',
+    ),
+  }
+}
+
+function agentKnowledgeCandidateRef(
+  knowledge: AgentCandidateKnowledge,
+): KnowledgeImprovementCandidateRef {
+  return {
+    schemaVersion: 1,
+    kind: 'knowledge-improvement-candidate',
+    runId: knowledge.candidate.runId,
+    candidateId: knowledge.candidate.candidateId,
+    goalHash: rawKnowledgeDigest(knowledge.candidate.goalHash, 'knowledge candidate goal'),
+    baseHash: rawKnowledgeDigest(knowledge.candidate.baseHash, 'knowledge candidate base'),
+    candidateHash: rawKnowledgeDigest(
+      knowledge.candidate.candidateHash,
+      'knowledge candidate snapshot',
+    ),
+    evidenceHash: rawKnowledgeDigest(
+      knowledge.candidate.evidenceHash,
+      'knowledge candidate evidence',
+    ),
+    promotionPlanHash: rawKnowledgeDigest(
+      knowledge.candidate.promotionPlanHash,
+      'knowledge candidate promotion plan',
+    ),
+  }
+}
+
+function knowledgeDigest(value: string, label: string): `sha256:${string}` {
+  const digest = value.startsWith('sha256:') ? value : `sha256:${value}`
+  if (!/^sha256:[a-f0-9]{64}$/.test(digest)) {
+    throw new Error(`${label} is not a SHA-256 digest`)
+  }
+  return digest as `sha256:${string}`
+}
+
+function rawKnowledgeDigest(value: string, label: string): string {
+  return knowledgeDigest(value, label).slice('sha256:'.length)
+}
+
 function emptySpent(): KnowledgeImprovementJobMeasurement['supervisedSpent'] {
   return { iterations: 0, inputTokens: 0, outputTokens: 0, usd: 0, ms: 0 }
 }
diff --git a/src/knowledge/index.ts b/src/knowledge/index.ts
index 0297ca41..d9920117 100644
--- a/src/knowledge/index.ts
+++ b/src/knowledge/index.ts
@@ -1,5 +1,6 @@
 export {
   type AgentKnowledgeReadinessCheckOptions,
+  type ApprovedKnowledgeImprovementCandidate,
   createAgentKnowledgeReadinessCheck,
   type KnowledgeImprovementJobMeasurement,
   type KnowledgeImprovementJobResult,
diff --git a/src/knowledge/supervised-update.ts b/src/knowledge/supervised-update.ts
index 9e7149e7..b1df286b 100644
--- a/src/knowledge/supervised-update.ts
+++ b/src/knowledge/supervised-update.ts
@@ -1,3 +1,4 @@
+import type { RagKnowledgeUpdateResult } from '@tangle-network/agent-knowledge'
 import { researcherProfile } from '../profiles/researcher'
 import type { DeliverableSpec } from '../runtime/supervise/completion-gate'
 import type { ExecutorConfig } from '../runtime/supervise/runtime'
@@ -51,7 +52,7 @@ export interface SupervisedKnowledgeUpdateResult {
   applied: boolean
   summary: string
   supervised: SupervisedResult<unknown>
-  metadata: Record<string, unknown>
+  metadata: NonNullable<RagKnowledgeUpdateResult['metadata']>
 }
 
 export interface SupervisedKnowledgeUpdateOptions {
diff --git a/src/runtime/environment-provider.ts b/src/runtime/environment-provider.ts
index c6a62a4b..adcdcfff 100644
--- a/src/runtime/environment-provider.ts
+++ b/src/runtime/environment-provider.ts
@@ -612,22 +612,25 @@ function sandboxInstanceAsEnvironment(
           },
         }
       : {}),
-    ...(hasCheckpoint(box)
-      ? {
-          async checkpoint(options?: CheckpointRequest): Promise<CheckpointRef> {
-            const result = await box.checkpoint(options as never)
-            return { id: checkpointIdFromResult(result), provider: providerName }
-          },
-        }
-      : {}),
-    ...(hasFork(box)
-      ? {
-          async fork(checkpoint: CheckpointRef, options?: ForkRequest): Promise<AgentEnvironment> {
-            const forked = await box.fork(checkpoint.id, options as never)
-            return sandboxInstanceAsEnvironment(forked, providerName, client)
-          },
-        }
-      : {}),
+    async checkpoint(options?: CheckpointRequest): Promise<CheckpointRef> {
+      const result = await box.snapshot({
+        ...(options?.name ? { tags: [options.name] } : {}),
+      })
+      return {
+        id: result.snapshotId,
+        provider: providerName,
+        ...(options?.metadata ? { metadata: options.metadata } : {}),
+      }
+    },
+    async fork(checkpoint: CheckpointRef, options?: ForkRequest): Promise<AgentEnvironment> {
+      const forked = await client.create({
+        fromSnapshot: checkpoint.id,
+        fromSandboxId: String(box.id),
+        ...(options?.name ? { name: options.name } : {}),
+        ...(options?.metadata ? { metadata: options.metadata } : {}),
+      })
+      return sandboxInstanceAsEnvironment(forked, providerName, client)
+    },
     async placement(): Promise<PlacementInfo> {
       return placementInfoFromLoopPlacement(client.describePlacement?.(box), box)
     },
@@ -1049,15 +1052,6 @@ function placementInfoFromLoopPlacement(
   }
 }
 
-function checkpointIdFromResult(result: unknown): string {
-  const record = result && typeof result === 'object' ? (result as Record<string, unknown>) : {}
-  const id = record.checkpointId ?? record.id
-  if (typeof id !== 'string' || id.length === 0) {
-    throw new ValidationError('sandboxClientAsProvider: checkpoint returned no checkpoint id')
-  }
-  return id
-}
-
 function defaultTangleSandboxCapabilities(): AgentEnvironmentCapabilities {
   return {
     profile: {
@@ -1159,18 +1153,6 @@ function hasExec(box: SandboxInstance): box is SandboxInstance & {
   return typeof (box as { exec?: unknown }).exec === 'function'
 }
 
-function hasCheckpoint(
-  box: SandboxInstance,
-): box is SandboxInstance & { checkpoint(options?: unknown): Promise<unknown> } {
-  return typeof (box as { checkpoint?: unknown }).checkpoint === 'function'
-}
-
-function hasFork(box: SandboxInstance): box is SandboxInstance & {
-  fork(checkpointId: string, options?: unknown): Promise<SandboxInstance>
-} {
-  return typeof (box as { fork?: unknown }).fork === 'function'
-}
-
 interface SandboxSessionLike {
   readonly id: string
   status(): Promise<unknown | null>
diff --git a/tests/candidate-bundle-builder.test.ts b/tests/candidate-bundle-builder.test.ts
index 985bc02c..fd3400a2 100644
--- a/tests/candidate-bundle-builder.test.ts
+++ b/tests/candidate-bundle-builder.test.ts
@@ -34,7 +34,7 @@ afterEach(() => {
 })
 
 describe('public agent candidate bundle builder', () => {
-  it('binds profile diffs, verified CodeSurface bytes, knowledge, and memory into one executable candidate', async () => {
+  it('binds profile diffs, verified CodeSurface bytes, and memory into one executable candidate', async () => {
     const fixture = createCandidateExecutionFixture(true)
     const adapter = gitWorktreeAdapter({
       repoRoot: fixture.task.stagingRoots.taskRoot,
@@ -71,7 +71,6 @@ describe('public agent candidate bundle builder', () => {
       },
     }
     const stored = new Map<Sha256Digest, Uint8Array>()
-    const knowledgeManifest = storeArtifact(stored, 'knowledge/manifest.json', '{"version":1}\n')
     const memorySeed = storeArtifact(stored, 'memory/seed.json', '{"entries":[]}\n')
     const input: BuildAgentCandidateBundleInput = {
       profile: { kind: 'profile-diffs', base, diffs: [diff] },
@@ -81,7 +80,6 @@ describe('public agent candidate bundle builder', () => {
         repository: { kind: 'github', owner: 'owner', repo: 'repo' },
       },
       execution: fixture.bundle.execution,
-      knowledge: { snapshotId: 'knowledge-snapshot-1', manifest: knowledgeManifest },
       memory: { mode: 'isolated', scope: 'task', seed: memorySeed },
       lineage: {
         source: 'compound',
diff --git a/tests/candidate-execution-claim.test.ts b/tests/candidate-execution-claim.test.ts
index 0f139045..f57cf006 100644
--- a/tests/candidate-execution-claim.test.ts
+++ b/tests/candidate-execution-claim.test.ts
@@ -2,7 +2,10 @@ import { spawn } from 'node:child_process'
 import { mkdtemp, readdir, readFile, rm } from 'node:fs/promises'
 import { tmpdir } from 'node:os'
 import { resolve } from 'node:path'
-import type { AgentCandidateArtifactRef } from '@tangle-network/agent-interface'
+import type {
+  AgentCandidateArtifactRef,
+  AgentCandidateFixedSpend,
+} from '@tangle-network/agent-interface'
 import { afterEach, describe, expect, it, vi } from 'vitest'
 
 import {
@@ -12,7 +15,6 @@ import {
   type AgentCandidateExecutionLease,
   type AgentCandidateExecutionRecoveryEvidence,
   type AgentCandidateExecutionTerminalResult,
-  type AgentCandidateExecutionUsage,
   InMemoryAgentCandidateExecutionClaimStore,
 } from '../src/candidate-execution/claim'
 import { FileAgentCandidateExecutionClaimStore } from '../src/candidate-execution/claim-file-store'
@@ -51,7 +53,7 @@ describe('candidate execution claim lifecycle', () => {
     const claimedAtMs = Date.now()
     vi.spyOn(Date, 'now').mockReturnValue(claimedAtMs)
 
-    const requested = candidateExecutionClaim(prepared)
+    const requested = candidateExecutionClaim(prepared, preparationEvidenceFor(prepared))
     const ownerWindowMs = candidateExecutionOwnerWindowMs(
       fixture.task.limits.timeoutMs,
       cleanupTimeoutMs,
@@ -88,7 +90,7 @@ describe('candidate execution claim lifecycle', () => {
     )
     vi.spyOn(Date, 'now').mockReturnValue(reservationExpiresAtMs - ownerWindowMs + 1)
 
-    expect(() => candidateExecutionClaim(prepared)).toThrow(
+    expect(() => candidateExecutionClaim(prepared, preparationEvidenceFor(prepared))).toThrow(
       /full execution and cleanup owner window/,
     )
   })
@@ -519,7 +521,7 @@ describe('candidate execution claim lifecycle', () => {
     ).toMatchObject({ acquired: false, detail: 'retry-lineage-mismatch' })
   })
 
-  it('persists claim7, pending1, terminal3, phase, staged, and full usage across stores', async () => {
+  it('persists claim8, pending2, terminal4, phase, staged, and full usage across stores', async () => {
     const directory = await tempDirectory()
     const store = new FileAgentCandidateExecutionClaimStore({ directory })
     const acquired = await acquire(store, claim())
@@ -539,9 +541,9 @@ describe('candidate execution claim lifecycle', () => {
       staged: terminal,
       terminal,
     })
-    expect(files.find(({ name }) => name.endsWith('.claim.json'))?.text).toContain('"version":7')
-    expect(files.find(({ name }) => name.includes('transition-2'))?.text).toContain('"version":1')
-    expect(files.find(({ name }) => name.endsWith('.terminal.json'))?.text).toContain('"version":3')
+    expect(files.find(({ name }) => name.endsWith('.claim.json'))?.text).toContain('"version":8')
+    expect(files.find(({ name }) => name.includes('transition-2'))?.text).toContain('"version":2')
+    expect(files.find(({ name }) => name.endsWith('.terminal.json'))?.text).toContain('"version":4')
     expect(files.map(({ text }) => text).join('\n')).not.toContain(acquired.lease.token)
   })
 
@@ -642,6 +644,10 @@ function claim(
     retryPolicy: 'none',
     bundleDigest: sha256('a'),
     executionPlanDigest: sha256('b'),
+    preparationEvidence: {
+      executionPlan: artifact('b'),
+      materializationReceipt: artifact('e'),
+    },
     retryLineageDigest: sha256('c'),
     leaseExpiresAtMs: FUTURE_EXPIRY_MS,
     resultTimeoutMs: 60_000,
@@ -656,10 +662,27 @@ function retryClaim(
   return claim({ maxAttempts: 3, retryPolicy: 'pre-model-infrastructure-only', ...overrides })
 }
 
+function preparationEvidenceFor(
+  prepared: Parameters<typeof candidateExecutionClaim>[0],
+): AgentCandidateExecutionClaim['preparationEvidence'] {
+  return {
+    executionPlan: {
+      locator: { kind: 's3', bucket: 'candidate-evidence', key: 'execution-plan.json' },
+      sha256: prepared.executionPlan.value.digest,
+      byteLength: prepared.executionPlan.bytes.byteLength,
+    },
+    materializationReceipt: {
+      locator: { kind: 's3', bucket: 'candidate-evidence', key: 'materialization-receipt.json' },
+      sha256: prepared.materializationReceipt.digest,
+      byteLength: prepared.materializationReceipt.bytes.byteLength,
+    },
+  }
+}
+
 function usage(
   modelCalls = 0,
-  overrides: Partial<AgentCandidateExecutionUsage> = {},
-): AgentCandidateExecutionUsage {
+  overrides: Partial<AgentCandidateFixedSpend> = {},
+): AgentCandidateFixedSpend {
   return {
     costUsdNanos: modelCalls * 123_456,
     inputTokens: modelCalls * 101,
@@ -738,7 +761,7 @@ function recoveryEvidence(
   requested: AgentCandidateExecutionClaim,
   overrides: {
     failureClass?: AgentCandidateExecutionFailureClass
-    usage?: AgentCandidateExecutionUsage
+    usage?: AgentCandidateFixedSpend
     modelSettlement?: AgentCandidateArtifactRef
   } = {},
 ): AgentCandidateExecutionRecoveryEvidence {
diff --git a/tests/candidate-execution-core.test.ts b/tests/candidate-execution-core.test.ts
index 5b3cb361..4619b703 100644
--- a/tests/candidate-execution-core.test.ts
+++ b/tests/candidate-execution-core.test.ts
@@ -13,7 +13,7 @@ import { join } from 'node:path'
 
 import type {
   AgentCandidateGitPatch,
-  AgentCandidateWorkspaceManifestMaterialV1,
+  AgentCandidateWorkspaceManifestMaterial,
 } from '@tangle-network/agent-interface'
 import { afterEach, describe, expect, it } from 'vitest'
 
@@ -101,7 +101,7 @@ describe('candidate canonical bytes and artifacts', () => {
   })
 
   it('requires workspace manifest bytes to equal canonical material, not only a locator', async () => {
-    const material: AgentCandidateWorkspaceManifestMaterialV1 = {
+    const material: AgentCandidateWorkspaceManifestMaterial = {
       schemaVersion: 1,
       kind: 'agent-candidate-workspace-manifest',
       files: [],
@@ -196,7 +196,7 @@ describe('materialized workspace identity', () => {
     writeFileSync(join(root, '.git', 'ignored'), 'Git internals are not uploaded')
     writeFileSync(join(root, 'run.js'), 'ok\n', { mode: 0o755 })
     const bytes = Buffer.from('ok\n')
-    const material: AgentCandidateWorkspaceManifestMaterialV1 = {
+    const material: AgentCandidateWorkspaceManifestMaterial = {
       schemaVersion: 1,
       kind: 'agent-candidate-workspace-manifest',
       files: [
@@ -231,7 +231,7 @@ describe('materialized workspace identity', () => {
     writeFileSync(join(root, 'source'), 'x')
     chmodSync(join(root, 'source'), 0o644)
     symlinkSync('source', join(root, 'symlink'))
-    const material: AgentCandidateWorkspaceManifestMaterialV1 = {
+    const material: AgentCandidateWorkspaceManifestMaterial = {
       schemaVersion: 1,
       kind: 'agent-candidate-workspace-manifest',
       files: [],
diff --git a/tests/candidate-execution-execute.test.ts b/tests/candidate-execution-execute.test.ts
index 82be48bc..a2d330e1 100644
--- a/tests/candidate-execution-execute.test.ts
+++ b/tests/candidate-execution-execute.test.ts
@@ -22,6 +22,7 @@ import {
   candidateSha,
   cleanupCandidateFixtures,
   createCandidateExecutionFixture,
+  createCandidateOutputExecutionFixture,
   createCandidateOutputFixture,
   emptyCandidateSnapshot,
   redigestCandidateBundle,
@@ -49,26 +50,45 @@ async function terminalTrace(
   })
 }
 
+interface TestExecutor {
+  execute: AgentCandidateExecutorPort['execute']
+  stop: AgentCandidateExecutorPort['stop']
+  capture?: AgentCandidateExecutorPort['capture']
+}
+
 function options(
-  executor: AgentCandidateExecutorPort,
+  executor: TestExecutor,
   traceStore = new InMemoryTraceStore(),
   claimStore = new InMemoryAgentCandidateExecutionClaimStore(),
 ) {
   const outputs = createCandidateOutputFixture()
-  let request: AgentCandidateExecutorRequest | undefined
+  const requests = new Map<string, AgentCandidateExecutorRequest>()
   return {
     executor: {
       execute: async (...args: Parameters<AgentCandidateExecutorPort['execute']>) => {
-        request = args[0]
+        requests.set(args[0].executionId, args[0])
         return await executor.execute(...args)
       },
-      stopAndCapture: async (...args: Parameters<AgentCandidateExecutorPort['stopAndCapture']>) => {
-        const capture = await executor.stopAndCapture(...args)
-        if (capture.taskOutcome || !request) return capture
+      stop: executor.stop,
+      capture: async (
+        stopRequest: Parameters<AgentCandidateExecutorPort['capture']>[0],
+        context: Parameters<AgentCandidateExecutorPort['capture']>[1],
+      ) => {
+        const capture = executor.capture ? await executor.capture(stopRequest, context) : {}
+        if (capture.taskOutcome) return capture
+        const request = requests.get(stopRequest.executionId)
+        if (!request) throw new Error('fixture capture has no execution request')
+        const outcome = request.executionPlan.value.material.task.outcome
+        if (outcome.kind !== 'workspace') {
+          throw new Error('fixture fallback requires a workspace outcome')
+        }
+        const repository = request.executionPlan.value.material.task.repository
+        if (!repository) throw new Error('fixture fallback requires repository identity')
         return {
           ...capture,
           taskOutcome: {
-            resultTree: request.executionPlan.value.material.task.repository.baseTree,
+            kind: 'workspace' as const,
+            resultTree: repository.baseTree,
             afterState: request.executionPlan.value.material.task.workspace.material,
             archive: Buffer.from('fixture unchanged task archive', 'utf8'),
             gitDiff: Buffer.alloc(0),
@@ -99,7 +119,7 @@ describe('atomic prepared candidate execution', () => {
     const traceStore = new InMemoryTraceStore()
     let observed: AgentCandidateExecutorRequest | undefined
     let stopped = 0
-    const executor: AgentCandidateExecutorPort = {
+    const executor: TestExecutor = {
       execute: async (request, context) => {
         observed = request
         expect(context.traceStore).not.toBe(traceStore)
@@ -126,8 +146,11 @@ describe('atomic prepared candidate execution', () => {
           termination: { kind: 'exit', exitCode: 0 },
         }
       },
-      stopAndCapture: async () => {
+      stop: async () => {
         stopped++
+        return { stopped: true }
+      },
+      capture: async () => {
         if (!observed) throw new Error('candidate executor did not receive its request')
         const changedBytes = Buffer.from('export const value = 2\n')
         const taskRoot = fixture.task.stagingRoots.taskRoot
@@ -148,8 +171,8 @@ describe('atomic prepared candidate execution', () => {
           })),
         )
         return {
-          stopped: true,
           taskOutcome: {
+            kind: 'workspace',
             resultTree,
             afterState: captured.snapshot.material,
             archive: captured.archive,
@@ -168,9 +191,13 @@ describe('atomic prepared candidate execution', () => {
       succeeded: true,
       receipt: {
         value: {
-          schemaVersion: 2,
-          usage: { modelCalls: 0, costUsd: 0 },
-          fixedUsage: { modelCalls: 0, costUsdNanos: 0 },
+          schemaVersion: 1,
+          executorCapture: expect.objectContaining({
+            sha256: expect.stringMatching(/^sha256:/),
+          }),
+          modelSettlement: {
+            material: { usage: { modelCalls: 0, costUsdNanos: 0 } },
+          },
           benchmarkResult: {
             material: {
               score: 1,
@@ -184,6 +211,54 @@ describe('atomic prepared candidate execution', () => {
     await expect(
       executionOptions.outputArtifacts.read(result.artifacts.runReceipt),
     ).resolves.toEqual(result.receipt.bytes)
+    expect(result.receipt.value.executorCapture).toEqual(result.artifacts.executorCapture)
+  })
+
+  it('grades and receipts exact normal-agent output through the shared execution path', async () => {
+    const fixture = createCandidateOutputExecutionFixture('application/json', 1_024)
+    const prepared = await prepareAgentCandidateExecution(
+      await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
+      fixture.task,
+      fixture.ports,
+    )
+    const traceStore = new InMemoryTraceStore()
+    const output = Buffer.from('{"recommendation":"add focused failure recovery"}\n', 'utf8')
+    const executionOptions = options(
+      {
+        execute: async (request) => {
+          await terminalTrace(request, traceStore)
+          return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
+        },
+        stop: async () => ({ stopped: true }),
+        capture: async () => ({
+          taskOutcome: { kind: 'output', bytes: output },
+        }),
+      },
+      traceStore,
+    )
+    let gradedOutput = Buffer.alloc(0)
+    executionOptions.grader = {
+      ...executionOptions.grader,
+      run: async (input) => {
+        if (input.outcome.kind !== 'output') throw new Error('expected output outcome')
+        gradedOutput = Buffer.from(input.outcome.bytes)
+        return await createCandidateOutputFixture().grader.run(input)
+      },
+    }
+
+    const result = await executePreparedAgentCandidate(prepared, executionOptions)
+    if (!result.succeeded) throw new Error(result.reason)
+    const captured = result.receipt.value.taskOutcome.material.outcome
+    if (captured.kind !== 'output') throw new Error('expected output receipt')
+    expect(gradedOutput).toEqual(output)
+    await expect(executionOptions.outputArtifacts.read(captured.artifact)).resolves.toEqual(
+      Uint8Array.from(output),
+    )
+    expect(result.receipt.value.benchmarkResult.material).toMatchObject({
+      taskOutcomeDigest: result.receipt.value.taskOutcome.digest,
+      score: 1,
+      passed: true,
+    })
   })
 
   it('authors paid model spans only from the closed router ledger', async () => {
@@ -237,17 +312,24 @@ describe('atomic prepared candidate execution', () => {
               termination: { kind: 'exit', exitCode: 0 },
             }
           },
-          stopAndCapture: async () => ({ stopped: true }),
+          stop: async () => ({ stopped: true }),
         },
         traceStore,
       ),
     )
     if (!result.succeeded) throw new Error(result.reason)
     expect(result.receipt.value).toMatchObject({
-      usage: { modelCalls: 1, inputTokens: 10, outputTokens: 5, costUsd: 0.01 },
       modelSettlement: {
         material: {
-          schemaVersion: 2,
+          schemaVersion: 1,
+          usage: {
+            modelCalls: 1,
+            inputTokens: 10,
+            outputTokens: 5,
+            cachedInputTokens: 2,
+            reasoningTokens: 1,
+            costUsdNanos: 10_000_000,
+          },
           calls: [
             {
               generationId: 'generation-paid-success',
@@ -305,7 +387,8 @@ describe('atomic prepared candidate execution', () => {
           executions++
           throw new Error('must not execute')
         },
-        stopAndCapture: async () => ({ stopped: true }),
+        stop: async () => ({ stopped: true }),
+        capture: async () => ({}),
       }),
     )
     expect(result).toMatchObject({
@@ -365,7 +448,7 @@ describe('atomic prepared candidate execution', () => {
           execute: async () => {
             throw new Error('must not execute')
           },
-          stopAndCapture: async () => ({ stopped: true }),
+          stop: async () => ({ stopped: true }),
         },
         new InMemoryTraceStore(),
         claimStore,
@@ -422,7 +505,7 @@ describe('atomic prepared candidate execution', () => {
             executions++
             throw new Error('must not execute')
           },
-          stopAndCapture: async () => ({ stopped: true }),
+          stop: async () => ({ stopped: true }),
         },
         new InMemoryTraceStore(),
         claimStore,
@@ -502,10 +585,11 @@ describe('atomic prepared candidate execution', () => {
           await terminalTrace(request, traceStore)
           return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
         },
-        stopAndCapture: async () => {
+        stop: async () => {
           now += cleanupTimeoutMs - 1
-          return { stopped: true, taskOutcome: unchangedTaskOutcomeCapture(fixture) }
+          return { stopped: true }
         },
+        capture: async () => ({ taskOutcome: unchangedTaskOutcomeCapture(fixture) }),
       },
       grader: {
         ...outputs.grader,
@@ -549,8 +633,8 @@ describe('atomic prepared candidate execution', () => {
             await terminalTrace(request, traceStore)
             return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
           },
-          stopAndCapture: async () => ({
-            stopped: true,
+          stop: async () => ({ stopped: true }),
+          capture: async () => ({
             taskOutcome: unchangedTaskOutcomeCapture(fixture),
           }),
         },
@@ -606,8 +690,8 @@ describe('atomic prepared candidate execution', () => {
             await terminalTrace(request, traceStore)
             return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
           },
-          stopAndCapture: async () => ({
-            stopped: true,
+          stop: async () => ({ stopped: true }),
+          capture: async () => ({
             taskOutcome: unchangedTaskOutcomeCapture(fixture),
           }),
         },
@@ -669,7 +753,8 @@ describe('atomic prepared candidate execution', () => {
         execute: async () => {
           throw new Error('must not execute')
         },
-        stopAndCapture: async () => ({ stopped: true }),
+        stop: async () => ({ stopped: true }),
+        capture: async () => ({}),
       }),
     )
     expect(result).toMatchObject({
@@ -711,7 +796,7 @@ describe('atomic prepared candidate execution', () => {
             await terminalTrace(request, traceStore)
             return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
           },
-          stopAndCapture: async () => ({ stopped: true }),
+          stop: async () => ({ stopped: true }),
         },
         traceStore,
       ),
@@ -733,7 +818,8 @@ describe('atomic prepared candidate execution', () => {
           await terminalTrace(request, traceStore)
           return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
         },
-        stopAndCapture: async () => ({ stopped: true }),
+        stop: async () => ({ stopped: true }),
+        capture: async () => ({}),
       },
       traceStore,
       claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
@@ -742,7 +828,7 @@ describe('atomic prepared candidate execution', () => {
     expect(result).toMatchObject({
       succeeded: false,
       reason: expect.stringMatching(/without a captured task outcome/),
-      usage: { modelCalls: 0, costUsd: 0 },
+      usage: { modelCalls: 0, costUsdNanos: 0 },
     })
   })
 
@@ -762,9 +848,9 @@ describe('atomic prepared candidate execution', () => {
           await terminalTrace(request, traceStore)
           return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
         },
-        stopAndCapture: async () =>
+        stop: async () => ({ stopped: true }),
+        capture: async () =>
           ({
-            stopped: true,
             taskOutcome: unchangedTaskOutcomeCapture(fixture),
             score: 1,
           }) as never,
@@ -783,7 +869,7 @@ describe('atomic prepared candidate execution', () => {
     expect(result).toMatchObject({
       succeeded: false,
       reason: expect.stringMatching(/unknown field score/),
-      usage: { modelCalls: 0, costUsd: 0 },
+      usage: { modelCalls: 0, costUsdNanos: 0 },
     })
     expect(graderCalls).toBe(0)
   })
@@ -808,8 +894,8 @@ describe('atomic prepared candidate execution', () => {
             score: 1,
           } as never
         },
-        stopAndCapture: async () => ({
-          stopped: true,
+        stop: async () => ({ stopped: true }),
+        capture: async () => ({
           taskOutcome: unchangedTaskOutcomeCapture(fixture),
         }),
       },
@@ -827,7 +913,7 @@ describe('atomic prepared candidate execution', () => {
     expect(result).toMatchObject({
       succeeded: false,
       reason: expect.stringMatching(/unknown field score/),
-      usage: { modelCalls: 0, costUsd: 0 },
+      usage: { modelCalls: 0, costUsdNanos: 0 },
     })
     expect(graderCalls).toBe(0)
   })
@@ -850,7 +936,7 @@ describe('atomic prepared candidate execution', () => {
     const didStart = new Promise<void>((resolve) => {
       started = resolve
     })
-    const executor: AgentCandidateExecutorPort = {
+    const executor: TestExecutor = {
       execute: async (request) => {
         executions++
         started()
@@ -858,7 +944,7 @@ describe('atomic prepared candidate execution', () => {
         await terminalTrace(request, traceStore)
         return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
       },
-      stopAndCapture: async () => ({ stopped: true }),
+      stop: async () => ({ stopped: true }),
     }
     const first = executePreparedAgentCandidate(prepared, options(executor, traceStore, claimStore))
     await didStart
@@ -934,14 +1020,14 @@ describe('atomic prepared candidate execution', () => {
     const didStart = new Promise<void>((resolve) => {
       started = resolve
     })
-    const executor: AgentCandidateExecutorPort = {
+    const executor: TestExecutor = {
       execute: async (request) => {
         started()
         await wait
         await terminalTrace(request, traceStore)
         return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
       },
-      stopAndCapture: async () => ({ stopped: true }),
+      stop: async () => ({ stopped: true }),
     }
 
     const first = executePreparedAgentCandidate(
@@ -1008,13 +1094,13 @@ describe('atomic prepared candidate execution', () => {
       prepared,
       options({
         execute: async () => Promise.reject(new Error(`container failed with ${secret}`)),
-        stopAndCapture: async () => ({ stopped: true }),
+        stop: async () => ({ stopped: true }),
       }),
     )
     expect(result).toMatchObject({
       succeeded: false,
       reason: expect.not.stringContaining(secret),
-      usage: { modelCalls: 1, inputTokens: 10, outputTokens: 5, costUsd: 0.01 },
+      usage: { modelCalls: 1, inputTokens: 10, outputTokens: 5, costUsdNanos: 10_000_000 },
     })
     expect(result.reason).toContain('[redacted:candidate-access]')
     expect({ closed, settlementReads }).toEqual({ closed: true, settlementReads: 1 })
@@ -1058,7 +1144,7 @@ describe('atomic prepared candidate execution', () => {
               })
             })
           },
-          stopAndCapture: async (_request, context) => {
+          stop: async (_request, context) => {
             expect(context.reason).toBe('timeout')
             expect(context.signal).toBe(observedSignal)
             expect(context.deadlineAtMs).toBe(startedAt + 15)
@@ -1106,7 +1192,7 @@ describe('atomic prepared candidate execution', () => {
             await terminalTrace(request, traceStore, 105)
             return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
           },
-          stopAndCapture: async () => {
+          stop: async () => {
             markStopStarted()
             await new Promise((resolve) => setTimeout(resolve, 30))
             return { stopped: true }
@@ -1142,7 +1228,7 @@ describe('atomic prepared candidate execution', () => {
             await terminalTrace(request, traceStore)
             return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
           },
-          stopAndCapture: async () => await new Promise<never>(() => undefined),
+          stop: async () => await new Promise<never>(() => undefined),
         },
         traceStore,
         claimStore,
@@ -1178,7 +1264,7 @@ describe('atomic prepared candidate execution', () => {
             await terminalTrace(request, traceStore)
             return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
           },
-          stopAndCapture: async () => {
+          stop: async () => {
             throw new Error('container death not observed')
           },
         },
@@ -1216,7 +1302,7 @@ describe('atomic prepared candidate execution', () => {
           execute: async () => {
             throw new Error('infrastructure failed')
           },
-          stopAndCapture: async () => ({ stopped: true }),
+          stop: async () => ({ stopped: true }),
         },
         new InMemoryTraceStore(),
         claimStore,
@@ -1245,7 +1331,7 @@ describe('atomic prepared candidate execution', () => {
             retryExecutions++
             throw new Error('retry must not execute')
           },
-          stopAndCapture: async () => ({ stopped: true }),
+          stop: async () => ({ stopped: true }),
         },
         new InMemoryTraceStore(),
         claimStore,
@@ -1285,7 +1371,7 @@ describe('atomic prepared candidate execution', () => {
           execute: async () => {
             throw new Error('executor must not run before model activation')
           },
-          stopAndCapture: async () => ({ stopped: true }),
+          stop: async () => ({ stopped: true }),
         },
         firstTraceStore,
         claimStore,
@@ -1314,7 +1400,7 @@ describe('atomic prepared candidate execution', () => {
             await terminalTrace(request, traceStore)
             return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
           },
-          stopAndCapture: async () => ({ stopped: true }),
+          stop: async () => ({ stopped: true }),
         },
         traceStore,
         claimStore,
@@ -1371,12 +1457,14 @@ describe('atomic prepared candidate execution', () => {
               termination: { kind: 'exit', exitCode: 0 },
             }
           },
-          stopAndCapture: async () => {
+          stop: async () => {
             order.push('process-stop')
+            return { stopped: true }
+          },
+          capture: async () => {
             if (!('content' in after.archive))
               throw new Error('fixture memory archive is not embedded')
             return {
-              stopped: true,
               memoryAfter: {
                 afterState: after.material,
                 archive: Buffer.from(after.archive.content, 'base64'),
@@ -1443,8 +1531,8 @@ describe('atomic prepared candidate execution', () => {
           await terminalTrace(request, traceStore)
           return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
         },
-        stopAndCapture: async () => ({
-          stopped: true,
+        stop: async () => ({ stopped: true }),
+        capture: async () => ({
           taskOutcome: unchangedTaskOutcomeCapture(fixture),
           memoryAfter: { afterState, archive: gzipSync(memoryBytes) },
         }),
@@ -1517,7 +1605,7 @@ describe('atomic prepared candidate execution', () => {
           })
           return { executionId: request.executionId, termination: { kind: 'exit', exitCode: 0 } }
         },
-        stopAndCapture: async () => ({ stopped: true }),
+        stop: async () => ({ stopped: true }),
       },
       traceStore,
     )
@@ -1558,7 +1646,7 @@ describe('atomic prepared candidate execution', () => {
           executions++
           throw new Error('must not execute')
         },
-        stopAndCapture: async () => ({ stopped: true }),
+        stop: async () => ({ stopped: true }),
       }),
     )
     expect(result).toMatchObject({ succeeded: false, reason: expect.stringMatching(/collides/) })
diff --git a/tests/candidate-execution-executor-capture.test.ts b/tests/candidate-execution-executor-capture.test.ts
new file mode 100644
index 00000000..2303dc98
--- /dev/null
+++ b/tests/candidate-execution-executor-capture.test.ts
@@ -0,0 +1,43 @@
+import { describe, expect, it } from 'vitest'
+
+import {
+  sealAgentCandidateExecutorFinalCapture,
+  sealAgentCandidateExecutorStopAcknowledgement,
+} from '../src/candidate-execution/executor-capture'
+
+describe('candidate executor capture', () => {
+  it('checks the signed output limit before detaching exact bytes', () => {
+    expect(() =>
+      sealAgentCandidateExecutorFinalCapture(
+        {
+          taskOutcome: { kind: 'output', bytes: new Uint8Array(5) },
+        },
+        { kind: 'output', mediaType: 'application/octet-stream', maxBytes: 4 },
+      ),
+    ).toThrow(/4-byte maximum/)
+
+    const source = Uint8Array.from([1, 2, 3])
+    const capture = sealAgentCandidateExecutorFinalCapture(
+      { taskOutcome: { kind: 'output', bytes: source } },
+      { kind: 'output', mediaType: 'application/octet-stream', maxBytes: 3 },
+    )
+    source[0] = 9
+    expect(capture.taskOutcome).toMatchObject({ kind: 'output', bytes: Uint8Array.from([1, 2, 3]) })
+  })
+
+  it('rejects a capture that disagrees with the signed result kind', () => {
+    expect(() =>
+      sealAgentCandidateExecutorFinalCapture(
+        { taskOutcome: { kind: 'output', bytes: Uint8Array.from([1]) } },
+        { kind: 'workspace' },
+      ),
+    ).toThrow(/signed outcome kind/)
+  })
+
+  it('seals stop independently from replayable capture evidence', () => {
+    expect(() => sealAgentCandidateExecutorStopAcknowledgement({ stopped: true })).not.toThrow()
+    expect(() =>
+      sealAgentCandidateExecutorStopAcknowledgement({ stopped: true, taskOutcome: {} }),
+    ).toThrow(/unknown field taskOutcome/)
+  })
+})
diff --git a/tests/candidate-execution-finalize.test.ts b/tests/candidate-execution-finalize.test.ts
index 33707c6b..edc2866e 100644
--- a/tests/candidate-execution-finalize.test.ts
+++ b/tests/candidate-execution-finalize.test.ts
@@ -1,6 +1,6 @@
 import { InMemoryTraceStore } from '@tangle-network/agent-eval'
 import { afterEach, describe, expect, it } from 'vitest'
-
+import { sealAgentCandidateExecutorFinalCapture } from '../src/candidate-execution/executor-capture'
 import { finalizeAgentCandidateRun } from '../src/candidate-execution/finalize'
 import {
   type SealedAgentCandidateModelSettlement,
@@ -9,7 +9,7 @@ import {
 import {
   persistCandidateBenchmarkResult,
   persistCandidateModelSettlement,
-  persistVerifiedCandidateTaskOutcome,
+  persistVerifiedAgentCandidateExecutorCapture,
 } from '../src/candidate-execution/outcome-evidence'
 import { prepareAgentCandidateExecution } from '../src/candidate-execution/prepare'
 import { assertPreparedCandidateIntegrity } from '../src/candidate-execution/prepared-state'
@@ -94,24 +94,29 @@ async function finalizePrepared(
       },
     )
   const outputs = overrides.outputs ?? createCandidateOutputFixture()
-  const finalCapture = {
-    stopped: true as const,
-    taskOutcome: {
-      resultTree: state.executionPlan.value.material.task.repository.baseTree,
-      afterState: state.executionPlan.value.material.task.workspace.material,
-      archive: Buffer.from('fixture unchanged task archive', 'utf8'),
-      gitDiff: Buffer.alloc(0),
-    },
+  const expectedOutcome = state.executionPlan.value.material.task.outcome
+  if (expectedOutcome.kind !== 'workspace') {
+    throw new Error('finalization fixture requires a workspace outcome')
   }
-  const [modelSettlement, taskOutcome] = await Promise.all([
+  const repository = state.executionPlan.value.material.task.repository
+  if (!repository) throw new Error('finalization fixture requires repository identity')
+  const finalCapture = sealAgentCandidateExecutorFinalCapture(
+    {
+      taskOutcome: {
+        kind: 'workspace' as const,
+        resultTree: repository.baseTree,
+        afterState: state.executionPlan.value.material.task.workspace.material,
+        archive: Buffer.from('fixture unchanged task archive', 'utf8'),
+        gitDiff: Buffer.alloc(0),
+      },
+    },
+    expectedOutcome,
+  )
+  const [modelSettlement, verifiedCapture] = await Promise.all([
     persistCandidateModelSettlement(state, settlement, outputs.outputArtifacts),
-    persistVerifiedCandidateTaskOutcome(
-      state,
-      finalCapture.taskOutcome,
-      outputs.outputArtifacts,
-      [],
-    ),
+    persistVerifiedAgentCandidateExecutorCapture(state, finalCapture, outputs.outputArtifacts, []),
   ])
+  const { persistedCapture, taskOutcome } = verifiedCapture
   const benchmarkResult = await persistCandidateBenchmarkResult(
     state,
     capture.termination,
@@ -127,6 +132,7 @@ async function finalizePrepared(
     settlement,
     {
       finalCapture,
+      persistedCapture,
       modelSettlement,
       taskOutcome,
       benchmarkResult,
@@ -194,7 +200,7 @@ async function traceStore(
 }
 
 describe('protected candidate run finalization', () => {
-  it('derives exact V2 evidence and durable trace bytes from tied agent-eval spans', async () => {
+  it('derives exact evidence and durable trace bytes from tied agent-eval spans', async () => {
     const execution = await prepared()
     const store = await traceStore(execution)
     const outputs = createCandidateOutputFixture()
@@ -206,40 +212,31 @@ describe('protected candidate run finalization', () => {
     )
     expect(result.succeeded).toBe(true)
     if (!result.succeeded) return
-    expect(result.receipt.value.usage).toEqual({
-      costUsd: 0.01,
-      inputTokens: 10,
-      outputTokens: 5,
-      cachedInputTokens: 2,
-      modelCalls: 1,
-    })
     expect(result.receipt.value.trace).toMatchObject({ eventCount: 3, modelCallCount: 1 })
     expect(result.receipt.bytes.byteLength).toBeGreaterThan(0)
     const task = assertPreparedCandidateIntegrity(execution).executionPlan.value.material.task
+    if (task.outcome.kind !== 'workspace') throw new Error('expected workspace task outcome')
+    if (!task.repository) throw new Error('expected task repository identity')
     expect(result.receipt.value).toMatchObject({
-      schemaVersion: 2,
-      fixedUsage: {
-        costUsdNanos: 10_000_000,
-        inputTokens: 10,
-        outputTokens: 5,
-        cachedInputTokens: 2,
-        modelCalls: 1,
-      },
+      schemaVersion: 1,
       taskOutcome: {
         artifact: result.artifacts.taskOutcome,
         material: {
           executionPlanDigest: execution.executionPlan.value.digest,
-          baseRepository: {
-            identity: task.repository.identity,
-            rootIdentity: task.repository.rootIdentity,
-            commit: task.repository.baseCommit,
-            tree: task.repository.baseTree,
-          },
-          resultRepository: {
-            identity: task.repository.identity,
-            rootIdentity: task.repository.rootIdentity,
-            commit: expect.stringMatching(/^[0-9a-f]{40}$/),
-            tree: task.repository.baseTree,
+          outcome: {
+            kind: 'workspace',
+            baseRepository: {
+              identity: task.repository.identity,
+              rootIdentity: task.repository.rootIdentity,
+              commit: task.repository.baseCommit,
+              tree: task.repository.baseTree,
+            },
+            resultRepository: {
+              identity: task.repository.identity,
+              rootIdentity: task.repository.rootIdentity,
+              commit: expect.stringMatching(/^[0-9a-f]{40}$/),
+              tree: task.repository.baseTree,
+            },
           },
         },
       },
@@ -261,6 +258,7 @@ describe('protected candidate run finalization', () => {
             inputTokens: 10,
             outputTokens: 5,
             cachedInputTokens: 2,
+            reasoningTokens: 0,
             modelCalls: 1,
           },
         },
@@ -271,7 +269,11 @@ describe('protected candidate run finalization', () => {
         result.receipt.bytes,
       ),
       expect(
-        outputs.outputArtifacts.read(result.receipt.value.taskOutcome.material.gitDiff.artifact),
+        outputs.outputArtifacts.read(
+          result.receipt.value.taskOutcome.material.outcome.kind === 'workspace'
+            ? result.receipt.value.taskOutcome.material.outcome.gitDiff.artifact
+            : result.receipt.value.taskOutcome.material.outcome.artifact,
+        ),
       ).resolves.toEqual(new Uint8Array()),
       expect(
         outputs.outputArtifacts.read(result.receipt.value.benchmarkResult.material.evidence),
@@ -455,7 +457,13 @@ describe('protected candidate run finalization', () => {
     )
     expect(result).toMatchObject({
       succeeded: true,
-      receipt: { value: { usage: { costUsd: 0.3, modelCalls: 2 } } },
+      receipt: {
+        value: {
+          modelSettlement: {
+            material: { usage: { costUsdNanos: 300_000_000, modelCalls: 2 } },
+          },
+        },
+      },
     })
   })
 })
diff --git a/tests/candidate-execution-model-port.test.ts b/tests/candidate-execution-model-port.test.ts
index e2464ff5..5f590f8b 100644
--- a/tests/candidate-execution-model-port.test.ts
+++ b/tests/candidate-execution-model-port.test.ts
@@ -1,5 +1,9 @@
 import { InMemoryTraceStore, type LlmSpan } from '@tangle-network/agent-eval'
-import type { AgentCandidateResolvedModel, Sha256Digest } from '@tangle-network/agent-interface'
+import type {
+  AgentCandidateModelSettlementCall,
+  AgentCandidateResolvedModel,
+  Sha256Digest,
+} from '@tangle-network/agent-interface'
 import { describe, expect, it } from 'vitest'
 import {
   appendAuthoritativeModelSettlementSpans,
@@ -17,7 +21,6 @@ import {
 import type {
   AgentCandidateModelPort,
   AgentCandidateProtectedModelActivation,
-  AgentCandidateProtectedModelCall,
   AgentCandidateProtectedModelSettlement,
 } from '../src/candidate-execution/types'
 
@@ -95,8 +98,8 @@ function settleInput(
 
 function modelCall(
   index: number,
-  overrides: Partial<AgentCandidateProtectedModelCall> = {},
-): AgentCandidateProtectedModelCall {
+  overrides: Partial<AgentCandidateModelSettlementCall> = {},
+): AgentCandidateModelSettlementCall {
   return {
     callId: `call-${index}`,
     generationId: `generation-${index}`,
@@ -115,7 +118,7 @@ function modelCall(
 }
 
 function settlement(
-  calls: readonly AgentCandidateProtectedModelCall[] = [],
+  calls: readonly AgentCandidateModelSettlementCall[] = [],
 ): AgentCandidateProtectedModelSettlement {
   return {
     preparationId: 'preparation-1',
diff --git a/tests/candidate-execution-outcome-evidence.test.ts b/tests/candidate-execution-outcome-evidence.test.ts
index 73fa0a41..49a0beaa 100644
--- a/tests/candidate-execution-outcome-evidence.test.ts
+++ b/tests/candidate-execution-outcome-evidence.test.ts
@@ -4,9 +4,10 @@ import { gunzipSync, gzipSync } from 'node:zlib'
 
 import { afterEach, describe, expect, it } from 'vitest'
 import { sha256Bytes } from '../src/candidate-execution/digest'
+import { sealAgentCandidateExecutorFinalCapture } from '../src/candidate-execution/executor-capture'
 import {
   persistCandidateBenchmarkResult,
-  persistVerifiedCandidateTaskOutcome,
+  persistVerifiedAgentCandidateExecutorCapture,
 } from '../src/candidate-execution/outcome-evidence'
 import { prepareAgentCandidateExecution } from '../src/candidate-execution/prepare'
 import { assertPreparedCandidateIntegrity } from '../src/candidate-execution/prepared-state'
@@ -14,6 +15,7 @@ import { verifyAgentCandidateBundle } from '../src/candidate-execution/verify'
 import {
   cleanupCandidateFixtures,
   createCandidateExecutionFixture,
+  createCandidateOutputExecutionFixture,
   createCandidateOutputFixture,
   unchangedTaskOutcomeCapture,
 } from './helpers/candidate-execution-fixture'
@@ -21,6 +23,71 @@ import {
 afterEach(cleanupCandidateFixtures)
 
 describe('candidate outcome evidence', () => {
+  it('persists exact bounded media output and rejects invalid captures', async () => {
+    const fixture = createCandidateOutputExecutionFixture('application/json', 32)
+    const state = await preparedState(fixture)
+    const outputs = createCandidateOutputFixture()
+    const source = Buffer.from('{"answer":42}\n', 'utf8')
+    const outcome = await persistTaskOutcome(
+      state,
+      { kind: 'output', bytes: source },
+      outputs.outputArtifacts,
+      [],
+    )
+
+    source.fill(0)
+    expect(outcome).toMatchObject({
+      kind: 'output',
+      spec: { mediaType: 'application/json', maxBytes: 32 },
+    })
+    if (outcome.kind !== 'output' || outcome.evidence.material.outcome.kind !== 'output') {
+      throw new Error('expected exact output evidence')
+    }
+    expect(Buffer.from(outcome.bytes).toString('utf8')).toBe('{"answer":42}\n')
+    await expect(
+      outputs.outputArtifacts.read(outcome.evidence.material.outcome.artifact),
+    ).resolves.toEqual(Uint8Array.from(Buffer.from('{"answer":42}\n', 'utf8')))
+
+    await expect(
+      persistTaskOutcome(
+        state,
+        { kind: 'output', bytes: new Uint8Array() },
+        outputs.outputArtifacts,
+        [],
+      ),
+    ).rejects.toThrow(/cannot be empty/)
+    await expect(
+      persistTaskOutcome(
+        state,
+        { kind: 'output', bytes: Buffer.alloc(33) },
+        outputs.outputArtifacts,
+        [],
+      ),
+    ).rejects.toThrow(/32-byte maximum/)
+    await expect(
+      persistTaskOutcome(
+        state,
+        { kind: 'output', bytes: Buffer.from('sk-output-secret-123456789') },
+        outputs.outputArtifacts,
+        ['sk-output-secret-123456789'],
+      ),
+    ).rejects.toThrow(/protected value/)
+    await expect(
+      persistTaskOutcome(
+        state,
+        {
+          kind: 'workspace',
+          resultTree: '0'.repeat(40),
+          afterState: fixture.task.workspace.material,
+          archive: Buffer.from('archive'),
+          gitDiff: new Uint8Array(),
+        },
+        outputs.outputArtifacts,
+        [],
+      ),
+    ).rejects.toThrow(/does not match the signed outcome kind/)
+  })
+
   it('rejects protected values in the canonical task manifest before any output write', async () => {
     const fixture = createCandidateExecutionFixture()
     const state = await preparedState(fixture)
@@ -35,12 +102,9 @@ describe('candidate outcome evidence', () => {
     }
 
     await expect(
-      persistVerifiedCandidateTaskOutcome(
-        state,
-        unchangedTaskOutcomeCapture(fixture),
-        outputArtifacts,
-        ['source.ts'],
-      ),
+      persistTaskOutcome(state, unchangedTaskOutcomeCapture(fixture), outputArtifacts, [
+        'source.ts',
+      ]),
     ).rejects.toThrow(/protected value/)
     expect(puts).toBe(0)
   })
@@ -67,7 +131,7 @@ describe('candidate outcome evidence', () => {
     const secret = 'sk-compressed-before-put-123456789'
 
     await expect(
-      persistVerifiedCandidateTaskOutcome(
+      persistTaskOutcome(
         state,
         {
           ...unchangedTaskOutcomeCapture(fixture),
@@ -87,7 +151,7 @@ describe('candidate outcome evidence', () => {
     const secret = 'sk-outcome-secret-123456789'
     const capture = unchangedTaskOutcomeCapture(fixture)
     await expect(
-      persistVerifiedCandidateTaskOutcome(
+      persistTaskOutcome(
         state,
         { ...capture, archive: Buffer.from(secret, 'utf8') },
         outputs.outputArtifacts,
@@ -95,12 +159,7 @@ describe('candidate outcome evidence', () => {
       ),
     ).rejects.toThrow(/protected value/)
 
-    const outcome = await persistVerifiedCandidateTaskOutcome(
-      state,
-      capture,
-      outputs.outputArtifacts,
-      [secret],
-    )
+    const outcome = await persistTaskOutcome(state, capture, outputs.outputArtifacts, [secret])
     await expect(
       persistCandidateBenchmarkResult(
         state,
@@ -131,7 +190,7 @@ describe('candidate outcome evidence', () => {
     const fixture = createCandidateExecutionFixture()
     const state = await preparedState(fixture)
     const outputs = createCandidateOutputFixture()
-    const outcome = await persistVerifiedCandidateTaskOutcome(
+    const outcome = await persistTaskOutcome(
       state,
       unchangedTaskOutcomeCapture(fixture),
       outputs.outputArtifacts,
@@ -161,7 +220,7 @@ describe('candidate outcome evidence', () => {
     const fixture = createCandidateExecutionFixture()
     const state = await preparedState(fixture)
     const outputs = createCandidateOutputFixture()
-    const outcome = await persistVerifiedCandidateTaskOutcome(
+    const outcome = await persistTaskOutcome(
       state,
       unchangedTaskOutcomeCapture(fixture),
       outputs.outputArtifacts,
@@ -211,7 +270,7 @@ describe('candidate outcome evidence', () => {
     const fixture = createCandidateExecutionFixture()
     const state = await preparedState(fixture)
     const outputs = createCandidateOutputFixture()
-    const outcome = await persistVerifiedCandidateTaskOutcome(
+    const outcome = await persistTaskOutcome(
       state,
       unchangedTaskOutcomeCapture(fixture),
       outputs.outputArtifacts,
@@ -272,3 +331,23 @@ async function preparedState(fixture: ReturnType<typeof createCandidateExecution
     ),
   )
 }
+
+async function persistTaskOutcome(
+  state: Parameters<typeof persistVerifiedAgentCandidateExecutorCapture>[0],
+  taskOutcome: import('../src/candidate-execution/types').AgentCandidateExecutorTaskOutcomeCapture,
+  outputArtifacts: Parameters<typeof persistVerifiedAgentCandidateExecutorCapture>[2],
+  protectedValues: readonly string[],
+) {
+  const capture = sealAgentCandidateExecutorFinalCapture(
+    { taskOutcome },
+    state.executionPlan.value.material.task.outcome,
+  )
+  return (
+    await persistVerifiedAgentCandidateExecutorCapture(
+      state,
+      capture,
+      outputArtifacts,
+      protectedValues,
+    )
+  ).taskOutcome
+}
diff --git a/tests/candidate-execution-prepare.test.ts b/tests/candidate-execution-prepare.test.ts
index c1bf8239..df7d0212 100644
--- a/tests/candidate-execution-prepare.test.ts
+++ b/tests/candidate-execution-prepare.test.ts
@@ -264,6 +264,7 @@ function fixture(active = false): {
       baseCommit: repository.commit,
       baseTree: repository.tree,
     },
+    outcome: { kind: 'workspace' },
     attempt: { number: 1, maxAttempts: 1, retryPolicy: 'none' },
     model: { requested: 'provider/model', reasoningEffort: 'high' },
     grader: {
@@ -327,6 +328,7 @@ describe('candidate execution preparation', () => {
       byteLength: Buffer.byteLength(value.task.instruction),
       delivery: value.bundle.execution.instructionDelivery,
     })
+    expect(plan.task.outcome).toEqual(value.task.outcome)
     expect(plan.task.repository).toEqual(value.task.repository)
     expect(plan.profile).toEqual({
       planDigest: prepared.profilePlan.value.digest,
@@ -394,7 +396,6 @@ describe('candidate execution preparation', () => {
         [field]: profileValue,
       } as AgentCandidateBundle['profile'],
     })
-    const verified = await verifyAgentCandidateBundle(value.bundle, value.ports)
     let stagingCalls = 0
     let reservationCalls = 0
     value.ports.workspaces.materialize = async () => {
@@ -405,7 +406,7 @@ describe('candidate execution preparation', () => {
       throw new Error('candidate must fail before reserving protected access')
     }
 
-    await expect(prepareAgentCandidateExecution(verified, value.task, value.ports)).rejects.toThrow(
+    await expect(verifyAgentCandidateBundle(value.bundle, value.ports)).rejects.toThrow(
       new RegExp(`non-empty AgentProfile fields: ${field}`),
     )
     expect(stagingCalls).toBe(0)
@@ -434,6 +435,7 @@ describe('candidate execution preparation', () => {
 
   it('rejects task Git drift, dirty profile staging, and unenforced model limits', async () => {
     const gitDrift = fixture()
+    if (!gitDrift.task.repository) throw new Error('expected repository identity')
     gitDrift.task.repository.baseTree = '0'.repeat(40)
     await expect(
       prepareAgentCandidateExecution(
@@ -443,6 +445,22 @@ describe('candidate execution preparation', () => {
       ),
     ).rejects.toThrow(/base tree/)
 
+    const outputGitDrift = fixture()
+    if (!outputGitDrift.task.repository) throw new Error('expected repository identity')
+    outputGitDrift.task.outcome = {
+      kind: 'output',
+      mediaType: 'application/json',
+      maxBytes: 1_024,
+    }
+    outputGitDrift.task.repository.baseTree = '0'.repeat(40)
+    await expect(
+      prepareAgentCandidateExecution(
+        await verifyAgentCandidateBundle(outputGitDrift.bundle, outputGitDrift.ports),
+        outputGitDrift.task,
+        outputGitDrift.ports,
+      ),
+    ).rejects.toThrow(/base tree/)
+
     const profileDrift = fixture()
     writeFileSync(join(profileDrift.task.stagingRoots.profileRoot, 'stale'), 'stale')
     await expect(
@@ -614,28 +632,20 @@ describe('candidate execution preparation', () => {
     ).rejects.toThrow(/do not match|unsupported mode/)
   })
 
-  it('resets task memory into an execution-scoped namespace and carries verified knowledge', async () => {
+  it('resets task memory into an execution-scoped namespace', async () => {
     const value = fixture()
     value.task.executionId = '..'
     const seedBytes = Buffer.from('seed-memory')
-    const knowledgeBytes = Buffer.from('{"documents":[]}')
     const seed = {
       locator: { kind: 's3' as const, bucket: 'test-artifacts', key: 'memory/seed.bin' },
       sha256: embeddedCandidateArtifact(seedBytes).sha256,
       byteLength: seedBytes.byteLength,
     }
-    const knowledge = {
-      locator: { kind: 's3' as const, bucket: 'test-artifacts', key: 'knowledge/manifest.json' },
-      sha256: embeddedCandidateArtifact(knowledgeBytes).sha256,
-      byteLength: knowledgeBytes.byteLength,
-    }
     value.bundle = redigestBundle(value.bundle, {
       memory: { mode: 'isolated', scope: 'task', seed },
-      knowledge: { snapshotId: 'knowledge-1', manifest: knowledge },
     })
     value.ports.artifacts.read = async (ref) => {
       if (ref.sha256 === seed.sha256) return seedBytes
-      if (ref.sha256 === knowledge.sha256) return knowledgeBytes
       throw new Error(`unexpected artifact ${ref.sha256}`)
     }
     let resetInput: Parameters<AgentCandidateExecutionPorts['memory']['reset']>[0] | undefined
@@ -663,11 +673,29 @@ describe('candidate execution preparation', () => {
       seedDigest: seed.sha256,
     })
     expect(JSON.stringify(prepared)).not.toContain('TANGLE_MEMORY_NAMESPACE')
-    expect(prepared.knowledge).toMatchObject({
-      snapshotId: 'knowledge-1',
-      manifestDigest: knowledge.sha256,
+  })
+
+  it('rejects snapshot-only knowledge before artifact access', async () => {
+    const value = fixture()
+    const bytes = Buffer.from('{"documents":[]}')
+    const manifest = {
+      locator: { kind: 's3' as const, bucket: 'test-artifacts', key: 'knowledge/manifest.json' },
+      sha256: embeddedCandidateArtifact(bytes).sha256,
+      byteLength: bytes.byteLength,
+    }
+    value.bundle = redigestBundle(value.bundle, {
+      knowledge: { snapshotId: 'knowledge-1', manifest },
     })
-    expect(Buffer.from(prepared.knowledge?.manifest ?? [])).toEqual(knowledgeBytes)
+    let reads = 0
+    value.ports.artifacts.read = async () => {
+      reads++
+      return bytes
+    }
+
+    await expect(verifyAgentCandidateBundle(value.bundle, value.ports)).rejects.toThrow(
+      /"knowledge"[\s\S]*"candidate"/,
+    )
+    expect(reads).toBe(0)
   })
 
   it('closes memory immediately when reset evidence fails before preparation can retain it', async () => {
diff --git a/tests/candidate-execution-recover.test.ts b/tests/candidate-execution-recover.test.ts
index 4e3e380b..a292d117 100644
--- a/tests/candidate-execution-recover.test.ts
+++ b/tests/candidate-execution-recover.test.ts
@@ -11,7 +11,10 @@ import {
 } from '../src/candidate-execution/claim'
 import { prepareAgentCandidateExecution } from '../src/candidate-execution/prepare'
 import { recoverExpiredAgentCandidateExecution } from '../src/candidate-execution/recover'
-import type { AgentCandidateOutputArtifactPort } from '../src/candidate-execution/types'
+import type {
+  AgentCandidateOutputArtifactPort,
+  PreparedAgentCandidateExecution,
+} from '../src/candidate-execution/types'
 import { verifyAgentCandidateBundle } from '../src/candidate-execution/verify'
 import {
   candidateSha,
@@ -33,7 +36,8 @@ describe('expired candidate recovery', () => {
     )
     let now = Date.now()
     const claimStore = new InMemoryAgentCandidateExecutionClaimStore({ now: () => now })
-    const claim = candidateExecutionClaim(prepared)
+    const outputArtifacts = outputArtifactPort()
+    const claim = await persistedClaim(prepared, outputArtifacts)
     const acquired = await claimStore.tryClaim(claim)
     expect(acquired.acquired).toBe(true)
     now = claim.leaseExpiresAtMs
@@ -44,8 +48,6 @@ describe('expired candidate recovery', () => {
       settlements++
       return await originalSettle(input)
     }
-    const outputArtifacts = outputArtifactPort()
-
     const result = await recoverExpiredAgentCandidateExecution({
       attempt: { executionId: claim.executionId, attempt: claim.attempt },
       claimStore,
@@ -57,7 +59,7 @@ describe('expired candidate recovery', () => {
         execute: async () => {
           throw new Error('recovery must not execute')
         },
-        stopAndCapture: async (request, context) => {
+        stop: async (request, context) => {
           stops++
           expect(request).toEqual({
             executionId: claim.executionId,
@@ -67,6 +69,7 @@ describe('expired candidate recovery', () => {
           expect(context.deadlineAtMs).toBe(claim.leaseExpiresAtMs)
           return { stopped: true }
         },
+        capture: async () => ({ evidence: Buffer.from('official recovery evidence') }),
       },
     })
     expect(result).toMatchObject({
@@ -101,9 +104,12 @@ describe('expired candidate recovery', () => {
         execute: async () => {
           throw new Error('recovery must not execute')
         },
-        stopAndCapture: async () => {
+        stop: async () => {
           throw new Error('terminal recovery must not stop twice')
         },
+        capture: async () => {
+          throw new Error('terminal recovery must not capture twice')
+        },
       },
     })
     expect(replay).toMatchObject({ finished: false, exactReplay: true })
@@ -119,10 +125,10 @@ describe('expired candidate recovery', () => {
     )
     let now = Date.now()
     const claimStore = new InMemoryAgentCandidateExecutionClaimStore({ now: () => now })
-    const claim = candidateExecutionClaim(prepared)
+    const outputArtifacts = outputArtifactPort()
+    const claim = await persistedClaim(prepared, outputArtifacts)
     await claimStore.tryClaim(claim)
     let stops = 0
-    const outputArtifacts = outputArtifactPort()
     const recover = () =>
       recoverExpiredAgentCandidateExecution({
         attempt: { executionId: claim.executionId, attempt: claim.attempt },
@@ -136,10 +142,11 @@ describe('expired candidate recovery', () => {
           execute: async () => {
             throw new Error('recovery must not execute')
           },
-          stopAndCapture: async () => {
+          stop: async () => {
             stops++
             throw new Error('process still live')
           },
+          capture: async () => ({}),
         },
       })
 
@@ -167,7 +174,8 @@ describe('expired candidate recovery', () => {
     )
     let now = Date.now()
     const claimStore = new InMemoryAgentCandidateExecutionClaimStore({ now: () => now })
-    const firstClaim = candidateExecutionClaim(first)
+    const outputArtifacts = outputArtifactPort()
+    const firstClaim = await persistedClaim(first, outputArtifacts)
     await claimStore.tryClaim(firstClaim)
     now = firstClaim.leaseExpiresAtMs
     const recovered = await recoverExpiredAgentCandidateExecution({
@@ -175,13 +183,14 @@ describe('expired candidate recovery', () => {
       claimStore,
       traceStore: new InMemoryTraceStore(),
       ports: fixture.ports,
-      outputArtifacts: outputArtifactPort(),
+      outputArtifacts,
       now: () => now,
       executor: {
         execute: async () => {
           throw new Error('recovery must not execute')
         },
-        stopAndCapture: async () => ({ stopped: true }),
+        stop: async () => ({ stopped: true }),
+        capture: async () => ({}),
       },
     })
     expect(recovered).toMatchObject({
@@ -201,9 +210,9 @@ describe('expired candidate recovery', () => {
       fixture.task,
       fixture.ports,
     )
-    await expect(claimStore.tryClaim(candidateExecutionClaim(second))).resolves.toMatchObject({
-      acquired: true,
-    })
+    await expect(
+      claimStore.tryClaim(await persistedClaim(second, outputArtifacts)),
+    ).resolves.toMatchObject({ acquired: true })
   })
 
   it('repeats idempotent cleanup after a partial recovery failure', async () => {
@@ -239,11 +248,20 @@ describe('expired candidate recovery', () => {
     )
     let now = Date.now()
     const claimStore = new InMemoryAgentCandidateExecutionClaimStore({ now: () => now })
-    const claim = candidateExecutionClaim(prepared)
+    const artifactStore = outputArtifactPort()
+    const persistedPurposes: string[] = []
+    const outputArtifacts: AgentCandidateOutputArtifactPort = {
+      ...artifactStore,
+      put: async (input) => {
+        persistedPurposes.push(input.purpose)
+        return artifactStore.put(input)
+      },
+    }
+    const claim = await persistedClaim(prepared, outputArtifacts)
     await claimStore.tryClaim(claim)
     now = claim.leaseExpiresAtMs
     let stops = 0
-    const outputArtifacts = outputArtifactPort()
+    let captures = 0
     const recover = () =>
       recoverExpiredAgentCandidateExecution({
         attempt: { executionId: claim.executionId, attempt: claim.attempt },
@@ -256,20 +274,26 @@ describe('expired candidate recovery', () => {
           execute: async () => {
             throw new Error('recovery must not execute')
           },
-          stopAndCapture: async () => {
+          stop: async () => {
             stops++
             return { stopped: true }
           },
+          capture: async () => {
+            captures++
+            return { evidence: Buffer.from(`capture-${captures}`) }
+          },
         },
       })
 
     await expect(recover()).rejects.toThrow(/cleanup could not be proven/)
+    expect(persistedPurposes).toContain('executor-capture')
     expect(
       await claimStore.getAttempt({ executionId: claim.executionId, attempt: 1 }),
     ).not.toHaveProperty('terminal')
     await expect(recover()).resolves.toMatchObject({ finished: true })
-    expect({ stops, settlements, memoryCloses }).toEqual({
+    expect({ stops, captures, settlements, memoryCloses }).toEqual({
       stops: 2,
+      captures: 2,
       settlements: 2,
       memoryCloses: 2,
     })
@@ -284,7 +308,8 @@ describe('expired candidate recovery', () => {
     )
     let now = Date.now()
     const claimStore = new InMemoryAgentCandidateExecutionClaimStore({ now: () => now })
-    const claim = candidateExecutionClaim(prepared)
+    const outputArtifacts = outputArtifactPort()
+    const claim = await persistedClaim(prepared, outputArtifacts)
     await claimStore.tryClaim(claim)
     now = claim.leaseExpiresAtMs
     const traceStore = new InMemoryTraceStore()
@@ -295,13 +320,13 @@ describe('expired candidate recovery', () => {
         claimStore,
         traceStore,
         ports: fixture.ports,
-        outputArtifacts: outputArtifactPort(),
+        outputArtifacts,
         now: () => now,
         executor: {
           execute: async () => {
             throw new Error('recovery must not execute')
           },
-          stopAndCapture: async (_request, context) => {
+          stop: async (_request, context) => {
             await context.traceStore.appendEvent({
               runId: claim.cleanup.traceRunId,
               eventId: 'late-secret',
@@ -311,6 +336,7 @@ describe('expired candidate recovery', () => {
             })
             return { stopped: true }
           },
+          capture: async () => ({}),
         },
       }),
     ).rejects.toThrow(/cleanup could not be proven/)
@@ -321,6 +347,25 @@ describe('expired candidate recovery', () => {
   })
 })
 
+async function persistedClaim(
+  prepared: PreparedAgentCandidateExecution,
+  outputArtifacts: AgentCandidateOutputArtifactPort,
+) {
+  const [executionPlan, materializationReceipt] = await Promise.all([
+    outputArtifacts.put({
+      executionId: prepared.executionId,
+      purpose: 'execution-plan',
+      bytes: prepared.executionPlan.bytes,
+    }),
+    outputArtifacts.put({
+      executionId: prepared.executionId,
+      purpose: 'materialization-receipt',
+      bytes: prepared.materializationReceipt.bytes,
+    }),
+  ])
+  return candidateExecutionClaim(prepared, { executionPlan, materializationReceipt })
+}
+
 function outputArtifactPort(): AgentCandidateOutputArtifactPort {
   const stored = new Map<string, Uint8Array>()
   return {
diff --git a/tests/candidate-execution-task-outcome.test.ts b/tests/candidate-execution-task-outcome.test.ts
index 4b907464..c065aa8c 100644
--- a/tests/candidate-execution-task-outcome.test.ts
+++ b/tests/candidate-execution-task-outcome.test.ts
@@ -67,15 +67,14 @@ function changedOutcome(root: string, baseTree: string) {
 describe('candidate task outcome verification', () => {
   it('proves a captured binary patch and after-state against the signed base tree', async () => {
     const fixture = createCandidateExecutionFixture()
-    const outcome = changedOutcome(
-      fixture.task.stagingRoots.taskRoot,
-      fixture.task.repository.baseTree,
-    )
+    if (!fixture.task.repository) throw new Error('expected repository identity')
+    const repository = fixture.task.repository
+    const outcome = changedOutcome(fixture.task.stagingRoots.taskRoot, repository.baseTree)
     await expect(
       verifyTaskOutcomePatch({
         repositoryRoot: fixture.task.stagingRoots.taskRoot,
-        baseCommit: fixture.task.repository.baseCommit,
-        baseTree: fixture.task.repository.baseTree,
+        baseCommit: repository.baseCommit,
+        baseTree: repository.baseTree,
         ...outcome,
       }),
     ).resolves.toMatchObject({
@@ -86,24 +85,23 @@ describe('candidate task outcome verification', () => {
 
   it('rejects a claimed result tree or after-state that does not match the patch', async () => {
     const fixture = createCandidateExecutionFixture()
-    const outcome = changedOutcome(
-      fixture.task.stagingRoots.taskRoot,
-      fixture.task.repository.baseTree,
-    )
+    if (!fixture.task.repository) throw new Error('expected repository identity')
+    const repository = fixture.task.repository
+    const outcome = changedOutcome(fixture.task.stagingRoots.taskRoot, repository.baseTree)
     await expect(
       verifyTaskOutcomePatch({
         repositoryRoot: fixture.task.stagingRoots.taskRoot,
-        baseCommit: fixture.task.repository.baseCommit,
-        baseTree: fixture.task.repository.baseTree,
+        baseCommit: repository.baseCommit,
+        baseTree: repository.baseTree,
         ...outcome,
-        resultTree: '0'.repeat(fixture.task.repository.baseTree.length),
+        resultTree: '0'.repeat(repository.baseTree.length),
       }),
     ).rejects.toThrow(/materialized tree/)
     await expect(
       verifyTaskOutcomePatch({
         repositoryRoot: fixture.task.stagingRoots.taskRoot,
-        baseCommit: fixture.task.repository.baseCommit,
-        baseTree: fixture.task.repository.baseTree,
+        baseCommit: repository.baseCommit,
+        baseTree: repository.baseTree,
         ...outcome,
         afterState: {
           ...outcome.afterState,
diff --git a/tests/candidate-execution-workspace-archive.test.ts b/tests/candidate-execution-workspace-archive.test.ts
index 0a2fb4de..52ddfbcd 100644
--- a/tests/candidate-execution-workspace-archive.test.ts
+++ b/tests/candidate-execution-workspace-archive.test.ts
@@ -7,6 +7,7 @@ import {
   mkdtempSync,
   readFileSync,
   rmSync,
+  statSync,
   symlinkSync,
   truncateSync,
   writeFileSync,
@@ -112,7 +113,7 @@ describe('candidate workspace archive', () => {
 
   it('restores a non-repository workspace and rejects archive drift', async () => {
     const source = temporaryRoot('candidate-workspace-files-')
-    writeFileSync(join(source, 'input.txt'), 'exact bytes', { mode: 0o644 })
+    writeFileSync(join(source, 'input.txt'), 'exact bytes', { mode: 0o664 })
     const captured = await captureAgentCandidateWorkspace(source)
     const port = createAgentCandidateWorkspacePort()
     const destination = join(temporaryRoot('candidate-workspace-parent-'), 'restored')
@@ -123,6 +124,7 @@ describe('candidate workspace archive', () => {
       destination,
     })
     expect(readFileSync(join(destination, 'input.txt'), 'utf8')).toBe('exact bytes')
+    expect(statSync(join(destination, 'input.txt')).mode & 0o777).toBe(0o664)
 
     const tampered = Uint8Array.from(captured.archive)
     const contentOffset = Buffer.from(tampered).indexOf('exact bytes')
diff --git a/tests/helpers/candidate-execution-fixture.ts b/tests/helpers/candidate-execution-fixture.ts
index 8f663049..18b2d473 100644
--- a/tests/helpers/candidate-execution-fixture.ts
+++ b/tests/helpers/candidate-execution-fixture.ts
@@ -161,7 +161,14 @@ export function candidateBundle(
             evaluation: { costUsd: 0, inputTokens: 0, outputTokens: 0, modelCalls: 0 },
           },
         }
-      : { source: 'human' as const },
+      : {
+          source: 'human' as const,
+          benchmark: {
+            name: 'development',
+            version: '1',
+            splitDigest: candidateSha('f'),
+          },
+        },
   }
   return { ...value, digest: canonicalCandidateDigest(value) }
 }
@@ -202,7 +209,12 @@ export interface CandidateExecutionFixture {
 export function unchangedTaskOutcomeCapture(
   fixture: CandidateExecutionFixture,
 ): AgentCandidateExecutorTaskOutcomeCapture {
+  if (fixture.task.outcome.kind !== 'workspace') {
+    throw new Error('unchanged workspace capture requires a workspace outcome')
+  }
+  if (!fixture.task.repository) throw new Error('workspace task is missing repository identity')
   return {
+    kind: 'workspace',
     resultTree: fixture.task.repository.baseTree,
     afterState: fixture.task.workspace.material,
     archive: Buffer.from(`task-archive:${fixture.task.workspace.digest}`, 'utf8'),
@@ -357,6 +369,7 @@ export function createCandidateExecutionFixture(active = false): CandidateExecut
       baseCommit: repository.commit,
       baseTree: repository.tree,
     },
+    outcome: { kind: 'workspace' },
     attempt: { number: 1, maxAttempts: 1, retryPolicy: 'none' },
     model: { requested: 'provider/model', reasoningEffort: 'high' },
     grader: {
@@ -400,3 +413,28 @@ export function createCandidateExecutionFixture(active = false): CandidateExecut
     ...(candidateRoot ? { candidateRoot } : {}),
   }
 }
+
+export function createCandidateOutputExecutionFixture(
+  mediaType = 'application/json',
+  maxBytes = 1_048_576,
+  withRepository = false,
+): CandidateExecutionFixture {
+  const fixture = createCandidateExecutionFixture()
+  if (withRepository) {
+    return {
+      ...fixture,
+      task: { ...fixture.task, outcome: { kind: 'output', mediaType, maxBytes } },
+    }
+  }
+  const taskRoot = temporaryRoot('candidate-output-task-')
+  const { repository: _repository, ...taskWithoutRepository } = fixture.task
+  return {
+    ...fixture,
+    task: {
+      ...taskWithoutRepository,
+      outcome: { kind: 'output', mediaType, maxBytes },
+      workspace: emptyCandidateSnapshot('output-task'),
+      stagingRoots: { ...fixture.task.stagingRoots, taskRoot },
+    },
+  }
+}
diff --git a/tests/improvement-cycle.test.ts b/tests/improvement-cycle.test.ts
index 93e2ba7b..6c27844c 100644
--- a/tests/improvement-cycle.test.ts
+++ b/tests/improvement-cycle.test.ts
@@ -1,27 +1,36 @@
 import { type AnalystFinding, InMemoryTraceStore } from '@tangle-network/agent-eval'
 import type {
+  CodeSurface,
   DispatchContext,
   JudgeConfig,
   MutableSurface,
   Scenario,
   SurfaceProposer,
 } from '@tangle-network/agent-eval/contract'
+import type { CandidateExecutionEvidence } from '@tangle-network/agent-interface'
 import { afterEach, describe, expect, it, vi } from 'vitest'
 import type { AnalystRegistryLike } from '../src/analyst-loop/types'
 import { buildAgentCandidateBundle } from '../src/candidate-execution/builder'
+import { sealAgentCandidateBundle } from '../src/candidate-execution/bundle'
 import { InMemoryAgentCandidateExecutionClaimStore } from '../src/candidate-execution/claim'
+import {
+  canonicalCandidateDocument,
+  embeddedCandidateArtifact,
+} from '../src/candidate-execution/digest'
 import { assertCandidateProfileBinding } from '../src/candidate-execution/profile'
 import type {
   AgentCandidateExecutorPort,
   AgentCandidateExecutorRequest,
 } from '../src/candidate-execution/types'
 import {
+  createAgentImprovementMeasuredComparison,
   createAgentImprovementProposal,
   executeApprovedAgentCandidate,
   proposeAgentImprovement,
   reviewAgentImprovementProposal,
   verifyAgentImprovementProposal,
   verifyAgentImprovementReview,
+  verifyCandidateExecutionEvidence,
 } from '../src/intelligence/improvement-cycle'
 import {
   candidateSha,
@@ -226,20 +235,22 @@ describe('agent improvement lifecycle', () => {
       now: () => new Date('2026-07-10T01:00:00.000Z'),
     })
 
-    expect(proposed.proposal.evaluation.gateDecision).toBe('ship')
-    expect(proposed.proposal.evaluation.lift).toBeGreaterThan(0)
+    expect(proposed.proposal.evaluation.decision.outcome).toBe('ship')
+    expect(proposed.proposal.evaluation.overall.delta).toBeGreaterThan(0)
+    expect(proposed.proposal.changedSurfaces).toEqual(['prompt'])
     expect(proposed.proposal.findings).toEqual([finding])
-    expect(proposed.proposal.candidateProfile.prompt?.systemPrompt).toBe('PROMOTED')
+    expect(proposed.proposal.candidateBundle?.profile.prompt?.systemPrompt).toBe('PROMOTED')
     expect(proposed.proposal.candidateBundle?.digest).toMatch(/^sha256:[a-f0-9]{64}$/)
     expect(verifyAgentImprovementProposal(proposed.proposal)).toEqual(proposed.proposal)
+    expect(() =>
+      verifyAgentImprovementProposal({ ...proposed.proposal, runId: 'tampered-run' }),
+    ).toThrow(/proposal digest does not match/)
     expect(
       createAgentImprovementProposal({
         runId: 'analysis-run-1',
-        surface: 'prompt',
         baselineProfile: profile,
-        candidateProfile: proposed.improvement.profile,
         findings: proposed.analysis.analystResult.findings,
-        evaluation: proposed.improvement.raw,
+        evaluation: proposed.proposal.evaluation,
         candidateBundle: proposed.proposal.candidateBundle,
         now: () => new Date('2026-07-10T01:00:00.000Z'),
       }),
@@ -247,32 +258,215 @@ describe('agent improvement lifecycle', () => {
     expect(() =>
       createAgentImprovementProposal({
         runId: 'analysis-run-unmeasured',
-        surface: 'prompt',
         baselineProfile: profile,
-        candidateProfile: {
+        findings: proposed.analysis.analystResult.findings,
+        evaluation: proposed.proposal.evaluation,
+        candidateBundle: alignedBundle(bundleInput, {
           ...proposed.improvement.profile,
           prompt: { systemPrompt: 'UNMEASURED' },
-        },
-        findings: proposed.analysis.analystResult.findings,
-        evaluation: proposed.improvement.raw,
+        }),
       }),
-    ).toThrow(/does not match the measured winner surface/)
+    ).toThrow(/does not bind the exact candidate bundle/)
     expect(() =>
       createAgentImprovementProposal({
-        runId: 'analysis-run-malformed-profile',
-        surface: 'agent-profile',
-        baselineProfile: profile,
-        candidateProfile: profile,
+        runId: 'analysis-run-baseline-drift',
+        baselineProfile: { ...profile, name: 'different-baseline' },
         findings: proposed.analysis.analystResult.findings,
-        evaluation: {
+        evaluation: proposed.proposal.evaluation,
+        candidateBundle: proposed.proposal.candidateBundle,
+      }),
+    ).toThrow(/does not bind the included baseline profile/)
+
+    const firstMeasuredCell = proposed.improvement.raw.raw.baselineOnHoldout.cells[0]
+    if (!firstMeasuredCell) throw new Error('expected a measured heldout cell')
+    const primaryObjective = Object.keys(firstMeasuredCell.judgeScores)[0]
+    const primaryScore = primaryObjective
+      ? firstMeasuredCell.judgeScores[primaryObjective]
+      : undefined
+    const primaryDimension = primaryScore ? Object.keys(primaryScore.dimensions)[0] : undefined
+    if (!primaryObjective || !primaryScore || !primaryDimension) {
+      throw new Error('expected a measured objective and dimension')
+    }
+    const addSecondJudge = (cell: typeof firstMeasuredCell) => ({
+      ...cell,
+      judgeScores: {
+        ...cell.judgeScores,
+        secondary: {
+          composite: cell.judgeScores[primaryObjective]!.composite,
+          dimensions: { ...cell.judgeScores[primaryObjective]!.dimensions },
+        },
+      },
+    })
+    const multiJudgeResult = {
+      ...proposed.improvement.raw,
+      raw: {
+        ...proposed.improvement.raw.raw,
+        baselineOnHoldout: {
+          ...proposed.improvement.raw.raw.baselineOnHoldout,
+          cells: proposed.improvement.raw.raw.baselineOnHoldout.cells.map(addSecondJudge),
+        },
+        winnerOnHoldout: {
+          ...proposed.improvement.raw.raw.winnerOnHoldout,
+          cells: proposed.improvement.raw.raw.winnerOnHoldout.cells.map(addSecondJudge),
+        },
+      },
+    }
+    const multiJudgeComparison = createAgentImprovementMeasuredComparison({
+      result: multiJudgeResult,
+      measuredSurface: 'prompt',
+      baselineProfile: profile,
+      candidateBundle: proposed.proposal.candidateBundle,
+    })
+    expect(multiJudgeComparison.objectives).toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({ kind: 'objective', name: primaryObjective }),
+        expect.objectContaining({ kind: 'objective', name: 'secondary' }),
+        expect.objectContaining({
+          kind: 'dimension',
+          objective: primaryObjective,
+          name: primaryDimension,
+        }),
+        expect.objectContaining({
+          kind: 'dimension',
+          objective: 'secondary',
+          name: primaryDimension,
+        }),
+      ]),
+    )
+
+    const compoundProfile = {
+      ...proposed.improvement.profile,
+      tools: { inspect_repository: true },
+    }
+    const compoundInput = alignedBundle(bundleInput, compoundProfile)
+    const compoundBundle = sealAgentCandidateBundle({
+      ...compoundInput,
+      profile: { ...compoundInput.profile, tools: compoundProfile.tools },
+    })
+    const compoundComparison = createAgentImprovementMeasuredComparison({
+      result: {
+        ...proposed.improvement.raw,
+        winner: {
+          ...proposed.improvement.raw.winner,
+          surface: JSON.stringify(compoundProfile),
+        },
+      },
+      measuredSurface: 'agent-profile',
+      baselineProfile: profile,
+      candidateBundle: compoundBundle,
+    })
+    const compoundProposal = createAgentImprovementProposal({
+      runId: 'analysis-run-compound-profile',
+      baselineProfile: profile,
+      findings: proposed.analysis.analystResult.findings,
+      evaluation: compoundComparison,
+      candidateBundle: compoundBundle,
+    })
+    expect(compoundProposal.changedSurfaces).toEqual(['prompt', 'tools'])
+    const { digest: _compoundDigest, ...compoundWithoutDigest } = compoundProposal
+    const omittedSurfaceProposal = canonicalCandidateDocument<typeof compoundProposal>({
+      ...compoundWithoutDigest,
+      changedSurfaces: ['prompt'],
+    }).value
+    expect(() => verifyAgentImprovementProposal(omittedSurfaceProposal)).toThrow(
+      /changed surfaces do not match/,
+    )
+
+    const codeFixture = createCandidateExecutionFixture(true)
+    const { digest: _codeDigest, ...codeBundleInput } = codeFixture.bundle
+    if (!codeFixture.task.repository) throw new Error('expected repository fixture')
+    const patch = Buffer.from('diff --git a/source.ts b/source.ts\n', 'utf8')
+    const candidateTree = 'f'.repeat(codeFixture.task.repository.baseTree.length)
+    const codeSurface: CodeSurface = {
+      kind: 'code',
+      worktreeRef: '/tmp/measured-code-winner',
+      baseRef: 'main',
+      baseCommit: codeFixture.task.repository.baseCommit,
+      baseTree: codeFixture.task.repository.baseTree,
+      candidateCommit: 'e'.repeat(codeFixture.task.repository.baseCommit.length),
+      candidateTree,
+      patch: {
+        format: 'git-diff-binary',
+        sha256: embeddedCandidateArtifact(patch).sha256,
+        byteLength: patch.byteLength,
+      },
+    }
+    const codeBundle = sealAgentCandidateBundle({
+      ...alignedBundle(codeBundleInput, profile),
+      code: {
+        kind: 'git-patch',
+        repository: { kind: 'github', owner: 'owner', repo: 'repo' },
+        baseCommit: codeSurface.baseCommit,
+        baseTree: codeSurface.baseTree,
+        candidateTree: codeSurface.candidateTree,
+        patch: { format: 'git-diff-binary', artifact: embeddedCandidateArtifact(patch) },
+      },
+    })
+    const codeResult = {
+      ...proposed.improvement.raw,
+      winner: { ...proposed.improvement.raw.winner, surface: codeSurface },
+    }
+    const codeComparison = createAgentImprovementMeasuredComparison({
+      result: codeResult,
+      measuredSurface: 'code',
+      baselineProfile: profile,
+      candidateBundle: codeBundle,
+    })
+    const codeProposal = createAgentImprovementProposal({
+      runId: 'analysis-run-code',
+      baselineProfile: profile,
+      findings: proposed.analysis.analystResult.findings,
+      evaluation: codeComparison,
+      candidateBundle: codeBundle,
+      now: () => new Date('2026-07-10T01:30:00.000Z'),
+    })
+    expect(codeProposal.changedSurfaces).toEqual(['code'])
+    expect(
+      reviewAgentImprovementProposal(codeProposal, {
+        decision: 'approve',
+        reviewedBy: 'operator@example.com',
+        reason: 'The measured code winner matches the sealed patch.',
+      }).decision,
+    ).toBe('approve')
+    expect(() =>
+      createAgentImprovementMeasuredComparison({
+        result: {
+          ...codeResult,
+          winner: {
+            ...codeResult.winner,
+            surface: { ...codeSurface, candidateTree: '0'.repeat(candidateTree.length) },
+          },
+        },
+        measuredSurface: 'code',
+        baselineProfile: profile,
+        candidateBundle: codeBundle,
+      }),
+    ).toThrow(/does not match the measured code winner/)
+    expect(() =>
+      createAgentImprovementMeasuredComparison({
+        result: {
+          ...proposed.improvement.raw,
+          winner: { ...proposed.improvement.raw.winner, surface: 'measured memory' },
+        },
+        measuredSurface: 'memory',
+        baselineProfile: profile,
+        candidateBundle: proposed.proposal.candidateBundle,
+      }),
+    ).toThrow(/memory improvement proposals require a content-addressed bundle binding/)
+    expect(() =>
+      createAgentImprovementMeasuredComparison({
+        result: {
           ...proposed.improvement.raw,
           winner: {
             ...proposed.improvement.raw.winner,
-            surface: '{not-json',
+            surface: '# Measured skill document',
           },
         },
+        measuredSurface: 'skills',
+        baselineProfile: profile,
+        candidateBundle: proposed.proposal.candidateBundle,
       }),
-    ).toThrow(/not valid JSON/)
+    ).toThrow(/skill-document improvement proposals require a content-addressed bundle binding/)
 
     const review = reviewAgentImprovementProposal(proposed.proposal, {
       decision: 'approve',
@@ -281,6 +475,9 @@ describe('agent improvement lifecycle', () => {
       now: () => new Date('2026-07-10T02:00:00.000Z'),
     })
     expect(verifyAgentImprovementReview(review)).toEqual(review)
+    expect(() =>
+      verifyAgentImprovementReview({ ...review, reason: 'Tampered after review.' }),
+    ).toThrow(/review digest does not match/)
 
     const traceStore = new InMemoryTraceStore()
     let request: AgentCandidateExecutorRequest | undefined
@@ -297,8 +494,8 @@ describe('agent improvement lifecycle', () => {
         })
         return { executionId: input.executionId, termination: { kind: 'exit', exitCode: 0 } }
       },
-      stopAndCapture: async () => ({
-        stopped: true,
+      stop: async () => ({ stopped: true }),
+      capture: async () => ({
         taskOutcome: unchangedTaskOutcomeCapture(fixture),
       }),
     }
@@ -319,14 +516,51 @@ describe('agent improvement lifecycle', () => {
 
     expect(request).toBeDefined()
     expect(executed.finalization.succeeded).toBe(true)
+    if (!executed.finalization.succeeded) throw new Error('expected successful finalization')
     expect(executed.evidence).toMatchObject({
       proposalDigest: proposed.proposal.digest,
       reviewDigest: review.digest,
-      bundleDigest: proposed.proposal.candidateBundle?.digest,
       executionId: fixture.task.executionId,
       succeeded: true,
-      runReceiptDigest: expect.stringMatching(/^sha256:/),
+      receipt: {
+        bundleDigest: proposed.proposal.candidateBundle?.digest,
+        executionPlanDigest: expect.stringMatching(/^sha256:/),
+        materializationReceiptDigest: expect.stringMatching(/^sha256:/),
+        digest: expect.stringMatching(/^sha256:/),
+      },
+      materializationReceipt: {
+        digest: executed.finalization.receipt.value.materializationReceiptDigest,
+        profilePlan: { material: { files: expect.any(Array) } },
+      },
     })
+    expect(
+      verifyCandidateExecutionEvidence([executed.evidence], {
+        proposal: proposed.proposal,
+        review,
+        expectedCount: 1,
+      }),
+    ).toEqual([executed.evidence])
+    expect(() =>
+      verifyCandidateExecutionEvidence([executed.evidence, executed.evidence], {
+        proposal: proposed.proposal,
+        review,
+        expectedCount: 2,
+      }),
+    ).toThrow(/reuses an execution id/)
+    expect(() =>
+      verifyCandidateExecutionEvidence([{ ...executed.evidence, succeeded: false }], {
+        proposal: proposed.proposal,
+        review,
+        expectedCount: 1,
+      }),
+    ).toThrow()
+    expect(() =>
+      verifyCandidateExecutionEvidence([forgeProfileActivation(executed.evidence)], {
+        proposal: proposed.proposal,
+        review,
+        expectedCount: 1,
+      }),
+    ).toThrow(/profile activation file path, mode, and content must match the canonical plan/)
   })
 
   it('records rejection and refuses to execute it', async () => {
@@ -365,7 +599,8 @@ describe('agent improvement lifecycle', () => {
             execute: async () => {
               throw new Error('must not execute')
             },
-            stopAndCapture: async () => ({ stopped: true }),
+            stop: async () => ({ stopped: true }),
+            capture: async () => ({}),
           },
           traceStore: new InMemoryTraceStore(),
           claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
@@ -435,7 +670,8 @@ describe('agent improvement lifecycle', () => {
             execute: async () => {
               throw new Error('must not execute')
             },
-            stopAndCapture: async () => ({ stopped: true }),
+            stop: async () => ({ stopped: true }),
+            capture: async () => ({}),
           },
           traceStore: new InMemoryTraceStore(),
           claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
@@ -468,3 +704,23 @@ describe('agent improvement lifecycle', () => {
     ).rejects.toThrow(/judge/i)
   })
 })
+
+function forgeProfileActivation(evidence: CandidateExecutionEvidence): CandidateExecutionEvidence {
+  const { digest: _activationDigest, ...activationWithoutDigest } = evidence.profileActivation
+  const first = evidence.profileActivation.files[0]
+  if (!first) throw new Error('expected a materialized profile file')
+  const profileActivation = canonicalCandidateDocument<
+    CandidateExecutionEvidence['profileActivation']
+  >({
+    ...activationWithoutDigest,
+    files: [
+      { ...first, content: `${first.content}\nforged` },
+      ...evidence.profileActivation.files.slice(1),
+    ],
+  }).value
+  const { digest: _evidenceDigest, ...evidenceWithoutDigest } = evidence
+  return canonicalCandidateDocument<CandidateExecutionEvidence>({
+    ...evidenceWithoutDigest,
+    profileActivation,
+  }).value
+}
diff --git a/tests/knowledge-improvement-job.test.ts b/tests/knowledge-improvement-job.test.ts
index 9e120534..80d88aae 100644
--- a/tests/knowledge-improvement-job.test.ts
+++ b/tests/knowledge-improvement-job.test.ts
@@ -1,16 +1,37 @@
-import { mkdir, mkdtemp, readFile, rm, writeFile } from 'node:fs/promises'
+import { mkdir, mkdtemp, readdir, readFile, rm, stat, writeFile } from 'node:fs/promises'
 import { tmpdir } from 'node:os'
-import { dirname, join } from 'node:path'
+import { dirname, join, relative } from 'node:path'
+import type {
+  AgentCandidateBundle,
+  AgentCandidateKnowledge,
+  AgentImprovementMeasuredComparison,
+  AgentProfile,
+} from '@tangle-network/agent-interface'
 import {
   addSourceText,
   defineReadinessSpec,
+  hashKnowledgeBase,
   initKnowledgeBase,
+  knowledgeImprovementCandidateRef,
+  withKnowledgeImprovementCandidate,
 } from '@tangle-network/agent-knowledge'
 import { describe, expect, it } from 'vitest'
+import { canonicalCandidateDigest } from '../src/candidate-execution/digest'
+import { agentCandidateProfileAsAgentProfile } from '../src/candidate-execution/profile'
+import {
+  createAgentImprovementProposal,
+  reviewAgentImprovementProposal,
+} from '../src/intelligence/improvement-cycle'
 import { runKnowledgeImprovementJob } from '../src/knowledge'
 import type { SuperviseOptions } from '../src/runtime/supervise/supervise'
 import type { SupervisorProfile } from '../src/runtime/supervise/supervisor-agent'
 import type { SupervisedResult } from '../src/runtime/supervise/types'
+import {
+  candidateBundle,
+  candidateSha,
+  createCandidateOutputFixture,
+  redigestCandidateBundle,
+} from './helpers/candidate-execution-fixture'
 
 async function withKb(fn: (root: string) => Promise<void>): Promise<void> {
   const root = await mkdtemp(join(tmpdir(), 'agent-runtime-knowledge-job-'))
@@ -64,10 +85,135 @@ async function writeRuntimeJobPage(root: string): Promise<void> {
   )
 }
 
+async function liveKnowledgeBytes(root: string): Promise<Record<string, string>> {
+  const paths = [join(root, 'knowledge'), join(root, 'raw')]
+  const sourceRegistry = join(root, '.agent-knowledge', 'sources.json')
+  const output: Record<string, string> = {}
+  for (const path of paths) await collectFiles(root, path, output)
+  try {
+    output[relative(root, sourceRegistry)] = (await readFile(sourceRegistry)).toString('base64')
+  } catch (error) {
+    if (!isMissingFile(error)) throw error
+  }
+  return output
+}
+
+async function collectFiles(
+  root: string,
+  path: string,
+  output: Record<string, string>,
+): Promise<void> {
+  let info: Awaited<ReturnType<typeof stat>>
+  try {
+    info = await stat(path)
+  } catch (error) {
+    if (isMissingFile(error)) return
+    throw error
+  }
+  if (info.isFile()) {
+    output[relative(root, path)] = (await readFile(path)).toString('base64')
+    return
+  }
+  for (const entry of (await readdir(path)).sort()) {
+    await collectFiles(root, join(path, entry), output)
+  }
+}
+
+function isMissingFile(error: unknown): boolean {
+  return error instanceof Error && 'code' in error && error.code === 'ENOENT'
+}
+
+function measuredKnowledgeComparison(
+  baselineProfile: AgentProfile,
+  bundle: AgentCandidateBundle,
+): AgentImprovementMeasuredComparison {
+  const confidenceInterval = {
+    level: 0.95,
+    lower: 0,
+    upper: 0,
+    method: 'paired-bootstrap' as const,
+    statistic: 'mean' as const,
+    resamples: 2_000,
+  }
+  return {
+    schemaVersion: 1,
+    kind: 'agent-improvement-measured-comparison',
+    benchmark: { name: 'development', version: '1', splitDigest: candidateSha('f') },
+    baselineProfileDigest: canonicalCandidateDigest(baselineProfile),
+    candidateBundleDigest: bundle.digest,
+    overall: {
+      name: 'composite',
+      baseline: 0,
+      candidate: 1,
+      delta: 1,
+      confidenceInterval: { ...confidenceInterval, lower: 1, upper: 1 },
+      n: 1,
+      direction: 'higher-is-better',
+      unit: 'score',
+    },
+    objectives: [
+      {
+        availability: 'measured',
+        kind: 'objective',
+        name: 'knowledge-readiness',
+        baseline: 0,
+        candidate: 1,
+        delta: 1,
+        confidenceInterval: { ...confidenceInterval, lower: 1, upper: 1 },
+        n: 1,
+        direction: 'higher-is-better',
+        unit: 'score',
+      },
+      {
+        availability: 'unavailable',
+        kind: 'cost',
+        name: 'cost',
+        direction: 'lower-is-better',
+        unit: 'usd',
+        reason: 'This deterministic fixture does not measure cost.',
+      },
+      {
+        availability: 'unavailable',
+        kind: 'latency',
+        name: 'latency',
+        direction: 'lower-is-better',
+        unit: 'milliseconds',
+        reason: 'This deterministic fixture does not measure latency.',
+      },
+    ],
+    candidate: { label: 'frozen knowledge candidate' },
+    decision: {
+      outcome: 'ship',
+      reasons: ['The paired held-out comparison approved this exact knowledge candidate.'],
+      contributingChecks: [{ name: 'heldout', passed: true }],
+    },
+    power: {
+      sufficient: true,
+      n: 1,
+      minimumDetectableDelta: 0,
+      confidenceLevel: 0.95,
+      scaleAssumed: true,
+      sharedScorerChannel: false,
+      reason: 'Fixture measurement is deterministic.',
+    },
+    provenance: {
+      kind: 'agent-eval-loop',
+      schema: '1.0.0',
+      runId: 'knowledge-heldout',
+      recordDigest: candidateSha('1'),
+      baselineContentHash: candidateSha('2'),
+      candidateContentHash: candidateSha('3'),
+    },
+    diff: 'knowledge candidate bytes changed',
+    evaluation: { generationsExplored: 1, durationMs: 1, totalCostUsd: 0 },
+  }
+}
+
 describe('runKnowledgeImprovementJob', () => {
-  it('runs agent-knowledge improvement through a supervised runtime updater', async () => {
+  it('leaves the live knowledge base byte-identical until approval', async () => {
     await withKb(async (root) => {
       const measurements: unknown[] = []
+      const before = await liveKnowledgeBytes(root)
       let captured:
         | {
             profile: SupervisorProfile
@@ -113,13 +259,30 @@ describe('runKnowledgeImprovementJob', () => {
         onMeasurement: (measurement) => measurements.push(measurement),
       })
 
-      expect(result.promoted).toBe(true)
-      expect(result.improvement.state.status).toBe('promoted')
+      expect(result.promoted).toBe(false)
+      expect(result.improvement.state.status).toBe('candidate-ready')
       expect(captured?.profile.name).toBe('knowledge-research-supervisor')
       expect(captured?.task).toContain('Goal: Add runtime job knowledge')
       expect(captured?.task).toContain('Knowledge base root:')
-      await expect(readFile(join(root, 'knowledge', 'runtime-job.md'), 'utf8')).resolves.toContain(
-        'candidate workspace',
+      await expect(
+        readFile(join(root, 'knowledge', 'runtime-job.md'), 'utf8'),
+      ).rejects.toMatchObject({
+        code: 'ENOENT',
+      })
+      expect(await liveKnowledgeBytes(root)).toEqual(before)
+      expect(result.candidateKnowledge?.candidate.candidateId).toBe(
+        result.improvement.candidate?.candidateId,
+      )
+      await withKnowledgeImprovementCandidate(
+        { root, candidate: knowledgeImprovementCandidateRef(result.improvement) },
+        async ({ root: candidateRoot }) => {
+          expect(result.candidateKnowledge?.candidate.candidateHash).toBe(
+            `sha256:${await hashKnowledgeBase(candidateRoot)}`,
+          )
+        },
+      )
+      expect(result.candidateKnowledge?.snapshot.material.files).toEqual(
+        expect.arrayContaining([expect.objectContaining({ path: 'knowledge/runtime-job.md' })]),
       )
       expect(result.measurement.updateCalls).toBe(1)
       expect(result.measurement.supervisedSpent).toMatchObject({
@@ -159,10 +322,127 @@ describe('runKnowledgeImprovementJob', () => {
         },
       })
 
-      expect(result.promoted).toBe(true)
-      await expect(readFile(join(root, 'knowledge', 'runtime-job.md'), 'utf8')).resolves.toContain(
-        'source-backed evidence',
+      expect(result.promoted).toBe(false)
+      await withKnowledgeImprovementCandidate(
+        { root, candidate: knowledgeImprovementCandidateRef(result.improvement) },
+        async ({ root: candidateRoot }) => {
+          await expect(
+            readFile(join(candidateRoot, 'knowledge', 'runtime-job.md'), 'utf8'),
+          ).resolves.toContain('source-backed evidence')
+        },
+      )
+      await expect(
+        readFile(join(root, 'knowledge', 'runtime-job.md'), 'utf8'),
+      ).rejects.toMatchObject({
+        code: 'ENOENT',
+      })
+    })
+  })
+
+  it('promotes only the frozen candidate bytes after an exact approved review', async () => {
+    await withKb(async (root) => {
+      const artifacts = createCandidateOutputFixture().outputArtifacts
+      const update = async (_profile: SupervisorProfile, task: unknown, opts: SuperviseOptions) => {
+        await writeRuntimeJobPage(rootFromTask(task))
+        await expect(opts.deliverable?.check({})).resolves.toBe(true)
+        return winner()
+      }
+      const proposed = await runKnowledgeImprovementJob({
+        root,
+        goal: 'Add runtime job knowledge',
+        runId: 'runtime-job-approved',
+        strict: true,
+        budget: { maxIterations: 2, maxTokens: 1000 },
+        readinessCheck: async ({ root: candidateRoot }) => ({
+          ready: await readFile(join(candidateRoot, 'knowledge', 'runtime-job.md'), 'utf8')
+            .then((text) => text.includes('source-backed evidence'))
+            .catch(() => false),
+        }),
+        runSupervised: update,
+        candidateArtifacts: artifacts,
+      })
+      const knowledge = proposed.candidateKnowledge
+      if (!knowledge) throw new Error('expected frozen knowledge candidate')
+      const candidateRef = knowledgeImprovementCandidateRef(proposed.improvement)
+      const candidateBytes = await withKnowledgeImprovementCandidate(
+        { root, candidate: candidateRef },
+        ({ root: candidateRoot }) => liveKnowledgeBytes(candidateRoot),
+      )
+      const liveBeforeApproval = await liveKnowledgeBytes(root)
+      const baseBundle = candidateBundle()
+      const baselineProfile = agentCandidateProfileAsAgentProfile(baseBundle.profile)
+      const bundle = redigestCandidateBundle(baseBundle, { knowledge })
+      const proposal = createAgentImprovementProposal({
+        runId: 'runtime-job-approved',
+        baselineProfile,
+        findings: [],
+        evaluation: measuredKnowledgeComparison(baselineProfile, bundle),
+        candidateBundle: bundle,
+        now: () => new Date('2026-07-13T01:00:00.000Z'),
+      })
+      const mismatchedKnowledge: AgentCandidateKnowledge = {
+        ...knowledge,
+        candidate: { ...knowledge.candidate, candidateHash: candidateSha('0') },
+      }
+      const mismatchedBundle = redigestCandidateBundle(bundle, {
+        knowledge: mismatchedKnowledge,
+      })
+      expect(() =>
+        createAgentImprovementProposal({
+          runId: 'runtime-job-mismatched',
+          baselineProfile,
+          findings: [],
+          evaluation: measuredKnowledgeComparison(baselineProfile, bundle),
+          candidateBundle: mismatchedBundle,
+        }),
+      ).toThrow(/measured comparison does not bind the exact candidate bundle/)
+      const review = reviewAgentImprovementProposal(proposal, {
+        decision: 'approve',
+        reviewedBy: 'operator@example.com',
+        reason: 'Approve the exact frozen knowledge candidate.',
+        now: () => new Date('2026-07-13T01:01:00.000Z'),
+      })
+
+      expect(await liveKnowledgeBytes(root)).toEqual(liveBeforeApproval)
+      const approvedUpdate = async () => {
+        throw new Error('candidate-ready promotion must not rerun the updater')
+      }
+      const promoteApproved = () =>
+        runKnowledgeImprovementJob({
+          root,
+          goal: 'Add runtime job knowledge',
+          runId: 'runtime-job-approved',
+          strict: true,
+          budget: { maxIterations: 2, maxTokens: 1000 },
+          readinessCheck: async () => ({ ready: true }),
+          runSupervised: approvedUpdate,
+          candidateArtifacts: artifacts,
+          approval: {
+            proposal,
+            review,
+            authorizeReview: async (candidateReview) => candidateReview.digest === review.digest,
+          },
+        })
+
+      await withKnowledgeImprovementCandidate(
+        { root, candidate: candidateRef },
+        async ({ root: candidateRoot }) => {
+          const frozenPage = join(candidateRoot, 'knowledge', 'runtime-job.md')
+          const frozenPageBytes = await readFile(frozenPage)
+          await writeFile(frozenPage, 'tampered frozen snapshot', 'utf8')
+          await expect(promoteApproved()).rejects.toThrow(/snapshot changed after approval/)
+          expect(await liveKnowledgeBytes(root)).toEqual(liveBeforeApproval)
+          await writeFile(frozenPage, frozenPageBytes)
+        },
       )
+
+      const promoted = await promoteApproved()
+
+      expect(promoted.promoted).toBe(true)
+      expect(promoted.improvement.state.status).toBe('promoted')
+      expect(promoted.measurement.updateCalls).toBe(0)
+      expect(await liveKnowledgeBytes(root)).toEqual(candidateBytes)
+      expect(`sha256:${await hashKnowledgeBase(root)}`).toBe(knowledge.candidate.candidateHash)
     })
   })
 })
diff --git a/tests/sandbox-approved-candidate.test.ts b/tests/sandbox-approved-candidate.test.ts
new file mode 100644
index 00000000..8a54abc3
--- /dev/null
+++ b/tests/sandbox-approved-candidate.test.ts
@@ -0,0 +1,377 @@
+import { InMemoryTraceStore } from '@tangle-network/agent-eval'
+import type {
+  AgentCandidateBundle,
+  AgentImprovementMeasuredComparison,
+  AgentProfile,
+} from '@tangle-network/agent-interface'
+import type { Process, ProcessStatus, SandboxInstance } from '@tangle-network/sandbox'
+import { afterEach, describe, expect, it, vi } from 'vitest'
+
+import { InMemoryAgentCandidateExecutionClaimStore } from '../src/candidate-execution/claim'
+import { canonicalCandidateDigest } from '../src/candidate-execution/digest'
+import type { AgentCandidateExecutorRequest } from '../src/candidate-execution/types'
+import {
+  createAgentImprovementProposal,
+  reviewAgentImprovementProposal,
+} from '../src/intelligence/improvement-cycle'
+import {
+  createSandboxApprovedCandidateExecutor,
+  sandboxApprovedCandidateExecutionSupport,
+} from '../src/intelligence/sandbox-approved-candidate'
+import {
+  candidateSha,
+  cleanupCandidateFixtures,
+  createCandidateOutputExecutionFixture,
+  createCandidateOutputFixture,
+} from './helpers/candidate-execution-fixture'
+
+afterEach(cleanupCandidateFixtures)
+
+describe('Sandbox approved candidate executor', () => {
+  it('runs one exact bounded-output task in a fresh strict-egress sandbox', async () => {
+    const fixture = createCandidateOutputExecutionFixture('application/json', 1_024)
+    const baselineProfile = { name: 'baseline' }
+    const proposal = createAgentImprovementProposal({
+      runId: 'sandbox-proposal-1',
+      baselineProfile,
+      findings: [],
+      evaluation: measuredProfileResult(baselineProfile, fixture.bundle),
+      candidateBundle: fixture.bundle,
+      now: () => new Date('2026-07-13T12:00:00.000Z'),
+    })
+    const review = reviewAgentImprovementProposal(proposal, {
+      decision: 'approve',
+      reviewedBy: 'operator@example.com',
+      reason: 'Held-back comparison passed.',
+      now: () => new Date('2026-07-13T12:01:00.000Z'),
+    })
+    const status: ProcessStatus = {
+      pid: 42,
+      command: 'codex',
+      cwd: '/workspace/task',
+      running: false,
+      exitCode: 0,
+      startedAt: new Date('2026-07-13T12:02:00.000Z'),
+      exitedAt: new Date('2026-07-13T12:02:01.000Z'),
+    }
+    const process = {
+      wait: vi.fn(async () => 0),
+      status: vi.fn(async () => status),
+      stdout: async function* () {
+        yield '{"accepted":true}\n'
+      },
+      kill: vi.fn(async () => undefined),
+    } as unknown as Process
+    const writes: Array<{ path: string; content: string; mode: number }> = []
+    const spawnExact = vi.fn(async () => {
+      spawned = true
+      return process
+    })
+    let spawned = false
+    let deleted = false
+    const sandbox = {
+      id: 'sandbox-candidate-1',
+      metadata: undefined,
+      fs: {
+        write: vi.fn(async (path: string, content: string, options: { mode: number }) => {
+          writes.push({ path, content, mode: options.mode })
+        }),
+      },
+      process: {
+        spawnExact,
+        list: vi.fn(async () => (spawned ? [status] : [])),
+        get: vi.fn(async () => process),
+      },
+      delete: vi.fn(async () => {
+        deleted = true
+      }),
+    } as unknown as SandboxInstance
+    const create = vi.fn(async () => sandbox)
+    const outputs = createCandidateOutputFixture()
+    const adapter = createSandboxApprovedCandidateExecutor({
+      client: {
+        create,
+        get: vi.fn(async () => (deleted ? null : sandbox)),
+        list: vi.fn(async () => (deleted ? [] : [sandbox])),
+      },
+      ports: fixture.ports,
+      ...outputs,
+      traceStore: new InMemoryTraceStore(),
+      claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
+      authorizeReview: async (candidateReview) => candidateReview.digest === review.digest,
+    })
+
+    const evidence = await adapter.execute({ proposal, review, task: fixture.task })
+
+    expect(create).toHaveBeenCalledWith(
+      expect.objectContaining({
+        image: `ghcr.io/example/task@sha256:${'b'.repeat(64)}`,
+        publicEdge: false,
+        ephemeral: true,
+        bare: true,
+        egressPolicy: {
+          mode: 'strict',
+          allowDomains: ['router.tangle.tools'],
+          includeImplicitDomains: false,
+        },
+      }),
+      expect.objectContaining({ signal: expect.any(AbortSignal), timeoutMs: 120_000 }),
+    )
+    expect(spawnExact).toHaveBeenCalledOnce()
+    const [executable, args, launch] = spawnExact.mock.calls[0]!
+    expect(executable).toBe('codex')
+    expect(args).toEqual(expect.any(Array))
+    expect(launch).toMatchObject({
+      cwd: '/workspace/task',
+      stdin: fixture.task.instruction,
+      timeoutMs: fixture.task.limits.timeoutMs,
+    })
+    expect(launch.env).not.toHaveProperty('PATH')
+    expect(launch.env).toHaveProperty('MODEL_GATEWAY_TOKEN', 'protected')
+    expect(writes.length).toBeGreaterThan(0)
+    expect(deleted).toBe(true)
+    expect(evidence).toMatchObject({
+      proposalDigest: proposal.digest,
+      reviewDigest: review.digest,
+      succeeded: true,
+      materializationReceipt: {
+        profilePlan: { material: { files: expect.any(Array) } },
+      },
+      profileActivation: {
+        profilePlan: { digest: evidence.materializationReceipt.profilePlan.digest },
+        files: expect.any(Array),
+      },
+      receipt: {
+        executorCapture: expect.objectContaining({
+          sha256: expect.stringMatching(/^sha256:/),
+        }),
+        taskOutcome: {
+          material: {
+            outcome: {
+              kind: 'output',
+              spec: { mediaType: 'application/json', maxBytes: 1_024 },
+            },
+          },
+        },
+      },
+    })
+    for (const file of evidence.profileActivation.files) {
+      const written = writes.find((entry) => entry.path === `/workspace/task/${file.path}`)
+      expect(written?.mode).toBe(file.mode)
+      expect(Buffer.from(written?.content ?? '', 'base64').toString('utf8')).toBe(file.content)
+    }
+    await expect(
+      adapter.executor.stop(
+        {
+          executionId: evidence.executionId,
+          executionPlanDigest: evidence.receipt.executionPlanDigest,
+        },
+        {
+          traceStore: new InMemoryTraceStore(),
+          reason: 'completed',
+          signal: new AbortController().signal,
+          deadlineAtMs: Date.now() + 1_000,
+        },
+      ),
+    ).resolves.toEqual({ stopped: true })
+  })
+
+  it('declares and enforces its bounded capability instead of claiming universality', async () => {
+    expect(sandboxApprovedCandidateExecutionSupport).toMatchObject({
+      outcomes: ['output'],
+      code: ['disabled'],
+      memory: ['disabled'],
+      knowledge: false,
+      profile: {
+        mcpTransports: ['stdio'],
+        remoteMcp: false,
+        tools: false,
+        permissions: false,
+        modes: false,
+        confidential: false,
+      },
+    })
+    const adapter = createSandboxApprovedCandidateExecutor({
+      client: {
+        create: vi.fn(async () => {
+          throw new Error('unsupported requests must fail before sandbox creation')
+        }),
+        get: vi.fn(async () => null),
+        list: vi.fn(async () => []),
+      },
+      ports: {} as never,
+      grader: {} as never,
+      outputArtifacts: {} as never,
+      traceStore: new InMemoryTraceStore(),
+      claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
+      authorizeReview: async () => true,
+    })
+    const request = minimalExecutorRequest()
+
+    await expect(
+      adapter.executor.execute(
+        withRequest(request, { taskOutcome: { kind: 'workspace' } }),
+        {} as never,
+      ),
+    ).rejects.toThrow(/workspace outcomes; use Pier/)
+    await expect(
+      adapter.executor.execute(withRequest(request, { codeKind: 'no-op' }), {} as never),
+    ).rejects.toThrow(/code workspaces; use Pier/)
+    await expect(
+      adapter.executor.execute(withRequest(request, { memoryMode: 'isolated' }), {} as never),
+    ).rejects.toThrow(/isolated memory/)
+    await expect(
+      adapter.executor.execute(
+        withRequest(request, { mediaType: 'application/octet-stream' }),
+        {} as never,
+      ),
+    ).rejects.toThrow(/only UTF-8 text and JSON/)
+  })
+})
+
+function measuredProfileResult(
+  baselineProfile: AgentProfile,
+  bundle: AgentCandidateBundle,
+): AgentImprovementMeasuredComparison {
+  return {
+    schemaVersion: 1,
+    kind: 'agent-improvement-measured-comparison',
+    benchmark: { name: 'development', version: '1', splitDigest: candidateSha('f') },
+    baselineProfileDigest: canonicalCandidateDigest(baselineProfile),
+    candidateBundleDigest: bundle.digest,
+    overall: {
+      name: 'composite',
+      baseline: 0,
+      candidate: 1,
+      delta: 1,
+      confidenceInterval: {
+        level: 0.95,
+        lower: 1,
+        upper: 1,
+        method: 'paired-bootstrap',
+        statistic: 'mean',
+        resamples: 2_000,
+      },
+      n: 3,
+      direction: 'higher-is-better',
+      unit: 'score',
+    },
+    objectives: [
+      {
+        kind: 'objective',
+        name: 'profile-quality',
+        availability: 'measured',
+        baseline: 0,
+        candidate: 1,
+        delta: 1,
+        confidenceInterval: {
+          level: 0.95,
+          lower: 1,
+          upper: 1,
+          method: 'paired-bootstrap',
+          statistic: 'mean',
+          resamples: 2_000,
+        },
+        n: 3,
+        direction: 'higher-is-better',
+        unit: 'score',
+      },
+      measuredResourceObjective('cost', 'usd'),
+      measuredResourceObjective('latency', 'milliseconds'),
+    ],
+    candidate: { label: 'measured profile' },
+    decision: {
+      outcome: 'ship',
+      reasons: ['paired heldout interval cleared the promotion threshold'],
+      contributingChecks: [{ name: 'heldout', passed: true }],
+    },
+    power: {
+      sufficient: true,
+      n: 3,
+      minimumDetectableDelta: 0.5,
+      confidenceLevel: 0.95,
+      scaleAssumed: true,
+      sharedScorerChannel: true,
+      reason: 'paired sample can detect the measured effect',
+    },
+    provenance: {
+      kind: 'agent-eval-loop',
+      schema: 'tangle.loop-provenance.v2',
+      runId: 'heldout-1',
+      recordDigest: candidateSha('1'),
+      baselineContentHash: candidateSha('2'),
+      candidateContentHash: candidateSha('3'),
+    },
+    diff: 'measured profile replacement',
+    evaluation: { generationsExplored: 1, durationMs: 1, totalCostUsd: 0 },
+  }
+}
+
+function measuredResourceObjective(
+  kind: 'cost' | 'latency',
+  unit: 'usd' | 'milliseconds',
+): AgentImprovementMeasuredComparison['objectives'][number] {
+  return {
+    kind,
+    name: kind,
+    availability: 'unavailable',
+    reason: `${kind} was not captured by this fixture`,
+    direction: 'lower-is-better' as const,
+    unit,
+  }
+}
+
+function minimalExecutorRequest(): AgentCandidateExecutorRequest {
+  return {
+    executionPlan: {
+      value: {
+        material: {
+          task: { outcome: { kind: 'output', mediaType: 'text/plain', maxBytes: 10 } },
+          codeKind: 'disabled',
+        },
+      },
+    },
+    inputs: {},
+    memory: { mode: 'disabled' },
+  } as unknown as AgentCandidateExecutorRequest
+}
+
+function withRequest(
+  request: AgentCandidateExecutorRequest,
+  change: {
+    taskOutcome?: { kind: 'workspace' }
+    codeKind?: 'no-op'
+    memoryMode?: 'isolated'
+    mediaType?: string
+  },
+): AgentCandidateExecutorRequest {
+  const material = request.executionPlan.value.material
+  return {
+    ...request,
+    executionPlan: {
+      ...request.executionPlan,
+      value: {
+        ...request.executionPlan.value,
+        material: {
+          ...material,
+          task: {
+            ...material.task,
+            outcome:
+              change.taskOutcome ??
+              (material.task.outcome.kind === 'output'
+                ? {
+                    ...material.task.outcome,
+                    mediaType: change.mediaType ?? material.task.outcome.mediaType,
+                  }
+                : material.task.outcome),
+          },
+          codeKind: change.codeKind ?? material.codeKind,
+        },
+      },
+    },
+    memory:
+      change.memoryMode === 'isolated'
+        ? ({ mode: 'isolated' } as AgentCandidateExecutorRequest['memory'])
+        : request.memory,
+  }
+}

From 9a2553ddc9a69799e41ac6d6ab7a37fa85b75951 Mon Sep 17 00:00:00 2001
From: Drew Stone <drewstone329@gmail.com>
Date: Mon, 13 Jul 2026 23:17:09 -0600
Subject: [PATCH 2/8] feat(candidate): adopt exact improvement contracts

---
 CLAUDE.md                                     |  14 +-
 docs/api/index.md                             | 246 +++++++++---------
 docs/api/intelligence.md                      |  92 +++----
 docs/api/mcp.md                               |  34 +--
 docs/api/primitive-catalog.md                 |  16 +-
 docs/canonical-api.md                         |   2 +-
 package.json                                  |  11 +-
 pnpm-lock.yaml                                |  95 ++-----
 scripts/verify-package-exports.mjs            |  58 ++++-
 src/candidate-execution/artifacts.ts          |  34 +--
 src/candidate-execution/builder.ts            |   2 +-
 src/candidate-execution/finalize.ts           |   4 +-
 src/candidate-execution/git-materialize.ts    |  18 +-
 src/candidate-execution/outcome-evidence.ts   |  16 +-
 src/candidate-execution/prepare.ts            |   6 +-
 src/candidate-execution/workspace-archive.ts  |  29 +--
 src/improvement/agentic-generator.ts          |  13 +-
 src/improvement/cleanup.ts                    |  18 ++
 src/improvement/improve.ts                    |  28 +-
 src/improvement/improvement-driver.ts         |  28 +-
 src/intelligence/improvement-cycle.ts         |  41 +--
 src/knowledge/improvement-job.ts              |  14 +-
 src/mcp/feedback-store.ts                     |   9 +-
 src/mcp/types.ts                              |   6 +-
 tests/candidate-execution-core.test.ts        |  10 +-
 tests/candidate-execution-execute.test.ts     |   6 +-
 tests/candidate-execution-finalize.test.ts    |   2 +-
 tests/candidate-execution-prepare.test.ts     |  10 +-
 .../candidate-execution-task-outcome.test.ts  |   2 +-
 tests/helpers/candidate-execution-fixture.ts  |  10 +-
 tests/knowledge-improvement-job.test.ts       |  58 +++--
 31 files changed, 466 insertions(+), 466 deletions(-)
 create mode 100644 src/improvement/cleanup.ts

diff --git a/CLAUDE.md b/CLAUDE.md
index e53da643..6d9622b0 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -41,19 +41,19 @@ This repo's bottleneck is agents paying a **re-discovery tax**: re-reading 15 fi
 
 **The anti-staleness law:** these maps are kept short and code-adjacent. If a map disagrees with the code, the **code wins** — fix the map in the *same* turn. Discovery is paid once, by whoever records it. When you learn something undocumented, write it to the map/memory before moving on. **For `docs/canonical-api.md` this is now ENFORCED, not aspirational:** the mechanical leaves (per-symbol signatures + `file:line`) are GENERATED into `docs/api/` by TypeDoc — never hand-edit them — and a CI freshness gate (`pnpm docs:check` → `scripts/check-docs-freshness.mjs`) turns a stale version pin, a broken `file:line` citation, or a decision-table symbol that no longer exists into a RED BUILD. The judgment layer (§2 decision table, when-to-use, every "Do NOT") stays hand-curated and the gate never touches it. Local fix path on a red docs gate: `pnpm run docs:api` then commit, or fix the cited line — see `docs/MAINTAINING.md`.
 
-## Repo layering — this package depends on agent-eval, never the reverse
+## Repo layering — runtime composes evaluation and knowledge
 
 ```
-agent-knowledge ──► agent-runtime (this repo — wraps the substrate)
-       │                  │
-       └──────────────────┴──► agent-eval (substrate — the bottom)
+agent-runtime (this repo)
+       ├──► agent-eval
+       └──► agent-knowledge ──► agent-eval
 ```
 
-agent-knowledge depends on agent-runtime (it imports `/loops`) and on agent-eval. agent-runtime depends ONLY on agent-eval and deliberately does NOT import agent-knowledge — domain stores/tools enter by injection (`createMcpServer({ feedbackStore })`, per `src/mcp/feedback-store.ts`), so importing the domain package would induce a dependency cycle.
+agent-runtime owns workers and orchestration. Its `/knowledge` entry point composes those workers with agent-knowledge's candidate, readiness, evaluation, and promotion operations. agent-knowledge does not import agent-runtime; runtime supplies execution through callbacks, which keeps the dependency direction acyclic.
 
-**Rule: agent-runtime depends on agent-eval. agent-eval MUST NOT import from agent-runtime.** No upward imports, no `peerDependencies` in agent-eval pointing here, no `import type { X } from '@tangle-network/agent-runtime'` inside agent-eval. A spotted upward import is a bug — file an issue and move the type into agent-eval. agent-eval is declared a **required `peerDependency`** (floor: whatever `package.json` `peerDependencies` says — do not restate the number here), not a hard dependency — keep it in sync with the `selfImprove`/`heldoutSignificance`/`loopDispatch` APIs the code uses.
+**Rule: neither agent-eval nor agent-knowledge may import agent-runtime.** No upward imports, no `peerDependencies` in either package pointing here, and no runtime types copied into those packages. agent-eval is declared a **required `peerDependency`** (floor: whatever `package.json` `peerDependencies` says — do not restate the number here), not a hard dependency — keep it in sync with the `selfImprove`/`heldoutSignificance`/`loopDispatch` APIs the code uses.
 
-**The runtime stays domain-clean by taking domain stores/tools as injected adapters, never importing a domain package** — the dependency arrow points up into this repo (agent-knowledge → agent-runtime), never down out of it.
+**Generic runtime and MCP modules stay domain-clean through injected adapters.** Only `src/knowledge/` imports agent-knowledge. Stores, tools, and product policy remain caller-owned even though runtime provides the standard knowledge-improvement composition.
 
 Substrate primitives CONSUMED from agent-eval: `DefaultVerdict`, `RunRecord`, `AgentEvalError` + taxonomy, `AnalystFinding`/`AnalystRunResult`/`FindingsDiff`, `TraceAnalystKindSpec`, `KnowledgeReadinessReport`, and the campaign types (`DispatchContext`/`ProfileDispatchFn`/`Scenario`, type-only).
 
diff --git a/docs/api/index.md b/docs/api/index.md
index 33090ff3..46effb65 100644
--- a/docs/api/index.md
+++ b/docs/api/index.md
@@ -3806,7 +3806,7 @@ Defined in: [candidate-execution/types.ts:537](https://github.com/tangle-network
 
 ### AgentCandidateWorkspaceArchiveLimits
 
-Defined in: [candidate-execution/workspace-archive.ts:52](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L52)
+Defined in: [candidate-execution/workspace-archive.ts:56](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L56)
 
 #### Properties
 
@@ -3814,49 +3814,49 @@ Defined in: [candidate-execution/workspace-archive.ts:52](https://github.com/tan
 
 > **maxArchiveBytes**: `number`
 
-Defined in: [candidate-execution/workspace-archive.ts:53](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L53)
+Defined in: [candidate-execution/workspace-archive.ts:57](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L57)
 
 ##### maxEmbeddedArtifactBytes
 
 > **maxEmbeddedArtifactBytes**: `number`
 
-Defined in: [candidate-execution/workspace-archive.ts:54](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L54)
+Defined in: [candidate-execution/workspace-archive.ts:58](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L58)
 
 ##### maxFiles
 
 > **maxFiles**: `number`
 
-Defined in: [candidate-execution/workspace-archive.ts:55](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L55)
+Defined in: [candidate-execution/workspace-archive.ts:59](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L59)
 
 ##### maxFileBytes
 
 > **maxFileBytes**: `number`
 
-Defined in: [candidate-execution/workspace-archive.ts:56](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L56)
+Defined in: [candidate-execution/workspace-archive.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L60)
 
 ##### maxTotalFileBytes
 
 > **maxTotalFileBytes**: `number`
 
-Defined in: [candidate-execution/workspace-archive.ts:57](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L57)
+Defined in: [candidate-execution/workspace-archive.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L61)
 
 ##### maxPathBytes
 
 > **maxPathBytes**: `number`
 
-Defined in: [candidate-execution/workspace-archive.ts:58](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L58)
+Defined in: [candidate-execution/workspace-archive.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L62)
 
 ##### maxRepositoryBundleBytes
 
 > **maxRepositoryBundleBytes**: `number`
 
-Defined in: [candidate-execution/workspace-archive.ts:59](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L59)
+Defined in: [candidate-execution/workspace-archive.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L63)
 
 ***
 
 ### CaptureAgentCandidateWorkspaceOptions
 
-Defined in: [candidate-execution/workspace-archive.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L98)
+Defined in: [candidate-execution/workspace-archive.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L102)
 
 #### Properties
 
@@ -3864,7 +3864,7 @@ Defined in: [candidate-execution/workspace-archive.ts:98](https://github.com/tan
 
 > `optional` **includeRepository?**: `boolean`
 
-Defined in: [candidate-execution/workspace-archive.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L100)
+Defined in: [candidate-execution/workspace-archive.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L104)
 
 Include Git HEAD so task preparation can prove its exact commit and tree.
 
@@ -3872,13 +3872,13 @@ Include Git HEAD so task preparation can prove its exact commit and tree.
 
 > `optional` **limits?**: `Partial`\<[`AgentCandidateWorkspaceArchiveLimits`](#agentcandidateworkspacearchivelimits)\>
 
-Defined in: [candidate-execution/workspace-archive.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L101)
+Defined in: [candidate-execution/workspace-archive.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L105)
 
 ##### artifactPersistence?
 
 > `optional` **artifactPersistence?**: `object`
 
-Defined in: [candidate-execution/workspace-archive.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L103)
+Defined in: [candidate-execution/workspace-archive.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L107)
 
 Use the evaluator-owned artifact store when manifest or archive bytes should not be embedded.
 
@@ -3898,7 +3898,7 @@ Use the evaluator-owned artifact store when manifest or archive bytes should not
 
 ### CreateAgentCandidateWorkspacePortOptions
 
-Defined in: [candidate-execution/workspace-archive.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L110)
+Defined in: [candidate-execution/workspace-archive.ts:114](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L114)
 
 #### Properties
 
@@ -3906,13 +3906,13 @@ Defined in: [candidate-execution/workspace-archive.ts:110](https://github.com/ta
 
 > `optional` **limits?**: `Partial`\<[`AgentCandidateWorkspaceArchiveLimits`](#agentcandidateworkspacearchivelimits)\>
 
-Defined in: [candidate-execution/workspace-archive.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L111)
+Defined in: [candidate-execution/workspace-archive.ts:115](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L115)
 
 ***
 
 ### CapturedAgentCandidateWorkspace
 
-Defined in: [candidate-execution/workspace-archive.ts:114](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L114)
+Defined in: [candidate-execution/workspace-archive.ts:118](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L118)
 
 #### Properties
 
@@ -3920,13 +3920,13 @@ Defined in: [candidate-execution/workspace-archive.ts:114](https://github.com/ta
 
 > `readonly` **snapshot**: `AgentCandidateWorkspaceSnapshotEvidence`
 
-Defined in: [candidate-execution/workspace-archive.ts:115](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L115)
+Defined in: [candidate-execution/workspace-archive.ts:119](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L119)
 
 ##### archive
 
 > `readonly` **archive**: `Uint8Array`
 
-Defined in: [candidate-execution/workspace-archive.ts:117](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L117)
+Defined in: [candidate-execution/workspace-archive.ts:121](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L121)
 
 Caller-owned bytes accepted by createAgentCandidateWorkspacePort.
 
@@ -5439,7 +5439,7 @@ Content type for the response.
 
 ### VerifyResult
 
-Defined in: [improvement/agentic-generator.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L73)
+Defined in: [improvement/agentic-generator.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L72)
 
 Outcome of verifying a candidate worktree. `feedback` (compiler errors,
  failing test output) is fed into the next shot when `ok` is false.
@@ -5450,19 +5450,19 @@ Outcome of verifying a candidate worktree. `feedback` (compiler errors,
 
 > **ok**: `boolean`
 
-Defined in: [improvement/agentic-generator.ts:74](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L74)
+Defined in: [improvement/agentic-generator.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L73)
 
 ##### feedback?
 
 > `optional` **feedback?**: `string`
 
-Defined in: [improvement/agentic-generator.ts:75](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L75)
+Defined in: [improvement/agentic-generator.ts:74](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L74)
 
 ***
 
 ### AgenticGeneratorShotReceipt
 
-Defined in: [improvement/agentic-generator.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L83)
+Defined in: [improvement/agentic-generator.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L82)
 
 `@tangle-network/agent-runtime` improvement — the CODE-surface proposer for
 agent-eval's improvement loop.
@@ -5480,25 +5480,25 @@ producer, which mutates an isolated git worktree via a pluggable
 
 > `readonly` **schemaVersion**: `1`
 
-Defined in: [improvement/agentic-generator.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L84)
+Defined in: [improvement/agentic-generator.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L83)
 
 ##### generation
 
 > `readonly` **generation**: `number` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L85)
+Defined in: [improvement/agentic-generator.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L84)
 
 ##### candidateIndex
 
 > `readonly` **candidateIndex**: `number` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L86)
+Defined in: [improvement/agentic-generator.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L85)
 
 ##### shot
 
 > `readonly` **shot**: `number`
 
-Defined in: [improvement/agentic-generator.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L88)
+Defined in: [improvement/agentic-generator.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L87)
 
 One-based shot number within this candidate.
 
@@ -5506,103 +5506,103 @@ One-based shot number within this candidate.
 
 > `readonly` **maxShots**: `number`
 
-Defined in: [improvement/agentic-generator.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L89)
+Defined in: [improvement/agentic-generator.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L88)
 
 ##### harness
 
 > `readonly` **harness**: [`LocalHarness`](mcp.md#localharness)
 
-Defined in: [improvement/agentic-generator.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L90)
+Defined in: [improvement/agentic-generator.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L89)
 
 ##### model
 
 > `readonly` **model**: `string` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L91)
+Defined in: [improvement/agentic-generator.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L90)
 
 ##### reasoningEffort
 
 > `readonly` **reasoningEffort**: `ReasoningEffort` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L92)
+Defined in: [improvement/agentic-generator.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L91)
 
 ##### promptSha256
 
 > `readonly` **promptSha256**: `` `sha256:${string}` ``
 
-Defined in: [improvement/agentic-generator.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L93)
+Defined in: [improvement/agentic-generator.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L92)
 
 ##### startedAt
 
 > `readonly` **startedAt**: `string`
 
-Defined in: [improvement/agentic-generator.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L94)
+Defined in: [improvement/agentic-generator.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L93)
 
 ##### completedAt
 
 > `readonly` **completedAt**: `string`
 
-Defined in: [improvement/agentic-generator.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L95)
+Defined in: [improvement/agentic-generator.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L94)
 
 ##### durationMs
 
 > `readonly` **durationMs**: `number`
 
-Defined in: [improvement/agentic-generator.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L96)
+Defined in: [improvement/agentic-generator.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L95)
 
 ##### exitCode
 
 > `readonly` **exitCode**: `number` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L97)
+Defined in: [improvement/agentic-generator.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L96)
 
 ##### timedOut
 
 > `readonly` **timedOut**: `boolean`
 
-Defined in: [improvement/agentic-generator.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L98)
+Defined in: [improvement/agentic-generator.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L97)
 
 ##### killedBySignal
 
 > `readonly` **killedBySignal**: `Signals` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:99](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L99)
+Defined in: [improvement/agentic-generator.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L98)
 
 ##### stdoutBytes
 
 > `readonly` **stdoutBytes**: `number` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L100)
+Defined in: [improvement/agentic-generator.ts:99](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L99)
 
 ##### stdoutSha256
 
 > `readonly` **stdoutSha256**: `` `sha256:${string}` `` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L101)
+Defined in: [improvement/agentic-generator.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L100)
 
 ##### stderrBytes
 
 > `readonly` **stderrBytes**: `number` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L102)
+Defined in: [improvement/agentic-generator.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L101)
 
 ##### stderrSha256
 
 > `readonly` **stderrSha256**: `` `sha256:${string}` `` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L103)
+Defined in: [improvement/agentic-generator.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L102)
 
 ##### usage
 
 > `readonly` **usage**: [`CodexTokenUsage`](mcp.md#codextokenusage) \| `null`
 
-Defined in: [improvement/agentic-generator.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L104)
+Defined in: [improvement/agentic-generator.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L103)
 
 ##### profileWorkspacePlanDigest
 
 > `readonly` **profileWorkspacePlanDigest**: `string` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L106)
+Defined in: [improvement/agentic-generator.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L105)
 
 Digest of the exact profile-file workspace plan applied for this shot.
 
@@ -5610,13 +5610,13 @@ Digest of the exact profile-file workspace plan applied for this shot.
 
 > `readonly` **profileWorkspaceFileCount**: `number`
 
-Defined in: [improvement/agentic-generator.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L107)
+Defined in: [improvement/agentic-generator.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L106)
 
 ##### costCallId
 
 > `readonly` **costCallId**: `string` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L109)
+Defined in: [improvement/agentic-generator.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L108)
 
 Shared run-ledger call id for this exact shot.
 
@@ -5624,7 +5624,7 @@ Shared run-ledger call id for this exact shot.
 
 > `readonly` **costBasis**: `"unknown"` \| `"provider-reported"` \| `"estimated-pricing"`
 
-Defined in: [improvement/agentic-generator.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L111)
+Defined in: [improvement/agentic-generator.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L110)
 
 Whether dollars came from the provider, the pricing table, or are unknown.
 
@@ -5632,13 +5632,13 @@ Whether dollars came from the provider, the pricing table, or are unknown.
 
 > `readonly` **costUsd**: `number` \| `null`
 
-Defined in: [improvement/agentic-generator.ts:112](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L112)
+Defined in: [improvement/agentic-generator.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L111)
 
 ##### costUsdKnown
 
 > `readonly` **costUsdKnown**: `boolean`
 
-Defined in: [improvement/agentic-generator.ts:114](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L114)
+Defined in: [improvement/agentic-generator.ts:113](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L113)
 
 True only for a provider-reported amount, never for a pricing estimate.
 
@@ -5646,19 +5646,19 @@ True only for a provider-reported amount, never for a pricing estimate.
 
 > `readonly` **evidence**: [`CodexExecutionEvidence`](mcp.md#codexexecutionevidence) \| `null`
 
-Defined in: [improvement/agentic-generator.ts:115](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L115)
+Defined in: [improvement/agentic-generator.ts:114](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L114)
 
 ##### error
 
 > `readonly` **error**: \{ `name`: `string`; `message`: `string`; \} \| `null`
 
-Defined in: [improvement/agentic-generator.ts:116](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L116)
+Defined in: [improvement/agentic-generator.ts:115](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L115)
 
 ***
 
 ### AgenticGeneratorOptions
 
-Defined in: [improvement/agentic-generator.ts:119](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L119)
+Defined in: [improvement/agentic-generator.ts:118](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L118)
 
 `@tangle-network/agent-runtime` improvement — the CODE-surface proposer for
 agent-eval's improvement loop.
@@ -5676,7 +5676,7 @@ producer, which mutates an isolated git worktree via a pluggable
 
 > `optional` **harness?**: [`LocalHarness`](mcp.md#localharness)
 
-Defined in: [improvement/agentic-generator.ts:121](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L121)
+Defined in: [improvement/agentic-generator.ts:120](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L120)
 
 Local coding harness to run in the worktree. Default `claude`.
 
@@ -5684,7 +5684,7 @@ Local coding harness to run in the worktree. Default `claude`.
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [improvement/agentic-generator.ts:124](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L124)
+Defined in: [improvement/agentic-generator.ts:123](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L123)
 
 Author profile rendered through the canonical harness mapper. Required
  for reproducible Codex so model and reasoning settings are explicit.
@@ -5693,7 +5693,7 @@ Author profile rendered through the canonical harness mapper. Required
 
 > `optional` **codexReproducible?**: `boolean`
 
-Defined in: [improvement/agentic-generator.ts:127](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L127)
+Defined in: [improvement/agentic-generator.ts:126](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L126)
 
 Run Codex with isolated configuration, exact prompt evidence, and required
  terminal token usage. Requires `harness: 'codex'` and `profile`.
@@ -5702,7 +5702,7 @@ Run Codex with isolated configuration, exact prompt evidence, and required
 
 > `optional` **codexReadDeniedPaths?**: readonly `string`[] \| ((`worktreePath`) => readonly `string`[])
 
-Defined in: [improvement/agentic-generator.ts:130](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L130)
+Defined in: [improvement/agentic-generator.ts:129](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L129)
 
 Absolute paths reproducible Codex must not read. A function can derive
  candidate-specific paths after the driver creates its worktree.
@@ -5711,7 +5711,7 @@ Absolute paths reproducible Codex must not read. A function can derive
 
 > `optional` **onShotCompleted?**: (`receipt`) => `void` \| `Promise`\<`void`\>
 
-Defined in: [improvement/agentic-generator.ts:133](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L133)
+Defined in: [improvement/agentic-generator.ts:132](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L132)
 
 Awaited once for every attempted author shot, including process failures.
  Throwing aborts the candidate so receipt persistence can fail closed.
@@ -5730,7 +5730,7 @@ Awaited once for every attempted author shot, including process failures.
 
 > `optional` **maximumCharge?**: `MaximumCharge`
 
-Defined in: [improvement/agentic-generator.ts:138](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L138)
+Defined in: [improvement/agentic-generator.ts:137](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L137)
 
 Optional hard upper bound passed to the run-wide CostLedger before each
  author shot. This MUST be enforced by the provider or executor; a planning
@@ -5741,7 +5741,7 @@ Optional hard upper bound passed to the run-wide CostLedger before each
 
 > `optional` **timeoutMs?**: `number`
 
-Defined in: [improvement/agentic-generator.ts:140](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L140)
+Defined in: [improvement/agentic-generator.ts:139](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L139)
 
 Per-shot wall-clock timeout (ms). Default = `runLocalHarness` default (5m).
 
@@ -5749,7 +5749,7 @@ Per-shot wall-clock timeout (ms). Default = `runLocalHarness` default (5m).
 
 > `optional` **buildPrompt?**: (`args`) => `string`
 
-Defined in: [improvement/agentic-generator.ts:143](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L143)
+Defined in: [improvement/agentic-generator.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L142)
 
 Build the harness task prompt from the report + findings. Override for
  domain phrasing; the default turns findings into a concrete coder task.
@@ -5774,7 +5774,7 @@ Build the harness task prompt from the report + findings. Override for
 
 > `optional` **verify?**: [`Verifier`](#verifier)
 
-Defined in: [improvement/agentic-generator.ts:149](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L149)
+Defined in: [improvement/agentic-generator.ts:148](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L148)
 
 Verify the worktree after each dirtying shot. When set, a candidate that
  fails verification is NOT returned — the failure feeds the next shot
@@ -5786,7 +5786,7 @@ Verify the worktree after each dirtying shot. When set, a candidate that
 
 > `optional` **runHarness?**: (`options`) => `Promise`\<[`LocalHarnessResult`](mcp.md#localharnessresult)\>
 
-Defined in: [improvement/agentic-generator.ts:151](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L151)
+Defined in: [improvement/agentic-generator.ts:150](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L150)
 
 Test seam — inject the harness runner (defaults to `runLocalHarness`).
 
@@ -5822,7 +5822,7 @@ Does NOT throw when:
 
 > `optional` **isDirty?**: (`worktreePath`) => `boolean`
 
-Defined in: [improvement/agentic-generator.ts:153](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L153)
+Defined in: [improvement/agentic-generator.ts:152](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L152)
 
 Test seam — inject the worktree-dirty check (defaults to `git status`).
 
@@ -5840,7 +5840,7 @@ Test seam — inject the worktree-dirty check (defaults to `git status`).
 
 ### ImproveSkillsOptions
 
-Defined in: [improvement/improve.ts:141](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L141)
+Defined in: [improvement/improve.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L142)
 
 #### Properties
 
@@ -5848,7 +5848,7 @@ Defined in: [improvement/improve.ts:141](https://github.com/tangle-network/agent
 
 > **document**: `string`
 
-Defined in: [improvement/improve.ts:143](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L143)
+Defined in: [improvement/improve.ts:144](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L144)
 
 The skill document's current text — the baseline `skillOptProposer` patches.
 
@@ -5856,7 +5856,7 @@ The skill document's current text — the baseline `skillOptProposer` patches.
 
 > `optional` **writeBack?**: (`winnerDocument`) => `void` \| `Promise`\<`void`\>
 
-Defined in: [improvement/improve.ts:147](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L147)
+Defined in: [improvement/improve.ts:148](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L148)
 
 Persist the shipped winner document (write the file the profile ref points at).
  Called only on a ship verdict. When omitted, the winner is still returned in
@@ -5876,7 +5876,7 @@ Persist the shipped winner document (write the file the profile ref points at).
 
 ### ImproveMemoryOptions
 
-Defined in: [improvement/improve.ts:150](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L150)
+Defined in: [improvement/improve.ts:151](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L151)
 
 #### Properties
 
@@ -5884,7 +5884,7 @@ Defined in: [improvement/improve.ts:150](https://github.com/tangle-network/agent
 
 > **document**: `string`
 
-Defined in: [improvement/improve.ts:152](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L152)
+Defined in: [improvement/improve.ts:153](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L153)
 
 Current durable memory text used as the measured baseline.
 
@@ -5892,7 +5892,7 @@ Current durable memory text used as the measured baseline.
 
 > `optional` **writeBack?**: (`winnerDocument`) => `void` \| `Promise`\<`void`\>
 
-Defined in: [improvement/improve.ts:154](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L154)
+Defined in: [improvement/improve.ts:155](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L155)
 
 Persist the promoted memory document. Never called on hold or error.
 
@@ -5910,7 +5910,7 @@ Persist the promoted memory document. Never called on hold or error.
 
 ### ImproveCodeOptions
 
-Defined in: [improvement/improve.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L157)
+Defined in: [improvement/improve.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L158)
 
 #### Properties
 
@@ -5918,7 +5918,7 @@ Defined in: [improvement/improve.ts:157](https://github.com/tangle-network/agent
 
 > **repoRoot**: `string`
 
-Defined in: [improvement/improve.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L159)
+Defined in: [improvement/improve.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L160)
 
 Repo root candidate worktrees fork from.
 
@@ -5926,7 +5926,7 @@ Repo root candidate worktrees fork from.
 
 > `optional` **baseRef?**: `string`
 
-Defined in: [improvement/improve.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L161)
+Defined in: [improvement/improve.ts:162](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L162)
 
 Base ref candidates fork from. Default `main`.
 
@@ -5934,7 +5934,7 @@ Base ref candidates fork from. Default `main`.
 
 > `optional` **worktreeDir?**: `string`
 
-Defined in: [improvement/improve.ts:163](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L163)
+Defined in: [improvement/improve.ts:164](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L164)
 
 Directory worktrees are created under. Default `<repoRoot>/.worktrees`.
 
@@ -5942,7 +5942,7 @@ Directory worktrees are created under. Default `<repoRoot>/.worktrees`.
 
 > `optional` **worktree?**: `WorktreeAdapter`
 
-Defined in: [improvement/improve.ts:166](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L166)
+Defined in: [improvement/improve.ts:167](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L167)
 
 Git-compatible adapter override, primarily for tests. Candidate advancement
  still requires normal Git worktree and commit semantics.
@@ -5951,7 +5951,7 @@ Git-compatible adapter override, primarily for tests. Candidate advancement
 
 > `optional` **harness?**: [`LocalHarness`](mcp.md#localharness)
 
-Defined in: [improvement/improve.ts:168](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L168)
+Defined in: [improvement/improve.ts:169](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L169)
 
 Coding harness the agentic generator runs in each worktree. Default `claude`.
 
@@ -5959,7 +5959,7 @@ Coding harness the agentic generator runs in each worktree. Default `claude`.
 
 > `optional` **verify?**: [`Verifier`](#verifier)
 
-Defined in: [improvement/improve.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L171)
+Defined in: [improvement/improve.ts:172](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L172)
 
 Verify a candidate worktree before it becomes a measurable surface; failures
  feed the next shot (see `agenticGenerator.verify` / `commandVerifier`).
@@ -5968,7 +5968,7 @@ Verify a candidate worktree before it becomes a measurable surface; failures
 
 > `optional` **timeoutMs?**: `number`
 
-Defined in: [improvement/improve.ts:173](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L173)
+Defined in: [improvement/improve.ts:174](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L174)
 
 Per-shot wall-clock timeout for the harness (ms).
 
@@ -5976,7 +5976,7 @@ Per-shot wall-clock timeout for the harness (ms).
 
 > `optional` **generator?**: [`CandidateGenerator`](#candidategenerator)
 
-Defined in: [improvement/improve.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L176)
+Defined in: [improvement/improve.ts:177](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L177)
 
 Byte-producer override — the test seam and the escape hatch for custom
  candidate production. When set, `harness`/`verify`/`timeoutMs` are unused.
@@ -5985,7 +5985,7 @@ Byte-producer override — the test seam and the escape hatch for custom
 
 ### ImproveResult
 
-Defined in: [improvement/improve.ts:179](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L179)
+Defined in: [improvement/improve.ts:180](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L180)
 
 #### Type Parameters
 
@@ -6003,7 +6003,7 @@ Defined in: [improvement/improve.ts:179](https://github.com/tangle-network/agent
 
 > **profile**: `AgentProfile`
 
-Defined in: [improvement/improve.ts:182](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L182)
+Defined in: [improvement/improve.ts:183](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L183)
 
 The profile after improvement: the winner surface applied back into the
  matching field when the gate shipped, else the input profile unchanged.
@@ -6012,7 +6012,7 @@ The profile after improvement: the winner surface applied back into the
 
 > **shipped**: `boolean`
 
-Defined in: [improvement/improve.ts:184](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L184)
+Defined in: [improvement/improve.ts:185](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L185)
 
 True when `gateDecision === 'ship'`.
 
@@ -6020,7 +6020,7 @@ True when `gateDecision === 'ship'`.
 
 > **lift**: `number`
 
-Defined in: [improvement/improve.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L186)
+Defined in: [improvement/improve.ts:187](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L187)
 
 Held-out lift (`winner − baseline` composite).
 
@@ -6028,7 +6028,7 @@ Held-out lift (`winner − baseline` composite).
 
 > **gateDecision**: `"ship"` \| `"hold"` \| `"need_more_work"` \| `"model_ceiling"` \| `"arch_ceiling"`
 
-Defined in: [improvement/improve.ts:188](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L188)
+Defined in: [improvement/improve.ts:189](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L189)
 
 The five-valued gate verdict from `selfImprove`.
 
@@ -6036,7 +6036,7 @@ The five-valued gate verdict from `selfImprove`.
 
 > **raw**: `SelfImproveResult`\<`TScenario`, `TArtifact`\>
 
-Defined in: [improvement/improve.ts:192](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L192)
+Defined in: [improvement/improve.ts:193](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L193)
 
 Full `selfImprove` result for advanced inspection. For code runs,
  `raw.winner.surface.worktreeRef` remains live after return whether the
@@ -6048,7 +6048,7 @@ Full `selfImprove` result for advanced inspection. For code runs,
 
 > **dispose**(): `Promise`\<`void`\>
 
-Defined in: [improvement/improve.ts:195](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L195)
+Defined in: [improvement/improve.ts:196](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L196)
 
 Release resources owned by this result. Idempotent; currently disposes
  the returned code worktree and is a no-op for profile-only surfaces.
@@ -6061,7 +6061,7 @@ Release resources owned by this result. Idempotent; currently disposes
 
 ### CandidateGenerator
 
-Defined in: [improvement/improvement-driver.ts:39](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L39)
+Defined in: [improvement/improvement-driver.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L41)
 
 The byte-producing seam — the ONE thing that differs between the cheap
  reflective path and the full agentic path. A generator makes (uncommitted)
@@ -6074,13 +6074,13 @@ The byte-producing seam — the ONE thing that differs between the cheap
 
 > **kind**: `string`
 
-Defined in: [improvement/improvement-driver.ts:40](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L40)
+Defined in: [improvement/improvement-driver.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L42)
 
 ##### proposesWithoutFindings?
 
 > `optional` **proposesWithoutFindings?**: `boolean`
 
-Defined in: [improvement/improvement-driver.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L51)
+Defined in: [improvement/improvement-driver.ts:53](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L53)
 
 Whether this generator can produce a candidate from an EMPTY findings set
  and no phase-2 report — i.e. it draws its change signal from the repo and
@@ -6099,7 +6099,7 @@ Whether this generator can produce a candidate from an EMPTY findings set
 
 > **generate**(`args`): `Promise`\<\{ `applied`: `boolean`; `summary`: `string`; \}\>
 
-Defined in: [improvement/improvement-driver.ts:52](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L52)
+Defined in: [improvement/improvement-driver.ts:54](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L54)
 
 ###### Parameters
 
@@ -6152,7 +6152,7 @@ Improvement-loop coordinates. Present when called through improvementDriver.
 
 ###### costLedger?
 
-`CostLedger`
+`CostLedgerHandle`
 
 Shared run-wide paid-call account supplied by agent-eval 0.117+.
 
@@ -6170,7 +6170,7 @@ Receipt attribution phase supplied alongside `costLedger`.
 
 ### ImprovementDriverOptions
 
-Defined in: [improvement/improvement-driver.ts:75](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L75)
+Defined in: [improvement/improvement-driver.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L77)
 
 #### Properties
 
@@ -6178,19 +6178,19 @@ Defined in: [improvement/improvement-driver.ts:75](https://github.com/tangle-net
 
 > **worktree**: `WorktreeAdapter`
 
-Defined in: [improvement/improvement-driver.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L76)
+Defined in: [improvement/improvement-driver.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L78)
 
 ##### generator
 
 > **generator**: [`CandidateGenerator`](#candidategenerator)
 
-Defined in: [improvement/improvement-driver.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L77)
+Defined in: [improvement/improvement-driver.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L79)
 
 ##### baseRef?
 
 > `optional` **baseRef?**: `string`
 
-Defined in: [improvement/improvement-driver.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L80)
+Defined in: [improvement/improvement-driver.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L82)
 
 Root ref for first-generation/direct callers. Default `main`.
  Later code generations retain the incumbent's original root.
@@ -6199,7 +6199,7 @@ Root ref for first-generation/direct callers. Default `main`.
 
 ### ManagedImprovementDriver
 
-Defined in: [improvement/improvement-driver.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L83)
+Defined in: [improvement/improvement-driver.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L85)
 
 #### Extends
 
@@ -6211,7 +6211,7 @@ Defined in: [improvement/improvement-driver.ts:83](https://github.com/tangle-net
 
 > **cleanup**(`retainWorktreeRefs?`): `Promise`\<`void`\>
 
-Defined in: [improvement/improvement-driver.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L85)
+Defined in: [improvement/improvement-driver.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L87)
 
 Remove every owned candidate except explicitly retained finalized winners.
 
@@ -6610,6 +6610,10 @@ Defined in: [knowledge/improvement-job.ts:83](https://github.com/tangle-network/
 
 > **outputTokens**: `number`
 
+###### usdKnown
+
+> **usdKnown**: `boolean`
+
 ###### usd
 
 > **usd**: `number`
@@ -6622,7 +6626,7 @@ Defined in: [knowledge/improvement-job.ts:83](https://github.com/tangle-network/
 
 ### KnowledgeImprovementJobResult
 
-Defined in: [knowledge/improvement-job.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L92)
+Defined in: [knowledge/improvement-job.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L93)
 
 #### Properties
 
@@ -6630,37 +6634,37 @@ Defined in: [knowledge/improvement-job.ts:92](https://github.com/tangle-network/
 
 > **improvement**: `KnowledgeImprovementResult`
 
-Defined in: [knowledge/improvement-job.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L93)
+Defined in: [knowledge/improvement-job.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L94)
 
 ##### candidateKnowledge?
 
 > `optional` **candidateKnowledge?**: `AgentCandidateKnowledge`
 
-Defined in: [knowledge/improvement-job.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L94)
+Defined in: [knowledge/improvement-job.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L95)
 
 ##### measurement
 
 > **measurement**: [`KnowledgeImprovementJobMeasurement`](#knowledgeimprovementjobmeasurement)
 
-Defined in: [knowledge/improvement-job.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L95)
+Defined in: [knowledge/improvement-job.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L96)
 
 ##### promoted
 
 > **promoted**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L96)
+Defined in: [knowledge/improvement-job.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L97)
 
 ##### blocked
 
 > **blocked**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L97)
+Defined in: [knowledge/improvement-job.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L98)
 
 ***
 
 ### AgentKnowledgeReadinessCheckOptions
 
-Defined in: [knowledge/improvement-job.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L100)
+Defined in: [knowledge/improvement-job.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L101)
 
 #### Properties
 
@@ -6668,37 +6672,37 @@ Defined in: [knowledge/improvement-job.ts:100](https://github.com/tangle-network
 
 > **goal**: `string`
 
-Defined in: [knowledge/improvement-job.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L101)
+Defined in: [knowledge/improvement-job.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L102)
 
 ##### readinessSpecs?
 
 > `optional` **readinessSpecs?**: readonly `KnowledgeReadinessSpec`[]
 
-Defined in: [knowledge/improvement-job.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L102)
+Defined in: [knowledge/improvement-job.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L103)
 
 ##### readinessTaskId?
 
 > `optional` **readinessTaskId?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L103)
+Defined in: [knowledge/improvement-job.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L104)
 
 ##### readiness?
 
 > `optional` **readiness?**: `Omit`\<`BuildEvalKnowledgeBundleOptions`, `"taskId"` \| `"index"` \| `"specs"`\>
 
-Defined in: [knowledge/improvement-job.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L104)
+Defined in: [knowledge/improvement-job.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L105)
 
 ##### strict?
 
 > `optional` **strict?**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L105)
+Defined in: [knowledge/improvement-job.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L106)
 
 ##### kbQuality?
 
 > `optional` **kbQuality?**: `KnowledgeBaseQualityOptions`
 
-Defined in: [knowledge/improvement-job.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L106)
+Defined in: [knowledge/improvement-job.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L107)
 
 ***
 
@@ -11035,7 +11039,7 @@ Defined in: [conversation/types.ts:237](https://github.com/tangle-network/agent-
 
 > **Verifier** = (`worktreePath`) => `Promise`\<[`VerifyResult`](#verifyresult)\> \| [`VerifyResult`](#verifyresult)
 
-Defined in: [improvement/agentic-generator.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L81)
+Defined in: [improvement/agentic-generator.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L80)
 
 Verifies the edited worktree. Sync or async; throws only on a setup fault
  (a candidate that fails verification returns `{ok:false}`, it does not
@@ -11057,7 +11061,7 @@ Verifies the edited worktree. Sync or async; throws only on a setup fault
 
 > **ImproveSurface** = `"prompt"` \| `"skills"` \| `"tools"` \| `"mcp"` \| `"hooks"` \| `"subagents"` \| `"agent-profile"` \| `"memory"` \| `"code"`
 
-Defined in: [improvement/improve.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L69)
+Defined in: [improvement/improve.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L70)
 
 The executable agent lever `improve` optimizes. Profile fields remain
  portable AgentProfile coordinates; implementation and orchestration files
@@ -11069,7 +11073,7 @@ The executable agent lever `improve` optimizes. Profile fields remain
 
 > **ImproveOptions**\<`TScenario`, `TArtifact`\> = `Omit`\<`SelfImproveOptions`\<`TScenario`, `TArtifact`\>, `"analyzeGeneration"` \| `"baselineSurface"` \| `"findings"` \| `"gate"` \| `"proposer"`\> & `object`
 
-Defined in: [improvement/improve.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L80)
+Defined in: [improvement/improve.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L81)
 
 #### Type Declaration
 
@@ -11903,7 +11907,7 @@ Hard cap on chained gateway hops; refused beyond this. Default keeps recursion b
 
 > `const` **AGENTIC\_PROFILE\_RESOURCE\_ROOT**: `".agent-runtime-profile-resources"` = `'.agent-runtime-profile-resources'`
 
-Defined in: [improvement/agentic-generator.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L69)
+Defined in: [improvement/agentic-generator.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L68)
 
 Dedicated ephemeral root for generic author-profile files. Every declared
 file must live below this root so cleanup cannot alter candidate-owned files.
@@ -12499,7 +12503,7 @@ Verifies every digest, resource, workspace, and Git object in a candidate bundle
 
 > **captureAgentCandidateWorkspace**(`rootInput`, `options?`): `Promise`\<[`CapturedAgentCandidateWorkspace`](#capturedagentcandidateworkspace)\>
 
-Defined in: [candidate-execution/workspace-archive.ts:121](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L121)
+Defined in: [candidate-execution/workspace-archive.ts:125](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L125)
 
 Capture one exact regular-file workspace for immutable candidate execution.
 
@@ -12523,7 +12527,7 @@ Capture one exact regular-file workspace for immutable candidate execution.
 
 > **captureAgentCandidateWorkspaceFiles**(`input`, `options?`): `Promise`\<[`CapturedAgentCandidateWorkspace`](#capturedagentcandidateworkspace)\>
 
-Defined in: [candidate-execution/workspace-archive.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L142)
+Defined in: [candidate-execution/workspace-archive.ts:146](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L146)
 
 Capture detached files returned by a remote executor into the standard archive.
 
@@ -12547,7 +12551,7 @@ readonly [`AgentCandidateExecutorWorkspaceFile`](#agentcandidateexecutorworkspac
 
 > **createAgentCandidateWorkspacePort**(`options?`): [`AgentCandidateWorkspacePort`](#agentcandidateworkspaceport)
 
-Defined in: [candidate-execution/workspace-archive.ts:204](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L204)
+Defined in: [candidate-execution/workspace-archive.ts:208](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/workspace-archive.ts#L208)
 
 Create the standard bounded materializer for candidate execution ports.
 
@@ -13052,7 +13056,7 @@ Wire integration:
 
 > **agenticGenerator**(`opts?`): [`CandidateGenerator`](#candidategenerator)
 
-Defined in: [improvement/agentic-generator.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L157)
+Defined in: [improvement/agentic-generator.ts:156](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L156)
 
 Full-agentic `CandidateGenerator` (the `shots=N, sandbox=on` setting): run a real coding harness inside the candidate worktree so the agent makes the change in place.
 
@@ -13072,7 +13076,7 @@ Full-agentic `CandidateGenerator` (the `shots=N, sandbox=on` setting): run a rea
 
 > **commandVerifier**(`command`, `args?`, `timeoutMs?`): [`Verifier`](#verifier)
 
-Defined in: [improvement/agentic-generator.ts:693](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L693)
+Defined in: [improvement/agentic-generator.ts:692](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/agentic-generator.ts#L692)
 
 A `Verifier` that runs a command in the worktree: exit 0 ⇒ ok, any other
  exit ⇒ failed with stdout+stderr as feedback. The common case — verify by
@@ -13145,7 +13149,7 @@ Build the starting instruction for a coder agent tasked with implementing a new
 
 > **applyImprovementWinnerToProfile**(`profile`, `surface`, `winner`): `AgentProfile`
 
-Defined in: [improvement/improve.ts:466](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L466)
+Defined in: [improvement/improve.ts:444](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L444)
 
 Apply a promoted winner surface back into the profile field for `surface`.
  Returns a shallow copy; never mutates the input profile.
@@ -13174,7 +13178,7 @@ Apply a promoted winner surface back into the profile field for `surface`.
 
 > **improve**\<`TScenario`, `TArtifact`\>(`profile`, `findings`, `opts`): `Promise`\<[`ImproveResult`](#improveresult)\<`TScenario`, `TArtifact`\>\>
 
-Defined in: [improvement/improve.ts:529](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L529)
+Defined in: [improvement/improve.ts:507](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L507)
 
 Run the held-out-gated self-improvement loop on ONE profile surface.
 
@@ -13226,7 +13230,7 @@ Optimize the system prompt, default holdout gate:
 
 > **improvementDriver**(`opts`): [`ManagedImprovementDriver`](#managedimprovementdriver)
 
-Defined in: [improvement/improvement-driver.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L89)
+Defined in: [improvement/improvement-driver.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improvement-driver.ts#L91)
 
 The one reflective/agentic improvement proposer (`SurfaceProposer`): owns the candidate worktree lifecycle and delegates HOW a change is produced to a pluggable `CandidateGenerator`.
 
@@ -13356,7 +13360,7 @@ Cheap no-sandbox `CandidateGenerator` (the `shots=1` setting): draft surface edi
 
 > **createAgentKnowledgeReadinessCheck**(`options`): [`KnowledgeReadinessCheck`](#knowledgereadinesscheck)
 
-Defined in: [knowledge/improvement-job.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L110)
+Defined in: [knowledge/improvement-job.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L111)
 
 Build the default readiness check backed by `@tangle-network/agent-knowledge` validation and scoring.
 
@@ -13376,7 +13380,7 @@ Build the default readiness check backed by `@tangle-network/agent-knowledge` va
 
 > **runKnowledgeImprovementJob**(`options`): `Promise`\<[`KnowledgeImprovementJobResult`](#knowledgeimprovementjobresult)\>
 
-Defined in: [knowledge/improvement-job.ts:141](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L141)
+Defined in: [knowledge/improvement-job.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L142)
 
 Produce a frozen KB candidate, and promote it only when an exact signed review is supplied.
 
diff --git a/docs/api/intelligence.md b/docs/api/intelligence.md
index a7040331..fabd8250 100644
--- a/docs/api/intelligence.md
+++ b/docs/api/intelligence.md
@@ -1078,7 +1078,7 @@ Intelligence-class spend ceiling. `0` refuses every intelligence spawn; `null` u
 
 ### ProposeAgentImprovementOptions
 
-Defined in: [intelligence/improvement-cycle.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L84)
+Defined in: [intelligence/improvement-cycle.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L85)
 
 #### Type Parameters
 
@@ -1096,31 +1096,31 @@ Defined in: [intelligence/improvement-cycle.ts:84](https://github.com/tangle-net
 
 > **runId**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L85)
+Defined in: [intelligence/improvement-cycle.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L86)
 
 ##### profile
 
 > **profile**: `AgentProfile`
 
-Defined in: [intelligence/improvement-cycle.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L86)
+Defined in: [intelligence/improvement-cycle.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L87)
 
 ##### analysis
 
 > **analysis**: `Omit`\<[`RunAnalystLoopOpts`](analyst-loop.md#runanalystloopopts), `"runId"` \| `"improvementAdapter"` \| `"autoApply"`\>
 
-Defined in: [intelligence/improvement-cycle.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L87)
+Defined in: [intelligence/improvement-cycle.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L88)
 
 ##### improvement
 
 > **improvement**: [`ImproveOptions`](index.md#improveoptions)\<`TScenario`, `TArtifact`\>
 
-Defined in: [intelligence/improvement-cycle.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L88)
+Defined in: [intelligence/improvement-cycle.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L89)
 
 ##### buildCandidate
 
 > **buildCandidate**: (`input`) => `AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput) \| `Promise`\<`AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)\>
 
-Defined in: [intelligence/improvement-cycle.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L90)
+Defined in: [intelligence/improvement-cycle.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L91)
 
 Freezes the measured winner into the exact bundle reviewed for execution.
 
@@ -1144,7 +1144,7 @@ Freezes the measured winner into the exact bundle reviewed for execution.
 
 > `optional` **now?**: () => `Date`
 
-Defined in: [intelligence/improvement-cycle.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L97)
+Defined in: [intelligence/improvement-cycle.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L98)
 
 ###### Returns
 
@@ -1154,7 +1154,7 @@ Defined in: [intelligence/improvement-cycle.ts:97](https://github.com/tangle-net
 
 ### ProposeAgentImprovementResult
 
-Defined in: [intelligence/improvement-cycle.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L100)
+Defined in: [intelligence/improvement-cycle.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L101)
 
 #### Type Parameters
 
@@ -1172,25 +1172,25 @@ Defined in: [intelligence/improvement-cycle.ts:100](https://github.com/tangle-ne
 
 > **analysis**: [`RunAnalystLoopResult`](analyst-loop.md#runanalystloopresult)
 
-Defined in: [intelligence/improvement-cycle.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L101)
+Defined in: [intelligence/improvement-cycle.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L102)
 
 ##### improvement
 
 > **improvement**: [`ImproveResult`](index.md#improveresult)\<`TScenario`, `TArtifact`\>
 
-Defined in: [intelligence/improvement-cycle.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L102)
+Defined in: [intelligence/improvement-cycle.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L103)
 
 ##### proposal
 
 > **proposal**: `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L103)
+Defined in: [intelligence/improvement-cycle.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L104)
 
 ***
 
 ### CreateAgentImprovementProposalOptions
 
-Defined in: [intelligence/improvement-cycle.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L106)
+Defined in: [intelligence/improvement-cycle.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L107)
 
 #### Properties
 
@@ -1198,37 +1198,37 @@ Defined in: [intelligence/improvement-cycle.ts:106](https://github.com/tangle-ne
 
 > **runId**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L107)
+Defined in: [intelligence/improvement-cycle.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L108)
 
 ##### baselineProfile
 
 > **baselineProfile**: `AgentProfile`
 
-Defined in: [intelligence/improvement-cycle.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L108)
+Defined in: [intelligence/improvement-cycle.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L109)
 
 ##### findings
 
 > **findings**: readonly `AnalystFinding`[]
 
-Defined in: [intelligence/improvement-cycle.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L109)
+Defined in: [intelligence/improvement-cycle.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L110)
 
 ##### evaluation
 
 > **evaluation**: `AgentImprovementMeasuredComparison`
 
-Defined in: [intelligence/improvement-cycle.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L110)
+Defined in: [intelligence/improvement-cycle.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L111)
 
 ##### candidateBundle
 
 > **candidateBundle**: `AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)
 
-Defined in: [intelligence/improvement-cycle.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L111)
+Defined in: [intelligence/improvement-cycle.ts:112](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L112)
 
 ##### now?
 
 > `optional` **now?**: () => `Date`
 
-Defined in: [intelligence/improvement-cycle.ts:112](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L112)
+Defined in: [intelligence/improvement-cycle.ts:113](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L113)
 
 ###### Returns
 
@@ -1238,7 +1238,7 @@ Defined in: [intelligence/improvement-cycle.ts:112](https://github.com/tangle-ne
 
 ### ReviewAgentImprovementInput
 
-Defined in: [intelligence/improvement-cycle.ts:125](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L125)
+Defined in: [intelligence/improvement-cycle.ts:126](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L126)
 
 #### Properties
 
@@ -1246,31 +1246,31 @@ Defined in: [intelligence/improvement-cycle.ts:125](https://github.com/tangle-ne
 
 > **decision**: `AgentImprovementReviewDecision`
 
-Defined in: [intelligence/improvement-cycle.ts:126](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L126)
+Defined in: [intelligence/improvement-cycle.ts:127](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L127)
 
 ##### reviewedBy
 
 > **reviewedBy**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:127](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L127)
+Defined in: [intelligence/improvement-cycle.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L128)
 
 ##### reason
 
 > **reason**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L128)
+Defined in: [intelligence/improvement-cycle.ts:129](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L129)
 
 ##### feedback?
 
 > `optional` **feedback?**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:129](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L129)
+Defined in: [intelligence/improvement-cycle.ts:130](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L130)
 
 ##### now?
 
 > `optional` **now?**: () => `Date`
 
-Defined in: [intelligence/improvement-cycle.ts:130](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L130)
+Defined in: [intelligence/improvement-cycle.ts:131](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L131)
 
 ###### Returns
 
@@ -1280,7 +1280,7 @@ Defined in: [intelligence/improvement-cycle.ts:130](https://github.com/tangle-ne
 
 ### ExecuteApprovedAgentCandidateOptions
 
-Defined in: [intelligence/improvement-cycle.ts:133](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L133)
+Defined in: [intelligence/improvement-cycle.ts:134](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L134)
 
 #### Properties
 
@@ -1288,19 +1288,19 @@ Defined in: [intelligence/improvement-cycle.ts:133](https://github.com/tangle-ne
 
 > **proposal**: `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:134](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L134)
+Defined in: [intelligence/improvement-cycle.ts:135](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L135)
 
 ##### review
 
 > **review**: `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:135](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L135)
+Defined in: [intelligence/improvement-cycle.ts:136](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L136)
 
 ##### authorizeReview
 
 > **authorizeReview**: (`review`, `proposal`) => `boolean` \| `Promise`\<`boolean`\>
 
-Defined in: [intelligence/improvement-cycle.ts:137](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L137)
+Defined in: [intelligence/improvement-cycle.ts:138](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L138)
 
 Product-owned authentication check for the persisted approval record.
 
@@ -1322,31 +1322,31 @@ Product-owned authentication check for the persisted approval record.
 
 > **task**: [`AgentCandidateTaskExecution`](index.md#agentcandidatetaskexecution)
 
-Defined in: [intelligence/improvement-cycle.ts:141](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L141)
+Defined in: [intelligence/improvement-cycle.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L142)
 
 ##### ports
 
 > **ports**: [`AgentCandidateExecutionPorts`](index.md#agentcandidateexecutionports)
 
-Defined in: [intelligence/improvement-cycle.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L142)
+Defined in: [intelligence/improvement-cycle.ts:143](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L143)
 
 ##### preparation?
 
 > `optional` **preparation?**: [`PrepareAgentCandidateExecutionOptions`](index.md#prepareagentcandidateexecutionoptions)
 
-Defined in: [intelligence/improvement-cycle.ts:143](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L143)
+Defined in: [intelligence/improvement-cycle.ts:144](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L144)
 
 ##### execution
 
 > **execution**: [`ExecutePreparedAgentCandidateOptions`](index.md#executepreparedagentcandidateoptions)
 
-Defined in: [intelligence/improvement-cycle.ts:144](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L144)
+Defined in: [intelligence/improvement-cycle.ts:145](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L145)
 
 ***
 
 ### VerifyCandidateExecutionEvidenceOptions
 
-Defined in: [intelligence/improvement-cycle.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L157)
+Defined in: [intelligence/improvement-cycle.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L158)
 
 #### Properties
 
@@ -1354,25 +1354,25 @@ Defined in: [intelligence/improvement-cycle.ts:157](https://github.com/tangle-ne
 
 > **proposal**: `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L158)
+Defined in: [intelligence/improvement-cycle.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L159)
 
 ##### review
 
 > **review**: `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L159)
+Defined in: [intelligence/improvement-cycle.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L160)
 
 ##### expectedCount
 
 > **expectedCount**: `number`
 
-Defined in: [intelligence/improvement-cycle.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L160)
+Defined in: [intelligence/improvement-cycle.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L161)
 
 ##### resolvedResources?
 
 > `optional` **resolvedResources?**: `ReadonlyMap`\<`` `sha256:${string}` ``, `string`\>
 
-Defined in: [intelligence/improvement-cycle.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L161)
+Defined in: [intelligence/improvement-cycle.ts:162](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L162)
 
 ***
 
@@ -3158,7 +3158,7 @@ Per-field overrides applied on top of a tier preset. Any subset of the
 
 > **ExecuteApprovedAgentCandidateResult** = \{ `finalization`: `Extract`\<[`AgentCandidateRunFinalization`](index.md#agentcandidaterunfinalization), \{ `succeeded`: `false`; \}\>; `evidence?`: `never`; \} \| \{ `finalization`: `Extract`\<[`AgentCandidateRunFinalization`](index.md#agentcandidaterunfinalization), \{ `succeeded`: `true`; \}\>; `evidence`: `CandidateExecutionEvidence`; \}
 
-Defined in: [intelligence/improvement-cycle.ts:147](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L147)
+Defined in: [intelligence/improvement-cycle.ts:148](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L148)
 
 ***
 
@@ -3558,7 +3558,7 @@ compile to `withAnalyst: true`, the tier's `fanout`, and `withLoops: true`.
 
 > **proposeAgentImprovement**\<`TScenario`, `TArtifact`\>(`options`): `Promise`\<[`ProposeAgentImprovementResult`](#proposeagentimprovementresult)\<`TScenario`, `TArtifact`\>\>
 
-Defined in: [intelligence/improvement-cycle.ts:165](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L165)
+Defined in: [intelligence/improvement-cycle.ts:166](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L166)
 
 Analyze one run and freeze its measured winner into one exact proposal.
 
@@ -3588,7 +3588,7 @@ Analyze one run and freeze its measured winner into one exact proposal.
 
 > **createAgentImprovementProposal**(`options`): `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:213](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L213)
+Defined in: [intelligence/improvement-cycle.ts:218](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L218)
 
 Freeze an already-measured improvement into the one reviewable proposal
 contract. Products that run analysis or evaluation in separate workers use
@@ -3610,7 +3610,7 @@ this constructor instead of rerunning either phase or rebuilding digests.
 
 > **reviewAgentImprovementProposal**(`inputProposal`, `input`): `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:252](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L252)
+Defined in: [intelligence/improvement-cycle.ts:257](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L257)
 
 Persist an approve/reject/change-request decision bound to one exact proposal.
 
@@ -3634,7 +3634,7 @@ Persist an approve/reject/change-request decision bound to one exact proposal.
 
 > **executeApprovedAgentCandidate**(`options`): `Promise`\<[`ExecuteApprovedAgentCandidateResult`](#executeapprovedagentcandidateresult)\>
 
-Defined in: [intelligence/improvement-cycle.ts:281](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L281)
+Defined in: [intelligence/improvement-cycle.ts:286](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L286)
 
 Verify, materialize, run, grade, and receipt only the exact approved bundle.
 
@@ -3654,7 +3654,7 @@ Verify, materialize, run, grade, and receipt only the exact approved bundle.
 
 > **verifyCandidateExecutionEvidence**(`input`, `options`): `CandidateExecutionEvidence`[]
 
-Defined in: [intelligence/improvement-cycle.ts:331](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L331)
+Defined in: [intelligence/improvement-cycle.ts:336](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L336)
 
 Verify approval, receipts, uniqueness, and the exact native profile files executed.
 
@@ -3678,7 +3678,7 @@ Verify approval, receipts, uniqueness, and the exact native profile files execut
 
 > **verifyAgentImprovementProposal**(`input`): `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:434](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L434)
+Defined in: [intelligence/improvement-cycle.ts:439](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L439)
 
 Validate a proposal's schema, profile, sealed bundle, and canonical digest.
 
@@ -3698,7 +3698,7 @@ Validate a proposal's schema, profile, sealed bundle, and canonical digest.
 
 > **verifyAgentImprovementReview**(`input`): `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:668](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L668)
+Defined in: [intelligence/improvement-cycle.ts:673](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L673)
 
 Validate a review's decision fields and canonical digest.
 
@@ -3718,7 +3718,7 @@ Validate a review's decision fields and canonical digest.
 
 > **createAgentImprovementMeasuredComparison**\<`TScenario`, `TArtifact`\>(`options`): `AgentImprovementMeasuredComparison`
 
-Defined in: [intelligence/improvement-cycle.ts:674](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L674)
+Defined in: [intelligence/improvement-cycle.ts:679](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L679)
 
 Convert agent-eval's paired result into the portable Interface comparison.
 
diff --git a/docs/api/mcp.md b/docs/api/mcp.md
index 4a72964c..d5324ded 100644
--- a/docs/api/mcp.md
+++ b/docs/api/mcp.md
@@ -432,7 +432,7 @@ readonly `string`[]
 
 ### InMemoryFeedbackStore
 
-Defined in: [mcp/feedback-store.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L42)
+Defined in: [mcp/feedback-store.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L41)
 
 **`Experimental`**
 
@@ -460,7 +460,7 @@ In-memory `FeedbackStore` — suitable for single-process use and tests.
 
 > **put**(`event`): `Promise`\<`void`\>
 
-Defined in: [mcp/feedback-store.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L45)
+Defined in: [mcp/feedback-store.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L44)
 
 **`Experimental`**
 
@@ -484,7 +484,7 @@ Append a new event. Never dedupes — every rating is its own event.
 
 > **list**(`filter?`): `Promise`\<[`FeedbackEvent`](#feedbackevent)[]\>
 
-Defined in: [mcp/feedback-store.ts:49](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L49)
+Defined in: [mcp/feedback-store.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L48)
 
 **`Experimental`**
 
@@ -2081,7 +2081,7 @@ machineId so workers don't compete with the orchestrator on the same VM.
 
 ### FeedbackEvent
 
-Defined in: [mcp/feedback-store.ts:21](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L21)
+Defined in: [mcp/feedback-store.ts:20](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L20)
 
 **`Experimental`**
 
@@ -2091,7 +2091,7 @@ Defined in: [mcp/feedback-store.ts:21](https://github.com/tangle-network/agent-r
 
 > **id**: `string`
 
-Defined in: [mcp/feedback-store.ts:22](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L22)
+Defined in: [mcp/feedback-store.ts:21](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L21)
 
 **`Experimental`**
 
@@ -2099,7 +2099,7 @@ Defined in: [mcp/feedback-store.ts:22](https://github.com/tangle-network/agent-r
 
 > **refersTo**: [`FeedbackRefersTo`](#feedbackrefersto)
 
-Defined in: [mcp/feedback-store.ts:23](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L23)
+Defined in: [mcp/feedback-store.ts:22](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L22)
 
 **`Experimental`**
 
@@ -2107,7 +2107,7 @@ Defined in: [mcp/feedback-store.ts:23](https://github.com/tangle-network/agent-r
 
 > **rating**: [`FeedbackRating`](#feedbackrating)
 
-Defined in: [mcp/feedback-store.ts:24](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L24)
+Defined in: [mcp/feedback-store.ts:23](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L23)
 
 **`Experimental`**
 
@@ -2115,7 +2115,7 @@ Defined in: [mcp/feedback-store.ts:24](https://github.com/tangle-network/agent-r
 
 > **by**: `"agent"` \| `"user"` \| `"downstream-judge"`
 
-Defined in: [mcp/feedback-store.ts:25](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L25)
+Defined in: [mcp/feedback-store.ts:24](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L24)
 
 **`Experimental`**
 
@@ -2123,7 +2123,7 @@ Defined in: [mcp/feedback-store.ts:25](https://github.com/tangle-network/agent-r
 
 > **capturedAt**: `string`
 
-Defined in: [mcp/feedback-store.ts:26](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L26)
+Defined in: [mcp/feedback-store.ts:25](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L25)
 
 **`Experimental`**
 
@@ -2131,7 +2131,7 @@ Defined in: [mcp/feedback-store.ts:26](https://github.com/tangle-network/agent-r
 
 > `optional` **namespace?**: `string`
 
-Defined in: [mcp/feedback-store.ts:27](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L27)
+Defined in: [mcp/feedback-store.ts:26](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L26)
 
 **`Experimental`**
 
@@ -2139,7 +2139,7 @@ Defined in: [mcp/feedback-store.ts:27](https://github.com/tangle-network/agent-r
 
 ### FeedbackStore
 
-Defined in: [mcp/feedback-store.ts:31](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L31)
+Defined in: [mcp/feedback-store.ts:30](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L30)
 
 **`Experimental`**
 
@@ -2149,7 +2149,7 @@ Defined in: [mcp/feedback-store.ts:31](https://github.com/tangle-network/agent-r
 
 > **put**(`event`): `Promise`\<`void`\>
 
-Defined in: [mcp/feedback-store.ts:33](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L33)
+Defined in: [mcp/feedback-store.ts:32](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L32)
 
 **`Experimental`**
 
@@ -2169,7 +2169,7 @@ Append a new event. Never dedupes — every rating is its own event.
 
 > **list**(`filter?`): `Promise`\<[`FeedbackEvent`](#feedbackevent)[]\>
 
-Defined in: [mcp/feedback-store.ts:38](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L38)
+Defined in: [mcp/feedback-store.ts:37](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L37)
 
 **`Experimental`**
 
@@ -5503,9 +5503,9 @@ Defined in: [mcp/types.ts:235](https://github.com/tangle-network/agent-runtime/b
 
 **`Experimental`**
 
-Loose shape of a research output over the wire — the substrate cannot
-import the `ResearchOutput` type from agent-knowledge without inducing
-a dependency cycle, so the MCP layer treats it structurally.
+Provider-neutral research output carried over the MCP boundary. The MCP
+layer accepts this structural shape instead of coupling its wire contract to
+one research implementation.
 
 #### Indexable
 
@@ -7568,7 +7568,7 @@ cross-sandbox copy step.
 
 > **eventToSnapshot**(`event`): [`DelegationFeedbackSnapshot`](#delegationfeedbacksnapshot)
 
-Defined in: [mcp/feedback-store.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L67)
+Defined in: [mcp/feedback-store.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/mcp/feedback-store.ts#L66)
 
 **`Experimental`**
 
diff --git a/docs/api/primitive-catalog.md b/docs/api/primitive-catalog.md
index eb935918..6dbe9f51 100644
--- a/docs/api/primitive-catalog.md
+++ b/docs/api/primitive-catalog.md
@@ -7,7 +7,7 @@
 
 # Primitive catalog — the never-stale anti-reinvention inventory
 
-> **GENERATED** from `@tangle-network/agent-runtime@0.94.8` and `@tangle-network/agent-eval@0.117.1` by `scripts/gen-primitive-catalog.mjs`. Do NOT hand-edit — run `pnpm run docs:api`. This is the mechanical companion to the JUDGMENT in `canonical-api.md` (§2 decision table + §1.5 AgentProfile law): that doc says WHICH primitive to reach for and what NOT to build; this catalog proves WHAT exists. Per-symbol signatures + `file:line` live in the per-module pages under `docs/api/`.
+> **GENERATED** from `@tangle-network/agent-runtime@0.94.9` and `@tangle-network/agent-eval@0.118.3` by `scripts/gen-primitive-catalog.mjs`. Do NOT hand-edit — run `pnpm run docs:api`. This is the mechanical companion to the JUDGMENT in `canonical-api.md` (§2 decision table + §1.5 AgentProfile law): that doc says WHICH primitive to reach for and what NOT to build; this catalog proves WHAT exists. Per-symbol signatures + `file:line` live in the per-module pages under `docs/api/`.
 
 ## 1. agent-runtime — own public surface
 
@@ -15,7 +15,7 @@ Every subpath this package declares in `package.json` `exports`. Reach for these
 
 ### Root — task lifecycle, conversation, RSI verbs, observability
 
-Import from `@tangle-network/agent-runtime` — 336 exports.
+Import from `@tangle-network/agent-runtime` — 338 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -113,6 +113,7 @@ Import from `@tangle-network/agent-runtime` — 336 exports.
 | `verifyAgentCandidateBundle` | function | Verifies every digest, resource, workspace, and Git object in a candidate bundle. |
 | `worktreeLoopRunner` | function | `code` mode on the GENERIC recursive path: author one `AgentProfile` per harness, run them as a |
 | `AGENT_CANDIDATE_EXECUTION_SUPPORT` | const | Surfaces admitted by Runtime's verifier before an environment adapter is selected. |
+| `AGENTIC_PROFILE_RESOURCE_ROOT` | const | Dedicated ephemeral root for generic author-profile files. Every declared |
 | `CANDIDATE_TRACE_ENV` | const | Environment keys used to propagate immutable candidate trace identity. |
 | `CANDIDATE_TRACE_TAGS` | const | Protected trace tags that bind a run to one prepared candidate execution. |
 | `DEFAULT_MAX_DEPTH` | const | Hard cap on chained gateway hops; refused beyond this. Default keeps recursion bounded. |
@@ -211,7 +212,7 @@ Import from `@tangle-network/agent-runtime` — 336 exports.
 | `VerifiedAgentCandidateTaskOutcome` | type | Branded task outcome that has survived independent evaluator verification. |
 | `Verifier` | type | Verifies the edited worktree. Sync or async; throws only on a setup fault |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentAdapter`, `AgentBackendContext`, `AgentBackendInput`, `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `AgentExecutionBackend`, `AgenticGeneratorOptions`, `AgentKnowledgeProvider`, `AgentKnowledgeReadinessCheckOptions`, `AgentProfileDiffProposal`, `AgentTaskContext`, `AgentTaskRunResult`, `AgentTaskSpec`, `ApprovedKnowledgeImprovementCandidate`, `BackendCallPolicy`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `ChatTurnHooks`, `ChatTurnResult`, `ControlBudget`, `ControlEvalResult`, `ControlRunResult`, `ControlStep`, `Conversation`, `ConversationDriveState`, `ConversationJournal`, `ConversationParticipant`, `ConversationPolicy`, `ConversationResult`, `ConversationTurn`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `D1StmtLike`, `DataAcquisitionPlan`, `DelegatedLoopResult`, `DisposePreparedAgentCandidateOptions`, `EvalRunEvent`, `EvalRunGeneration`, `EvalRunsExportConfig`, `EvalRunsExportResult`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `HaltContext`, `HaltSignal`, `ImproveCodeOptions`, `ImproveMemoryOptions`, `ImprovementDriverOptions`, `ImproveResult`, `ImproveSkillsOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `KnowledgeReadinessReport`, `KnowledgeRequirement`, `LoopRunnerCliArgs`, `LoopRunnerCliResult`, `ManagedImprovementDriver`, `OtelAttribute`, `OtelExporter`, `OtelSpan`, `PersonaConversationResult`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `ProfileDiffProposerOptions`, `RecoverExpiredAgentCandidateOptions`, `ResearchLoopResult`, `ResearchLoopRunnerOptions`, `ResolveAgentBackendOptions`, `ResolvedAgentCandidateContainer`, `ResolvedChatModel`, `RunChatTurnInput`, `RunConversationOptions`, `RunDelegatedLoopOptions`, `RunKnowledgeImprovementJobOptions`, `RunPersonaConfig`, `RunPersonaConversationOptions`, `RuntimeDecisionEvidenceRef`, `RuntimeDecisionPoint`, `RuntimeEventCollector`, `RuntimeEventOtelOptions`, `RuntimeHookContext`, `RuntimeHookErrorContext`, `RuntimeHookEvent`, `RuntimeRunHandle`, `RuntimeRunPersistenceAdapter`, `RuntimeRunRow`, `RuntimeSessionStore`, `RuntimeStreamEventCollector`, `RuntimeTelemetryOptions`, `RunToolLoopOptions`, `SanitizedKnowledgeReadinessReport`, `StreamToolLoopOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `ToolLoopResult`, `VerifiedAgentCandidate`, `VetoedFact`, `WorktreeLoopRunnerOptions`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`, `AgentRuntimeEvent`, `AgentRuntimeEventSink`, `AgentTaskStatus`, `AuthSource`, `ControlDecision`, `ConversationStreamEvent`, `DelegatedLoopMode`, `DelegatedLoopRegistry`, `DelegatedLoopRunner`, `ForwardHeaderName`, `HaltPredicate`, `HaltReason`, `ImproveOptions`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `ProfileDiffProposerContext`, `RuntimeDecisionKind`, `RuntimeHookTarget`, `RuntimeStreamEvent`, `StreamToolLoopYield`, `SupervisedKnowledgeUpdater`, `ToolLoopEvent`, `TurnOrder`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentAdapter`, `AgentBackendContext`, `AgentBackendInput`, `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `AgentExecutionBackend`, `AgenticGeneratorOptions`, `AgenticGeneratorShotReceipt`, `AgentKnowledgeProvider`, `AgentKnowledgeReadinessCheckOptions`, `AgentProfileDiffProposal`, `AgentTaskContext`, `AgentTaskRunResult`, `AgentTaskSpec`, `ApprovedKnowledgeImprovementCandidate`, `BackendCallPolicy`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `ChatTurnHooks`, `ChatTurnResult`, `ControlBudget`, `ControlEvalResult`, `ControlRunResult`, `ControlStep`, `Conversation`, `ConversationDriveState`, `ConversationJournal`, `ConversationParticipant`, `ConversationPolicy`, `ConversationResult`, `ConversationTurn`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `D1StmtLike`, `DataAcquisitionPlan`, `DelegatedLoopResult`, `DisposePreparedAgentCandidateOptions`, `EvalRunEvent`, `EvalRunGeneration`, `EvalRunsExportConfig`, `EvalRunsExportResult`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `HaltContext`, `HaltSignal`, `ImproveCodeOptions`, `ImproveMemoryOptions`, `ImprovementDriverOptions`, `ImproveResult`, `ImproveSkillsOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `KnowledgeReadinessReport`, `KnowledgeRequirement`, `LoopRunnerCliArgs`, `LoopRunnerCliResult`, `ManagedImprovementDriver`, `OtelAttribute`, `OtelExporter`, `OtelSpan`, `PersonaConversationResult`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `ProfileDiffProposerOptions`, `RecoverExpiredAgentCandidateOptions`, `ResearchLoopResult`, `ResearchLoopRunnerOptions`, `ResolveAgentBackendOptions`, `ResolvedAgentCandidateContainer`, `ResolvedChatModel`, `RunChatTurnInput`, `RunConversationOptions`, `RunDelegatedLoopOptions`, `RunKnowledgeImprovementJobOptions`, `RunPersonaConfig`, `RunPersonaConversationOptions`, `RuntimeDecisionEvidenceRef`, `RuntimeDecisionPoint`, `RuntimeEventCollector`, `RuntimeEventOtelOptions`, `RuntimeHookContext`, `RuntimeHookErrorContext`, `RuntimeHookEvent`, `RuntimeRunHandle`, `RuntimeRunPersistenceAdapter`, `RuntimeRunRow`, `RuntimeSessionStore`, `RuntimeStreamEventCollector`, `RuntimeTelemetryOptions`, `RunToolLoopOptions`, `SanitizedKnowledgeReadinessReport`, `StreamToolLoopOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `ToolLoopResult`, `VerifiedAgentCandidate`, `VetoedFact`, `WorktreeLoopRunnerOptions`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`, `AgentRuntimeEvent`, `AgentRuntimeEventSink`, `AgentTaskStatus`, `AuthSource`, `ControlDecision`, `ConversationStreamEvent`, `DelegatedLoopMode`, `DelegatedLoopRegistry`, `DelegatedLoopRunner`, `ForwardHeaderName`, `HaltPredicate`, `HaltReason`, `ImproveOptions`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `ProfileDiffProposerContext`, `RuntimeDecisionKind`, `RuntimeHookTarget`, `RuntimeStreamEvent`, `StreamToolLoopYield`, `SupervisedKnowledgeUpdater`, `ToolLoopEvent`, `TurnOrder`.
 
 ### Vertical agent — manifest + improvement adapter
 
@@ -1035,7 +1036,7 @@ Import from `@tangle-network/agent-runtime/mcp` — 176 exports.
 | `DetachedSessionRefParts` | interface | Decoded `DelegationRecord.detachedSessionRef`. `sandboxId` is absent between |
 | `DriveTurnCapableBox` | interface | The box surface detached turns need. `SandboxInstance` |
 | `FleetHandle` | interface | Minimal `SandboxFleet` surface the fleet executor calls. Declared |
-| `ResearchOutputShape` | interface | Loose shape of a research output over the wire — the substrate cannot |
+| `ResearchOutputShape` | interface | Provider-neutral research output carried over the MCP boundary. The MCP |
 | `SettledWorker` | interface | A worker the driver has drained via `await_event`. |
 | `TraceContext` | interface | Trace context propagation for MCP subprocess. |
 | `UiAuditorDelegationOutput` | interface | Wire-shape of a completed UI-audit delegation. The `findings` array |
@@ -1172,7 +1173,7 @@ Import from `@tangle-network/agent-eval` — 51 exports.
 
 ### CAMPAIGN — profile matrix, gates, improvement loop
 
-Import from `@tangle-network/agent-eval/campaign` — 379 exports.
+Import from `@tangle-network/agent-eval/campaign` — 380 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -1364,6 +1365,7 @@ Import from `@tangle-network/agent-eval/campaign` — 379 exports.
 | `UserStoryVerdict` | interface | A scored user story — the completion verdict plus its human title. |
 | `AxisVerdict` | type | Per-axis verdict from the good-direction paired bootstrap. |
 | `CampaignTokenUsage` | type | Token usage accumulated for a cell. Aliased to the canonical `RunTokenUsage` |
+| `CostLedgerHandle` | type | Public callback surface for a shared cost ledger. |
 | `CrossSurfaceAttemptCompleteness` | type | Whether one candidate attempt produced a usable executable outcome. |
 | `DispatchFn` | type | One function: scenario + ctx → artifact. Dispatcher chooses |
 | `FapoOptimizationLevel` | type | FAPO (Fully Autonomous Prompt Optimization) is an orchestration policy, not |
@@ -1389,8 +1391,8 @@ Import from `@tangle-network/agent-eval` — 5 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
-| `extractUsage` | function | Pull `{ input, output, cached? }` from a parsed chat-completions response |
+| `extractUsage` | function | Pull `{ input, output, cached?, cacheWrite? }` from a parsed response |
 | `extractUsageFromResponse` | function | Extract usage from an HTTP `Response` without consuming the caller's body: |
-| `extractUsageFromSse` | function | Sum token usage across an SSE response body. Each `data:` line is parsed and |
+| `extractUsageFromSse` | function | Extract token usage from a complete SSE response body using the shared SSE |
 
 **Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `LlmUsage`, `RunTokenUsage`.
diff --git a/docs/canonical-api.md b/docs/canonical-api.md
index 640eac8c..7cd54fc7 100644
--- a/docs/canonical-api.md
+++ b/docs/canonical-api.md
@@ -2,7 +2,7 @@
 
 <!-- This doc is the JUDGMENT layer: the mental model (§1), the AgentProfile law (§1.5), and the anti-reinvention decision table (§2) — WHICH primitive to reach for and what NOT to build. The export INVENTORY (WHAT exists) and per-symbol signatures + `file:line` are GENERATED into `docs/api/` (TypeDoc + `scripts/gen-primitive-catalog.mjs`, do NOT hand-edit) — that is the mechanical reference: `docs/api/primitive-catalog.md` is the never-stale list of every primitive to reuse. The freshness gate (`pnpm docs:freshness`) FAILS CI if a version pin, a cited `file:line`, a decision-table symbol, or the generated catalog drifts from source — see `docs/MAINTAINING.md`. Keep this file the small, hand-curated spine; never re-list the inventory here — point at the catalog. -->
 
-> **Version 0.94.8.** The export inventory + per-symbol signatures live in the generated `docs/api/` reference: **`docs/api/primitive-catalog.md`** is the never-stale, grouped list of every primitive to reuse (own surface + the agent-eval judge / authenticity / verification / statistics / campaign / token-usage surfaces), with each one's import path and one-line summary read live from source; the per-module pages hold the full signatures. The pinned substrate is agent-eval `>=0.117.1 <0.118.0`; the sandbox substrate that materializes profiles into harness shapes is `@tangle-network/sandbox` (peer `>=0.8.0 <1.0.0`). The neutral contract types (`AgentProfile`, `AgentProfileMcpServer`, `HarnessType`, `ReasoningEffort`, `Part`/`ToolPart`/`ToolState`, plus environment-provider types) are owned by **`@tangle-network/agent-interface`** (peer `>=0.25.0 <0.26.0`) — the single source of truth. Substrate primitives are re-exported through `@tangle-network/agent-eval/contract` (or `/campaign`), not local to this package — the catalog's §2 shows exactly which subpath each lives under.
+> **Version 0.94.9.** The export inventory + per-symbol signatures live in the generated `docs/api/` reference: **`docs/api/primitive-catalog.md`** is the never-stale, grouped list of every primitive to reuse (own surface + the agent-eval judge / authenticity / verification / statistics / campaign / token-usage surfaces), with each one's import path and one-line summary read live from source; the per-module pages hold the full signatures. The pinned substrate is agent-eval `>=0.118.3 <1.0.0`; the sandbox substrate that materializes profiles into harness shapes is `@tangle-network/sandbox` (peer `>=0.10.5 <1.0.0`). The neutral contract types (`AgentProfile`, `AgentProfileMcpServer`, `HarnessType`, `ReasoningEffort`, `Part`/`ToolPart`/`ToolState`, plus environment-provider types) are owned by **`@tangle-network/agent-interface`** (peer `>=0.26.0 <1.0.0`) — the single source of truth. Substrate primitives are re-exported through `@tangle-network/agent-eval/contract` (or `/campaign`), not local to this package — the catalog's §2 shows exactly which subpath each lives under.
 >
 > **`./loops` is the runtime barrel** — `package.json` maps it to `src/runtime/index.ts`. Everything below labelled `/loops` is the recursive-atom + loop-kernel surface.
 >
diff --git a/package.json b/package.json
index f7f9ba5e..215c5b1b 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-runtime",
-  "version": "0.94.8",
+  "version": "0.94.9",
   "description": "Shared task-lifecycle skeleton for agents: a recursive loop kernel for chat turns, one-shot tasks, and multi-attempt loops, with trace capture and eval-gated self-improvement. Domain behavior lives in adapters; scoring and ship-gates in @tangle-network/agent-eval.",
   "homepage": "https://github.com/tangle-network/agent-runtime#readme",
   "repository": {
@@ -104,8 +104,8 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.15",
-    "@tangle-network/agent-eval": "^0.115.1",
-    "@tangle-network/agent-interface": "file:/tmp/tangle-network-agent-interface-candidate.tgz",
+    "@tangle-network/agent-eval": "0.118.3",
+    "@tangle-network/agent-interface": "0.26.0",
     "@tangle-network/sandbox": "^0.10.5",
     "@types/node": "^25.9.3",
     "@types/tar-stream": "3.1.4",
@@ -122,6 +122,7 @@
     "minimumReleaseAgeExclude": [
       "@tangle-network/agent-eval",
       "@tangle-network/agent-interface",
+      "@tangle-network/agent-knowledge",
       "@tangle-network/agent-profile-materialize",
       "@tangle-network/sandbox"
     ],
@@ -135,8 +136,8 @@
   "license": "MIT",
   "packageManager": "pnpm@10.28.0",
   "peerDependencies": {
-    "@tangle-network/agent-eval": ">=0.115.1 <1.0.0",
-    "@tangle-network/agent-interface": ">=0.25.0 <1.0.0",
+    "@tangle-network/agent-eval": ">=0.118.3 <1.0.0",
+    "@tangle-network/agent-interface": ">=0.26.0 <1.0.0",
     "@tangle-network/sandbox": ">=0.10.5 <1.0.0",
     "playwright": "^1.40.0"
   },
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index e31dbae8..0f58435a 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -10,7 +10,7 @@ importers:
     dependencies:
       '@tangle-network/agent-knowledge':
         specifier: ^2.0.0
-        version: file:../tangle-network-agent-knowledge-candidate.tgz(typescript@5.9.3)
+        version: 2.0.0(typescript@5.9.3)
       '@tangle-network/agent-profile-materialize':
         specifier: 0.3.2
         version: 0.3.2
@@ -22,11 +22,11 @@ importers:
         specifier: ^2.4.15
         version: 2.4.15
       '@tangle-network/agent-eval':
-        specifier: 0.117.1
-        version: 0.117.1(typescript@5.9.3)
+        specifier: 0.118.3
+        version: 0.118.3(typescript@5.9.3)
       '@tangle-network/agent-interface':
-        specifier: file:/tmp/tangle-network-agent-interface-candidate.tgz
-        version: file:../tangle-network-agent-interface-candidate.tgz
+        specifier: 0.26.0
+        version: 0.26.0
       '@tangle-network/sandbox':
         specifier: ^0.10.5
         version: 0.10.5(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))
@@ -649,16 +649,14 @@ packages:
   '@tangle-network/agent-core@0.3.8':
     resolution: {integrity: sha512-bZfVpdiFjXbcQwSxSQABdtXEmUuapWChCcXrHkP6fAnwqEy9hWEfeZLlMZUy+1XaOfpTaMLUOEAwYhXfN018wQ==}
 
+  '@tangle-network/agent-core@0.4.11':
+    resolution: {integrity: sha512-5B1IjrJ8xDR7w8Hv/MSk2ixul6NEJQ5Ftzo+z6l7ipeFpc0yaf1+Kkml4HDUEmGW/rqdrNpBf9/MpT7i/SY0PA==}
+
   '@tangle-network/agent-core@0.4.9':
     resolution: {integrity: sha512-BFU4WV12Y6D28PX+o8LIVkIMD+cXwrL1uyV182l12Bn9zEu7VHt2oVMEEzbsN8L+X0xM9baEfMCHTFKFNJv7Aw==}
 
-  '@tangle-network/agent-eval@0.115.1':
-    resolution: {integrity: sha512-Yd487u8v0c8/+On9kaYGASS88oLhOJ5p0zwRMGdbWRSWC1UODbIgk1Fqdyq6EI60v+lZ+yDg/7LHkIrNi4kjPA==}
-    engines: {node: '>=20'}
-    hasBin: true
-
-  '@tangle-network/agent-eval@0.116.0':
-    resolution: {integrity: sha512-smOENca2M8s1JH/5vQc0vETR0LPhkZVC6tCtITDwwyof4toPPg4Rek9U5pRpykmvsGJrpiOzpVLPNBXzOVgYHQ==}
+  '@tangle-network/agent-eval@0.118.3':
+    resolution: {integrity: sha512-+HQaUHlvT5SH5jCcj7CgOfxnbOv4bqcOucE8r7ZHdo+JgaE9mRbXUTaaYe5wPSqHDUZb5SVtks70mz9olT5Lzg==}
     engines: {node: '>=20'}
     hasBin: true
 
@@ -671,22 +669,17 @@ packages:
   '@tangle-network/agent-interface@0.21.0':
     resolution: {integrity: sha512-jDxhVJgxymrvU1RLWxWKueuaWQIpBAfrW8BuVTB5m2Y4eFMLo1SawDBEMDLdZN4/Gf34xrFrsRk+PAj9brGKMQ==}
 
-  '@tangle-network/agent-interface@0.22.0':
-    resolution: {integrity: sha512-7fsJhNdvTmOB1X9E2owl06jzyrqaN+jMkOPVKbK7dvNqQvf627PowCtL/edhUqEEv+K0FWtkwG3R3xqjX7VNZg==}
-
   '@tangle-network/agent-interface@0.24.0':
     resolution: {integrity: sha512-iaHWNTYne49cBYXwb72NjGDw5bjT2KVJlfawXygbvkqnfUsQG57BS++Yjf/XYvwJJdDeEaGs+Xvy8SWYkvu0qA==}
 
   '@tangle-network/agent-interface@0.25.0':
     resolution: {integrity: sha512-bMKjyjk8bk2651HleiTEUOTPgNE2bWxCEl9VBVUyub/UkP87Y09byXOxeDEsNGc4TIT19Y5Uo0b+k1kRWXkMVQ==}
 
-  '@tangle-network/agent-interface@file:../tangle-network-agent-interface-candidate.tgz':
-    resolution: {integrity: sha512-GG3hZc/bUPQ6ssGSzPkhJcmxq2lUPllSXt5VsCvFVl2fd6wf6kBtDEv38JyUyeJEF/utTFsxNg4dNDQfEYhFJQ==, tarball: file:../tangle-network-agent-interface-candidate.tgz}
-    version: 0.24.0
+  '@tangle-network/agent-interface@0.26.0':
+    resolution: {integrity: sha512-/z4HavFr/9AbaHxi/13bFP9iSUt8oitZbw4wS9g9KAt2TeN2ec5/A2C106udWkKIETZLnu465TfU+K4KDVNkKw==}
 
-  '@tangle-network/agent-knowledge@file:../tangle-network-agent-knowledge-candidate.tgz':
-    resolution: {integrity: sha512-8qOttTET4huTNz8t4NHLl+iMuiVlbzcxFiiGmUNzbEsfsGqE8vDc4Jiu7CVEl2yqwfo8YEw6tirWCKyB8zUAQg==, tarball: file:../tangle-network-agent-knowledge-candidate.tgz}
-    version: 2.0.0
+  '@tangle-network/agent-knowledge@2.0.0':
+    resolution: {integrity: sha512-QquWbvefmeAjdntzoyIBtZzpNItfC/b8Nbpf15cSYxbmXEdpWJaxMR4tSNiLaf0Ruh4/qlNPrHGt0D4ARSB4YA==}
     engines: {node: '>=20.19.0'}
     hasBin: true
 
@@ -982,8 +975,8 @@ packages:
   graceful-fs@4.2.11:
     resolution: {integrity: sha512-RbJ5/jmFcNNCcDV5o9eTnBLJ/HszWV0P73bc+Ff4nS/rJj+YaS6IGyiOL0VoBYX+l1Wrl3k63h/KrH+nhJ0XvQ==}
 
-  hono@4.12.28:
-    resolution: {integrity: sha512-YwUvVpSF7m1yOblFPrU3Hbo8XhPheBoiyfGuII6z19LnOr6JpDnyyp7LFNrfV56wS8tpvtBFGRISHN02pDdLOA==}
+  hono@4.12.30:
+    resolution: {integrity: sha512-emn+JoJjrN9YTpRDS5it/UI2SO9BAE37T6I3d963RxcZ81G9A4pr2SZTEiiaiKbzx+NKRg5BZ89fCL7gCJCUog==}
     engines: {node: '>=16.9.0'}
 
   isows@1.0.7:
@@ -1746,35 +1739,23 @@ snapshots:
       '@tangle-network/agent-interface': 0.17.1
       zod: 4.4.3
 
-  '@tangle-network/agent-core@0.4.9':
+  '@tangle-network/agent-core@0.4.11':
     dependencies:
-      '@tangle-network/agent-interface': 0.24.0
+      '@tangle-network/agent-interface': 0.26.0
       zod: 4.4.3
 
-  '@tangle-network/agent-eval@0.115.1(typescript@5.9.3)':
+  '@tangle-network/agent-core@0.4.9':
     dependencies:
-      '@asteasolutions/zod-to-openapi': 8.5.0(zod@4.4.3)
-      '@ax-llm/ax': 19.0.45(zod@4.4.3)
-      '@hono/node-server': 2.0.8(hono@4.12.30)
-      '@tangle-network/agent-interface': 0.22.0
-      '@tangle-network/tcloud': 0.4.14(typescript@5.9.3)(zod@4.4.3)
-      hono: 4.12.30
+      '@tangle-network/agent-interface': 0.24.0
       zod: 4.4.3
-    transitivePeerDependencies:
-      - '@mastra/core'
-      - '@modelcontextprotocol/sdk'
-      - ai
-      - bufferutil
-      - openai
-      - typescript
-      - utf-8-validate
 
-  '@tangle-network/agent-eval@0.117.1(typescript@5.9.3)':
+  '@tangle-network/agent-eval@0.118.3(typescript@5.9.3)':
     dependencies:
       '@asteasolutions/zod-to-openapi': 8.5.0(zod@4.4.3)
       '@ax-llm/ax': 19.0.45(zod@4.4.3)
       '@hono/node-server': 2.0.8(hono@4.12.30)
-      '@tangle-network/agent-interface': 0.22.0
+      '@tangle-network/agent-core': 0.4.11
+      '@tangle-network/agent-interface': 0.26.0
       '@tangle-network/tcloud': 0.4.14(typescript@5.9.3)(zod@4.4.3)
       hono: 4.12.30
       zod: 4.4.3
@@ -1787,24 +1768,6 @@ snapshots:
       - typescript
       - utf-8-validate
 
-  '@tangle-network/agent-eval@0.116.0(typescript@5.9.3)':
-    dependencies:
-      '@asteasolutions/zod-to-openapi': 8.5.0(zod@4.4.3)
-      '@ax-llm/ax': 19.0.45(zod@4.4.3)
-      '@hono/node-server': 2.0.8(hono@4.12.28)
-      '@tangle-network/agent-interface': 0.22.0
-      '@tangle-network/tcloud': 0.4.14(typescript@5.9.3)(zod@4.4.3)
-      hono: 4.12.28
-      zod: 4.4.3
-    transitivePeerDependencies:
-      - '@mastra/core'
-      - '@modelcontextprotocol/sdk'
-      - ai
-      - bufferutil
-      - openai
-      - typescript
-      - utf-8-validate
-
   '@tangle-network/agent-interface@0.13.0':
     dependencies:
       zod: 4.4.3
@@ -1817,10 +1780,6 @@ snapshots:
     dependencies:
       zod: 4.4.3
 
-  '@tangle-network/agent-interface@0.22.0':
-    dependencies:
-      zod: 4.4.3
-
   '@tangle-network/agent-interface@0.24.0':
     dependencies:
       zod: 4.4.3
@@ -1829,14 +1788,14 @@ snapshots:
     dependencies:
       zod: 4.4.3
 
-  '@tangle-network/agent-interface@file:../tangle-network-agent-interface-candidate.tgz':
+  '@tangle-network/agent-interface@0.26.0':
     dependencies:
       '@noble/hashes': 1.8.0
       zod: 4.4.3
 
-  '@tangle-network/agent-knowledge@file:../tangle-network-agent-knowledge-candidate.tgz(typescript@5.9.3)':
+  '@tangle-network/agent-knowledge@2.0.0(typescript@5.9.3)':
     dependencies:
-      '@tangle-network/agent-eval': 0.116.0(typescript@5.9.3)
+      '@tangle-network/agent-eval': 0.118.3(typescript@5.9.3)
       proper-lockfile: 4.1.2
       zod: 4.4.3
     transitivePeerDependencies:
@@ -2134,7 +2093,7 @@ snapshots:
 
   graceful-fs@4.2.11: {}
 
-  hono@4.12.28: {}
+  hono@4.12.30: {}
 
   isows@1.0.7(ws@8.21.0):
     dependencies:
diff --git a/scripts/verify-package-exports.mjs b/scripts/verify-package-exports.mjs
index c28e1f85..42833acf 100644
--- a/scripts/verify-package-exports.mjs
+++ b/scripts/verify-package-exports.mjs
@@ -30,21 +30,31 @@ try {
   if (packageJson.peerDependenciesMeta?.['@tangle-network/agent-eval']?.optional) {
     throw new Error('@tangle-network/agent-eval must stay required: root and ./loops import it at runtime')
   }
-  const requiredExports = {
-    '.': ['import', 'types'],
-    './agent': ['import', 'types'],
-    './candidate-execution': ['import', 'types'],
-    './intelligence': ['import', 'types'],
-    './loops': ['import', 'types'],
-    './environment-provider': ['import', 'types'],
-    './profiles': ['import', 'types'],
-    './mcp': ['import', 'types'],
+  const packageExports = packageJson.exports
+  if (!packageExports || typeof packageExports !== 'object') {
+    throw new Error('packed package has no exports map')
   }
-
-  for (const [subpath, fields] of Object.entries(requiredExports)) {
-    const exportTarget = packageJson.exports?.[subpath]
-    if (!exportTarget) throw new Error(`missing package export ${subpath}`)
-    for (const field of fields) {
+  const requiredSubpaths = [
+    '.',
+    './agent',
+    './intelligence',
+    './loops',
+    './environment-provider',
+    './analyst-loop',
+    './lifecycle',
+    './knowledge',
+    './profiles',
+    './platform',
+    './candidate-execution',
+    './mcp',
+  ]
+  for (const subpath of requiredSubpaths) {
+    if (!(subpath in packageExports)) {
+      throw new Error(`packed package removed public export ${subpath}`)
+    }
+  }
+  for (const [subpath, exportTarget] of Object.entries(packageExports)) {
+    for (const field of ['import', 'types']) {
       const relativeTarget = exportTarget[field]
       if (typeof relativeTarget !== 'string') {
         throw new Error(`missing ${field} target for package export ${subpath}`)
@@ -194,6 +204,26 @@ try {
   run('pnpm', ['install', '--ignore-scripts', '--config.auto-install-peers=false'], appDir)
   run('pnpm', ['exec', 'tsc', '-p', 'tsconfig.json'], appDir)
 
+  run(
+    process.execPath,
+    [
+      '--input-type=module',
+      '--eval',
+      `
+        const { readFileSync } = await import('node:fs')
+        const packageJson = JSON.parse(
+          readFileSync('node_modules/@tangle-network/agent-runtime/package.json', 'utf8'),
+        )
+        for (const subpath of Object.keys(packageJson.exports)) {
+          const specifier =
+            subpath === '.' ? packageJson.name : packageJson.name + subpath.slice(1)
+          await import(specifier)
+        }
+      `,
+    ],
+    appDir,
+  )
+
   run(
     process.execPath,
     [
diff --git a/src/candidate-execution/artifacts.ts b/src/candidate-execution/artifacts.ts
index 5c8ab84e..4cbe4446 100644
--- a/src/candidate-execution/artifacts.ts
+++ b/src/candidate-execution/artifacts.ts
@@ -113,6 +113,23 @@ export async function readMaterializedWorkspaceFiles(
   )
 }
 
+export function candidateWorkspaceManifest(
+  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
+): AgentCandidateWorkspaceManifestMaterial {
+  return {
+    schemaVersion: 2,
+    kind: 'agent-candidate-workspace-manifest',
+    files: files
+      .map((file) => ({
+        path: file.path,
+        mode: file.mode,
+        sha256: sha256Bytes(file.bytes),
+        byteLength: file.bytes.byteLength,
+      }))
+      .sort((a, b) => (a.path < b.path ? -1 : a.path > b.path ? 1 : 0)),
+  }
+}
+
 function assertWorkspaceManifest(
   observed: AgentCandidateWorkspaceManifestMaterial,
   expected: AgentCandidateWorkspaceManifestMaterial,
@@ -163,7 +180,6 @@ async function scanWorkspace(
   if ((await realpath(absoluteRoot)) !== absoluteRoot) {
     throw new Error('workspace root has a symlinked path component')
   }
-  const files: AgentCandidateWorkspaceManifestMaterial['files'] = []
   const capturedFiles: Array<{ path: string; mode: number; bytes: Uint8Array }> = []
   let totalBytes = 0
 
@@ -190,7 +206,7 @@ async function scanWorkspace(
       if (!stats.isFile()) {
         throw new Error(`workspace contains a non-regular entry: ${relPath}`)
       }
-      if (limits && files.length >= limits.maxFiles) {
+      if (limits && capturedFiles.length >= limits.maxFiles) {
         throw new Error('workspace exceeds maxFiles')
       }
       const descriptor = await open(
@@ -220,12 +236,6 @@ async function scanWorkspace(
           relPath,
         )
         totalBytes += bytes.byteLength
-        files.push({
-          path: relPath,
-          mode,
-          sha256: sha256Bytes(bytes),
-          byteLength: bytes.byteLength,
-        })
         capturedFiles.push({ path: relPath, mode, bytes: Uint8Array.from(bytes) })
       } finally {
         await descriptor.close()
@@ -234,14 +244,8 @@ async function scanWorkspace(
   }
 
   await visit(absoluteRoot)
-  files.sort((a, b) => (a.path < b.path ? -1 : a.path > b.path ? 1 : 0))
-  capturedFiles.sort((a, b) => (a.path < b.path ? -1 : a.path > b.path ? 1 : 0))
   return {
-    manifest: {
-      schemaVersion: 1,
-      kind: 'agent-candidate-workspace-manifest',
-      files,
-    },
+    manifest: candidateWorkspaceManifest(capturedFiles),
     files: capturedFiles,
   }
 }
diff --git a/src/candidate-execution/builder.ts b/src/candidate-execution/builder.ts
index 4fb05fc2..6ff0bd60 100644
--- a/src/candidate-execution/builder.ts
+++ b/src/candidate-execution/builder.ts
@@ -82,7 +82,7 @@ export function buildAgentCandidateBundle(
   const compiledProfile = compileCandidateProfile(input.profile)
   const profileDiffIds = compiledProfile.profileDiffIds
   const bundle: AgentCandidateBundleInput = {
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-bundle',
     digestAlgorithm: 'rfc8785-sha256',
     profile: compiledProfile.profile,
diff --git a/src/candidate-execution/finalize.ts b/src/candidate-execution/finalize.ts
index 04142910..8fbd3365 100644
--- a/src/candidate-execution/finalize.ts
+++ b/src/candidate-execution/finalize.ts
@@ -156,7 +156,7 @@ export async function finalizeAgentCandidateRun(
       modelCallCount: modelSpans.length,
     }
     const document = canonicalCandidateDocument<AgentCandidateRunReceipt>({
-      schemaVersion: 1,
+      schemaVersion: 3,
       kind: 'agent-candidate-run',
       digestAlgorithm: 'rfc8785-sha256',
       bundleDigest: state.bundle.digest,
@@ -266,7 +266,7 @@ async function memoryReceipt(
   const afterState = capture.memoryAfter.afterState
   const manifestBytes = canonicalCandidateBytes(afterState)
   const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: afterState,
diff --git a/src/candidate-execution/git-materialize.ts b/src/candidate-execution/git-materialize.ts
index 6ce82eb5..8a530df1 100644
--- a/src/candidate-execution/git-materialize.ts
+++ b/src/candidate-execution/git-materialize.ts
@@ -10,8 +10,8 @@ import type {
   AgentCandidateWorkspaceManifestMaterial,
 } from '@tangle-network/agent-interface'
 
-import { verifyBytes } from './artifacts'
-import { canonicalCandidateBytes, sha256Bytes } from './digest'
+import { candidateWorkspaceManifest, verifyBytes } from './artifacts'
+import { canonicalCandidateBytes } from './digest'
 import type { AgentCandidateRepositoryPort } from './types'
 
 interface GitResult {
@@ -320,19 +320,9 @@ async function workspaceManifestFromGitTree(
   tree: string,
   environment: Record<string, string>,
 ): Promise<AgentCandidateWorkspaceManifestMaterial> {
-  const files = (await readCandidateGitTreeFiles(repositoryRoot, tree, environment)).map(
-    ({ path, mode, bytes }) => ({
-      path,
-      mode,
-      sha256: sha256Bytes(bytes),
-      byteLength: bytes.byteLength,
-    }),
+  return candidateWorkspaceManifest(
+    await readCandidateGitTreeFiles(repositoryRoot, tree, environment),
   )
-  return {
-    schemaVersion: 1,
-    kind: 'agent-candidate-workspace-manifest',
-    files,
-  }
 }
 
 export async function readCandidateGitTreeFiles(
diff --git a/src/candidate-execution/outcome-evidence.ts b/src/candidate-execution/outcome-evidence.ts
index 05c33973..5d7ec046 100644
--- a/src/candidate-execution/outcome-evidence.ts
+++ b/src/candidate-execution/outcome-evidence.ts
@@ -84,7 +84,7 @@ export async function persistCandidateModelSettlementEvidence(
   outputArtifacts: AgentCandidateOutputArtifactPort,
 ): Promise<PersistedAgentCandidateModelSettlement> {
   const material = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-model-settlement-material' as const,
     executionPlanDigest: identity.executionPlanDigest,
     preparationId: settlement.value.preparationId,
@@ -116,7 +116,7 @@ export async function persistCandidateModelSettlementEvidence(
   })
   return immutableCandidateValue(
     agentCandidateModelSettlementEvidenceSchema.parse({
-      schemaVersion: 1,
+      schemaVersion: 2,
       kind: 'agent-candidate-model-settlement',
       digest,
       material,
@@ -224,7 +224,7 @@ async function verifyWorkspaceTaskCapture(
   const manifestBytes = canonicalCandidateBytes(afterState)
   assertNoProtectedBytes(manifestBytes, protectedValues)
   const provisionalSnapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: afterState,
@@ -254,7 +254,7 @@ async function persistVerifiedWorkspaceTaskOutcome(
 ): Promise<VerifiedAgentCandidateTaskOutcome> {
   const { afterState, manifestBytes, patch, repository } = verified
   const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: afterState,
@@ -262,7 +262,7 @@ async function persistVerifiedWorkspaceTaskOutcome(
     archive: persisted.archive,
   })
   const material = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-task-outcome-material' as const,
     executionPlanDigest: state.executionPlan.value.digest,
     outcome: {
@@ -336,7 +336,7 @@ async function persistVerifiedOutputTaskOutcome(
     throw new Error('persisted task output changed the signed media constraints')
   }
   const material = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-task-outcome-material' as const,
     executionPlanDigest: state.executionPlan.value.digest,
     outcome: {
@@ -379,7 +379,7 @@ async function persistTaskOutcomeEvidence<Material extends AgentCandidateTaskOut
   })
   return immutableCandidateValue(
     agentCandidateTaskOutcomeEvidenceSchema.parse({
-      schemaVersion: 1,
+      schemaVersion: 2,
       kind: 'agent-candidate-task-outcome',
       digest,
       material,
@@ -501,7 +501,7 @@ async function verifyMemoryCapture(
   assertNoProtectedBytes(manifestBytes, protectedValues)
   assertNoProtectedBytes(memory.archive, protectedValues)
   const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: memory.afterState,
diff --git a/src/candidate-execution/prepare.ts b/src/candidate-execution/prepare.ts
index 845ada76..46407fb1 100644
--- a/src/candidate-execution/prepare.ts
+++ b/src/candidate-execution/prepare.ts
@@ -231,7 +231,7 @@ export async function prepareAgentCandidateExecution(
     )
     const routes = modelRoutes(bundle.profile, task.model.requested)
     const executionMaterial: AgentCandidateExecutionPlanMaterial = {
-      schemaVersion: 1,
+      schemaVersion: 2,
       kind: 'agent-candidate-execution-plan-material',
       bundleDigest: bundle.digest,
       executionId: task.executionId,
@@ -292,7 +292,7 @@ export async function prepareAgentCandidateExecution(
     }
     const executionPlan: AgentCandidateExecutionPlanEvidence =
       agentCandidateExecutionPlanEvidenceSchema.parse({
-        schemaVersion: 1,
+        schemaVersion: 2,
         kind: 'agent-candidate-execution-plan',
         digest: executionDigest,
         material: executionMaterial,
@@ -302,7 +302,7 @@ export async function prepareAgentCandidateExecution(
     const entrypoint = candidateEntrypointReceipt(candidate)
     const materializationReceipt = canonicalCandidateDocument<AgentCandidateMaterializationReceipt>(
       {
-        schemaVersion: 1,
+        schemaVersion: 2,
         kind: 'agent-candidate-materialization',
         digestAlgorithm: 'rfc8785-sha256',
         bundleDigest: bundle.digest,
diff --git a/src/candidate-execution/workspace-archive.ts b/src/candidate-execution/workspace-archive.ts
index 2975c6ac..c02c21c6 100644
--- a/src/candidate-execution/workspace-archive.ts
+++ b/src/candidate-execution/workspace-archive.ts
@@ -18,12 +18,16 @@ import { isAnyArrayBuffer, isSharedArrayBuffer } from 'node:util/types'
 
 import type {
   AgentCandidateCapturedArtifact,
-  AgentCandidateWorkspaceManifestMaterial,
   AgentCandidateWorkspaceSnapshotEvidence,
 } from '@tangle-network/agent-interface'
 import { type Entry, extract, type Pack, pack } from 'tar-stream'
 
-import { captureMaterializedWorkspace, verifyBytes, verifyMaterializedWorkspace } from './artifacts'
+import {
+  candidateWorkspaceManifest,
+  captureMaterializedWorkspace,
+  verifyBytes,
+  verifyMaterializedWorkspace,
+} from './artifacts'
 import {
   canonicalCandidateBytes,
   canonicalCandidateDigest,
@@ -158,7 +162,7 @@ async function captureWorkspaceFiles(
   limits: AgentCandidateWorkspaceArchiveLimits,
   artifactPersistence: CaptureAgentCandidateWorkspaceOptions['artifactPersistence'],
 ): Promise<CapturedAgentCandidateWorkspace> {
-  const material = workspaceManifest(files)
+  const material = candidateWorkspaceManifest(files)
   const manifest = canonicalCandidateBytes(material)
   const archive = await encodeWorkspaceArchive(files, repository, limits.maxArchiveBytes)
   if (
@@ -173,7 +177,7 @@ async function captureWorkspaceFiles(
   const manifestArtifact = await captureWorkspaceArtifact('manifest', manifest, artifactPersistence)
   const archiveArtifact = await captureWorkspaceArtifact('archive', archive, artifactPersistence)
   const snapshot = deepFreezeCandidate({
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: canonicalCandidateDigest(material),
     material,
@@ -737,26 +741,11 @@ async function materializeRepository(
   }
 }
 
-function workspaceManifest(
-  files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
-): AgentCandidateWorkspaceManifestMaterial {
-  return {
-    schemaVersion: 1,
-    kind: 'agent-candidate-workspace-manifest',
-    files: files.map((file) => ({
-      path: file.path,
-      mode: file.mode,
-      sha256: sha256Bytes(file.bytes),
-      byteLength: file.bytes.byteLength,
-    })),
-  }
-}
-
 function assertArchiveMatchesSnapshot(
   files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
   snapshot: AgentCandidateWorkspaceSnapshotEvidence,
 ): void {
-  const material = workspaceManifest(files)
+  const material = candidateWorkspaceManifest(files)
   const materialBytes = canonicalCandidateBytes(material)
   verifyBytes(
     materialBytes,
diff --git a/src/improvement/agentic-generator.ts b/src/improvement/agentic-generator.ts
index f8e4e391..703774ed 100644
--- a/src/improvement/agentic-generator.ts
+++ b/src/improvement/agentic-generator.ts
@@ -38,7 +38,6 @@ import { existsSync, readFileSync, rmSync } from 'node:fs'
 import { join, resolve, sep } from 'node:path'
 import type {
   AnalystFinding,
-  CostLedger,
   CostReceipt,
   CostReceiptInput,
   MaximumCharge,
@@ -195,11 +194,6 @@ export function agenticGenerator(opts: AgenticGeneratorOptions = {}): CandidateG
       costLedger,
       costPhase,
     }) {
-      if (opts.codexReproducible && !costLedger) {
-        throw new Error(
-          'agenticGenerator: reproducible Codex requires the run-wide CostLedger supplied by agent-eval 0.117+',
-        )
-      }
       const basePrompt = appendProfileResourcePaths(
         buildPrompt({ report, findings }),
         profileResourcePlan,
@@ -264,7 +258,12 @@ export function agenticGenerator(opts: AgenticGeneratorOptions = {}): CandidateG
           }
 
           if (opts.codexReproducible) {
-            const ledger = costLedger as CostLedger
+            const ledger = costLedger
+            if (!ledger) {
+              throw new Error(
+                'agenticGenerator: reproducible Codex requires the run-wide CostLedger supplied by agent-eval',
+              )
+            }
             const model = opts.profile?.model?.default
             if (!model) {
               throw new Error('agenticGenerator: reproducible Codex requires profile.model.default')
diff --git a/src/improvement/cleanup.ts b/src/improvement/cleanup.ts
new file mode 100644
index 00000000..bbe0d18d
--- /dev/null
+++ b/src/improvement/cleanup.ts
@@ -0,0 +1,18 @@
+export async function rethrowAfterCleanup(
+  cause: unknown,
+  cleanup: () => Promise<void>,
+  context: string,
+): Promise<never> {
+  const cleanupErrors: unknown[] = []
+  for (let attempt = 0; attempt < 2; attempt += 1) {
+    try {
+      await cleanup()
+    } catch (cleanupCause) {
+      cleanupErrors.push(cleanupCause)
+      continue
+    }
+    if (cleanupErrors.length === 0) throw cause
+    throw new AggregateError([cause, ...cleanupErrors], `${context}; cleanup retry succeeded`)
+  }
+  throw new AggregateError([cause, ...cleanupErrors], `${context}; cleanup failed`)
+}
diff --git a/src/improvement/improve.ts b/src/improvement/improve.ts
index 72a83616..945eb9b1 100644
--- a/src/improvement/improve.ts
+++ b/src/improvement/improve.ts
@@ -56,6 +56,7 @@ import { ConfigError } from '../errors'
 import type { LocalHarness } from '../mcp/local-harness'
 import { assertModelAllowed } from '../runtime/supervise/model-policy'
 import { agenticGenerator, type Verifier } from './agentic-generator'
+import { rethrowAfterCleanup } from './cleanup'
 import {
   type CandidateGenerator,
   improvementDriver,
@@ -336,29 +337,6 @@ interface PreparedCodeRun {
   cleanup(retainedWinner?: MutableSurface): Promise<void>
 }
 
-/** Preserve the primary failure while making two best-effort cleanup attempts.
- * A failed first attempt is retained in the thrown AggregateError even when the
- * retry succeeds, so callers can diagnose degraded cleanup without losing the
- * error that caused cleanup to run. */
-async function rethrowAfterCleanup(
-  cause: unknown,
-  cleanup: () => Promise<void>,
-  message: string,
-): Promise<never> {
-  const cleanupErrors: unknown[] = []
-  for (let attempt = 0; attempt < 2; attempt += 1) {
-    try {
-      await cleanup()
-    } catch (cleanupCause) {
-      cleanupErrors.push(cleanupCause)
-      continue
-    }
-    if (cleanupErrors.length === 0) throw cause
-    throw new AggregateError([cause, ...cleanupErrors], `${message}; the cleanup retry succeeded`)
-  }
-  throw new AggregateError([cause, ...cleanupErrors], message)
-}
-
 async function discardPreparedBaseline(
   worktree: WorktreeAdapter,
   baselineWorktree: Worktree,
@@ -367,7 +345,7 @@ async function discardPreparedBaseline(
   return rethrowAfterCleanup(
     cause,
     () => worktree.discard(baselineWorktree),
-    'improve(): code preparation failed and its baseline worktree could not be cleaned',
+    'improve(): code preparation failed',
   )
 }
 
@@ -616,7 +594,7 @@ export async function improve<TScenario extends Scenario, TArtifact>(
     return rethrowAfterCleanup(
       cause,
       () => preparedCode.cleanup(),
-      'improve(): code improvement failed and its worktrees could not be cleaned',
+      'improve(): code improvement failed',
     )
   }
 
diff --git a/src/improvement/improvement-driver.ts b/src/improvement/improvement-driver.ts
index b3ea5c8a..cdc3480d 100644
--- a/src/improvement/improvement-driver.ts
+++ b/src/improvement/improvement-driver.ts
@@ -21,9 +21,10 @@
  */
 
 import { spawnSync } from 'node:child_process'
-import type { AnalystFinding, CostLedger } from '@tangle-network/agent-eval'
+import type { AnalystFinding } from '@tangle-network/agent-eval'
 import {
   type CodeSurface,
+  type CostLedgerHandle,
   type LabeledScenarioStore,
   type ProposeContext,
   type SurfaceProposer,
@@ -31,6 +32,7 @@ import {
   type Worktree,
   type WorktreeAdapter,
 } from '@tangle-network/agent-eval/campaign'
+import { rethrowAfterCleanup } from './cleanup'
 
 /** The byte-producing seam — the ONE thing that differs between the cheap
  *  reflective path and the full agentic path. A generator makes (uncommitted)
@@ -66,7 +68,7 @@ export interface CandidateGenerator {
     generation?: number
     candidateIndex?: number
     /** Shared run-wide paid-call account supplied by agent-eval 0.117+. */
-    costLedger?: CostLedger
+    costLedger?: CostLedgerHandle
     /** Receipt attribution phase supplied alongside `costLedger`. */
     costPhase?: string
   }): Promise<{ applied: boolean; summary: string }>
@@ -146,24 +148,14 @@ export function improvementDriver(opts: ImprovementDriverOptions): ManagedImprov
           owned.delete(wt.path)
           owned.set(surface.worktreeRef, wt)
         } catch (err) {
-          const cleanupErrors: unknown[] = []
-          for (let attempt = 0; attempt < 2; attempt += 1) {
-            try {
+          const failure = err instanceof Error ? err.message : String(err)
+          return rethrowAfterCleanup(
+            err,
+            async () => {
               await opts.worktree.discard(wt)
               owned.delete(wt.path)
-              break
-            } catch (cause) {
-              cleanupErrors.push(cause)
-            }
-          }
-          if (cleanupErrors.length === 0) throw err
-          const failure = err instanceof Error ? err.message : String(err)
-          const cleanupSucceeded = !owned.has(wt.path)
-          throw new AggregateError(
-            [err, ...cleanupErrors],
-            cleanupSucceeded
-              ? `improvementDriver: ${failure}; candidate cleanup retry succeeded`
-              : `improvementDriver: ${failure}; candidate worktree could not be cleaned`,
+            },
+            `improvementDriver: ${failure}`,
           )
         }
       }
diff --git a/src/intelligence/improvement-cycle.ts b/src/intelligence/improvement-cycle.ts
index 051f806d..2ba4ea24 100644
--- a/src/intelligence/improvement-cycle.ts
+++ b/src/intelligence/improvement-cycle.ts
@@ -65,6 +65,7 @@ import {
   verifiedResourceTextByDigest,
   verifyAgentCandidateBundle,
 } from '../candidate-execution/verify'
+import { rethrowAfterCleanup } from '../improvement/cleanup'
 import {
   applyImprovementWinnerToProfile,
   type ImproveOptions,
@@ -183,26 +184,30 @@ export async function proposeAgentImprovement<TScenario extends Scenario, TArtif
     'proposeAgentImprovement findings',
   )
   const improvement = await improve(options.profile, [...findings], options.improvement)
-  if (!improvement.shipped) {
-    throw new Error('agent improvement proposal requires a passing measured comparison')
-  }
-  const candidateBundle = sealBuiltCandidate(
-    await options.buildCandidate({ analysis, improvement }),
-  )
-  const proposal = createAgentImprovementProposal({
-    runId: options.runId,
-    baselineProfile: options.profile,
-    findings,
-    evaluation: createAgentImprovementMeasuredComparison({
-      result: improvement.raw,
-      measuredSurface: options.improvement.surface ?? 'prompt',
+  try {
+    if (!improvement.shipped) {
+      throw new Error('agent improvement proposal requires a passing measured comparison')
+    }
+    const candidateBundle = sealBuiltCandidate(
+      await options.buildCandidate({ analysis, improvement }),
+    )
+    const proposal = createAgentImprovementProposal({
+      runId: options.runId,
       baselineProfile: options.profile,
+      findings,
+      evaluation: createAgentImprovementMeasuredComparison({
+        result: improvement.raw,
+        measuredSurface: options.improvement.surface ?? 'prompt',
+        baselineProfile: options.profile,
+        candidateBundle,
+      }),
       candidateBundle,
-    }),
-    candidateBundle,
-    ...(options.now ? { now: options.now } : {}),
-  })
-  return { analysis, improvement, proposal }
+      ...(options.now ? { now: options.now } : {}),
+    })
+    return { analysis, improvement, proposal }
+  } catch (cause) {
+    return rethrowAfterCleanup(cause, () => improvement.dispose(), 'proposeAgentImprovement failed')
+  }
 }
 
 /**
diff --git a/src/knowledge/improvement-job.ts b/src/knowledge/improvement-job.ts
index 660efbe8..0766ef78 100644
--- a/src/knowledge/improvement-job.ts
+++ b/src/knowledge/improvement-job.ts
@@ -84,6 +84,7 @@ export interface KnowledgeImprovementJobMeasurement {
     iterations: number
     inputTokens: number
     outputTokens: number
+    usdKnown: boolean
     usd: number
     ms: number
   }
@@ -386,7 +387,7 @@ function rawKnowledgeDigest(value: string, label: string): string {
 }
 
 function emptySpent(): KnowledgeImprovementJobMeasurement['supervisedSpent'] {
-  return { iterations: 0, inputTokens: 0, outputTokens: 0, usd: 0, ms: 0 }
+  return { iterations: 0, inputTokens: 0, outputTokens: 0, usdKnown: true, usd: 0, ms: 0 }
 }
 
 function addSpent(
@@ -394,9 +395,10 @@ function addSpent(
   result: SupervisedResult<unknown>,
 ): void {
   const spent = result.spentTotal
-  target.iterations += spent.iterations ?? 0
-  target.inputTokens += spent.tokens?.input ?? 0
-  target.outputTokens += spent.tokens?.output ?? 0
-  target.usd += spent.usd ?? 0
-  target.ms += spent.ms ?? 0
+  target.iterations += spent.iterations
+  target.inputTokens += spent.tokens.input
+  target.outputTokens += spent.tokens.output
+  target.usdKnown = target.usdKnown && spent.usdKnown !== false
+  target.usd += spent.usd
+  target.ms += spent.ms
 }
diff --git a/src/mcp/feedback-store.ts b/src/mcp/feedback-store.ts
index 1ebbbfa6..50f6124a 100644
--- a/src/mcp/feedback-store.ts
+++ b/src/mcp/feedback-store.ts
@@ -2,11 +2,10 @@
  *
  * Feedback persistence surface for the MCP layer.
  *
- * The substrate cannot import `@tangle-network/agent-knowledge` (it would
- * induce a dependency cycle), so the store is an abstract interface. The
- * default implementation is in-memory; consumers wire their own adapter
- * (a real KbStore-backed sink, an HTTP relay to gtm-agent's knowledge
- * service, etc.) via `createMcpServer({ feedbackStore })`.
+ * Feedback storage is product policy, so the MCP layer depends on this narrow
+ * interface instead of choosing a knowledge store. The default implementation
+ * is in-memory; consumers wire their own durable adapter via
+ * `createMcpServer({ feedbackStore })`.
  *
  * Feedback events are append-only: every rating is a new event with a
  * fresh id, even when the same delegation is rated multiple times. The
diff --git a/src/mcp/types.ts b/src/mcp/types.ts
index 1cc95f95..576eb140 100644
--- a/src/mcp/types.ts
+++ b/src/mcp/types.ts
@@ -226,9 +226,9 @@ export interface DelegateUiAuditResult {
 }
 
 /**
- * Loose shape of a research output over the wire — the substrate cannot
- * import the `ResearchOutput` type from agent-knowledge without inducing
- * a dependency cycle, so the MCP layer treats it structurally.
+ * Provider-neutral research output carried over the MCP boundary. The MCP
+ * layer accepts this structural shape instead of coupling its wire contract to
+ * one research implementation.
  *
  * @experimental
  */
diff --git a/tests/candidate-execution-core.test.ts b/tests/candidate-execution-core.test.ts
index 4619b703..e94803b2 100644
--- a/tests/candidate-execution-core.test.ts
+++ b/tests/candidate-execution-core.test.ts
@@ -102,7 +102,7 @@ describe('candidate canonical bytes and artifacts', () => {
 
   it('requires workspace manifest bytes to equal canonical material, not only a locator', async () => {
     const material: AgentCandidateWorkspaceManifestMaterial = {
-      schemaVersion: 1,
+      schemaVersion: 2,
       kind: 'agent-candidate-workspace-manifest',
       files: [],
     }
@@ -111,7 +111,7 @@ describe('candidate canonical bytes and artifacts', () => {
     await expect(
       verifyWorkspaceSnapshotArtifacts(
         {
-          schemaVersion: 1,
+          schemaVersion: 2,
           kind: 'agent-candidate-workspace-snapshot',
           digest: manifest.sha256,
           material,
@@ -124,7 +124,7 @@ describe('candidate canonical bytes and artifacts', () => {
     await expect(
       verifyWorkspaceSnapshotArtifacts(
         {
-          schemaVersion: 1,
+          schemaVersion: 2,
           kind: 'agent-candidate-workspace-snapshot',
           digest: manifest.sha256,
           material,
@@ -197,7 +197,7 @@ describe('materialized workspace identity', () => {
     writeFileSync(join(root, 'run.js'), 'ok\n', { mode: 0o755 })
     const bytes = Buffer.from('ok\n')
     const material: AgentCandidateWorkspaceManifestMaterial = {
-      schemaVersion: 1,
+      schemaVersion: 2,
       kind: 'agent-candidate-workspace-manifest',
       files: [
         {
@@ -232,7 +232,7 @@ describe('materialized workspace identity', () => {
     chmodSync(join(root, 'source'), 0o644)
     symlinkSync('source', join(root, 'symlink'))
     const material: AgentCandidateWorkspaceManifestMaterial = {
-      schemaVersion: 1,
+      schemaVersion: 2,
       kind: 'agent-candidate-workspace-manifest',
       files: [],
     }
diff --git a/tests/candidate-execution-execute.test.ts b/tests/candidate-execution-execute.test.ts
index a2d330e1..942aa3bc 100644
--- a/tests/candidate-execution-execute.test.ts
+++ b/tests/candidate-execution-execute.test.ts
@@ -191,7 +191,7 @@ describe('atomic prepared candidate execution', () => {
       succeeded: true,
       receipt: {
         value: {
-          schemaVersion: 1,
+          schemaVersion: 3,
           executorCapture: expect.objectContaining({
             sha256: expect.stringMatching(/^sha256:/),
           }),
@@ -321,7 +321,7 @@ describe('atomic prepared candidate execution', () => {
     expect(result.receipt.value).toMatchObject({
       modelSettlement: {
         material: {
-          schemaVersion: 1,
+          schemaVersion: 2,
           usage: {
             modelCalls: 1,
             inputTokens: 10,
@@ -1506,7 +1506,7 @@ describe('atomic prepared candidate execution', () => {
     }
     const memoryBytes = Buffer.from(secret, 'utf8')
     const afterState = {
-      schemaVersion: 1 as const,
+      schemaVersion: 2 as const,
       kind: 'agent-candidate-workspace-manifest' as const,
       files: [
         {
diff --git a/tests/candidate-execution-finalize.test.ts b/tests/candidate-execution-finalize.test.ts
index edc2866e..819e5f78 100644
--- a/tests/candidate-execution-finalize.test.ts
+++ b/tests/candidate-execution-finalize.test.ts
@@ -218,7 +218,7 @@ describe('protected candidate run finalization', () => {
     if (task.outcome.kind !== 'workspace') throw new Error('expected workspace task outcome')
     if (!task.repository) throw new Error('expected task repository identity')
     expect(result.receipt.value).toMatchObject({
-      schemaVersion: 1,
+      schemaVersion: 3,
       taskOutcome: {
         artifact: result.artifacts.taskOutcome,
         material: {
diff --git a/tests/candidate-execution-prepare.test.ts b/tests/candidate-execution-prepare.test.ts
index df7d0212..4f0ac943 100644
--- a/tests/candidate-execution-prepare.test.ts
+++ b/tests/candidate-execution-prepare.test.ts
@@ -64,7 +64,7 @@ function snapshot(
   files: Array<{ path: string; mode: 0o644 | 0o755 }>,
 ): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: files
       .map((file) => {
@@ -83,7 +83,7 @@ function snapshot(
   }
   const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
   return {
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: manifest.sha256,
     material,
@@ -97,7 +97,7 @@ function bundle(
   active?: { commit: string; tree: string; workspace: AgentCandidateWorkspaceSnapshotEvidence },
 ): AgentCandidateBundle {
   const value = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-bundle' as const,
     digestAlgorithm: 'rfc8785-sha256' as const,
     profile: {
@@ -165,13 +165,13 @@ function redigestBundle(
 
 function emptySnapshot(label: string): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: [],
   }
   const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
   return {
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: manifest.sha256,
     material,
diff --git a/tests/candidate-execution-task-outcome.test.ts b/tests/candidate-execution-task-outcome.test.ts
index c065aa8c..1e8db32b 100644
--- a/tests/candidate-execution-task-outcome.test.ts
+++ b/tests/candidate-execution-task-outcome.test.ts
@@ -50,7 +50,7 @@ function changedOutcome(root: string, baseTree: string) {
   const resultTree = git(root, ['write-tree'], undefined, environment)
   const resultBytes = Buffer.from('export const value = 2\n')
   const afterState = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: [
       {
diff --git a/tests/helpers/candidate-execution-fixture.ts b/tests/helpers/candidate-execution-fixture.ts
index 18b2d473..9e2eb320 100644
--- a/tests/helpers/candidate-execution-fixture.ts
+++ b/tests/helpers/candidate-execution-fixture.ts
@@ -68,7 +68,7 @@ function snapshot(
   files: Array<{ path: string; mode: 0o644 | 0o755 }>,
 ): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: files
       .map((file) => {
@@ -87,7 +87,7 @@ function snapshot(
   }
   const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
   return {
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: manifest.sha256,
     material,
@@ -105,7 +105,7 @@ export function candidateBundle(
   },
 ): AgentCandidateBundle {
   const value = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-bundle' as const,
     digestAlgorithm: 'rfc8785-sha256' as const,
     profile: {
@@ -184,13 +184,13 @@ export function redigestCandidateBundle(
 
 export function emptyCandidateSnapshot(label: string): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
-    schemaVersion: 1 as const,
+    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: [],
   }
   const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
   return {
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: manifest.sha256,
     material,
diff --git a/tests/knowledge-improvement-job.test.ts b/tests/knowledge-improvement-job.test.ts
index 80d88aae..2ec22661 100644
--- a/tests/knowledge-improvement-job.test.ts
+++ b/tests/knowledge-improvement-job.test.ts
@@ -49,7 +49,13 @@ function winner(): SupervisedResult<unknown> {
     out: { ok: true },
     outRef: 'out:1',
     tree: { nodes: [] },
-    spentTotal: { iterations: 2, tokens: { input: 30, output: 12 }, usd: 0.004, ms: 75 },
+    spentTotal: {
+      iterations: 2,
+      tokens: { input: 30, output: 12 },
+      usdKnown: false,
+      usd: 0.004,
+      ms: 75,
+    },
   } as unknown as SupervisedResult<unknown>
 }
 
@@ -289,6 +295,7 @@ describe('runKnowledgeImprovementJob', () => {
         iterations: 2,
         inputTokens: 30,
         outputTokens: 12,
+        usdKnown: false,
         usd: 0.004,
         ms: 75,
       })
@@ -407,7 +414,10 @@ describe('runKnowledgeImprovementJob', () => {
       const approvedUpdate = async () => {
         throw new Error('candidate-ready promotion must not rerun the updater')
       }
-      const promoteApproved = () =>
+      const runApproved = (
+        approvedProposal: AgentImprovementProposal,
+        approvedReview: AgentImprovementReview,
+      ) =>
         runKnowledgeImprovementJob({
           root,
           goal: 'Add runtime job knowledge',
@@ -418,23 +428,41 @@ describe('runKnowledgeImprovementJob', () => {
           runSupervised: approvedUpdate,
           candidateArtifacts: artifacts,
           approval: {
-            proposal,
-            review,
-            authorizeReview: async (candidateReview) => candidateReview.digest === review.digest,
+            proposal: approvedProposal,
+            review: approvedReview,
+            authorizeReview: async (candidateReview) =>
+              candidateReview.digest === approvedReview.digest,
           },
         })
+      const promoteApproved = () => runApproved(proposal, review)
 
-      await withKnowledgeImprovementCandidate(
-        { root, candidate: candidateRef },
-        async ({ root: candidateRoot }) => {
-          const frozenPage = join(candidateRoot, 'knowledge', 'runtime-job.md')
-          const frozenPageBytes = await readFile(frozenPage)
-          await writeFile(frozenPage, 'tampered frozen snapshot', 'utf8')
-          await expect(promoteApproved()).rejects.toThrow(/snapshot changed after approval/)
-          expect(await liveKnowledgeBytes(root)).toEqual(liveBeforeApproval)
-          await writeFile(frozenPage, frozenPageBytes)
-        },
+      await expect(
+        withKnowledgeImprovementCandidate(
+          { root, candidate: candidateRef },
+          async ({ root: candidateRoot }) => {
+            const frozenPage = join(candidateRoot, 'knowledge', 'runtime-job.md')
+            await writeFile(frozenPage, 'tampered frozen snapshot', 'utf8')
+          },
+        ),
+      ).rejects.toThrow(/snapshot changed during use/)
+      expect(await liveKnowledgeBytes(root)).toEqual(liveBeforeApproval)
+
+      const mismatchedProposal = createAgentImprovementProposal({
+        runId: 'runtime-job-mismatched',
+        baselineProfile,
+        findings: [],
+        evaluation: measuredKnowledgeComparison(baselineProfile, mismatchedBundle),
+        candidateBundle: mismatchedBundle,
+      })
+      const mismatchedReview = reviewAgentImprovementProposal(mismatchedProposal, {
+        decision: 'approve',
+        reviewedBy: 'operator@example.com',
+        reason: 'Attempt to approve a candidate identity that was never measured.',
+      })
+      await expect(runApproved(mismatchedProposal, mismatchedReview)).rejects.toThrow(
+        /does not match the measured candidate/,
       )
+      expect(await liveKnowledgeBytes(root)).toEqual(liveBeforeApproval)
 
       const promoted = await promoteApproved()
 

From 2cab99c0ad5d63bc8ea8a1cae1ebdef25cee61ce Mon Sep 17 00:00:00 2001
From: Drew Stone <drewstone329@gmail.com>
Date: Wed, 15 Jul 2026 10:38:00 -0600
Subject: [PATCH 3/8] refactor(contract): remove candidate schema versions

---
 bench/pier_agents/candidate_contract.py       | 65 ++++++++++---------
 bench/pier_agents/tangle_candidate_test.py    | 23 ++++---
 src/candidate-execution/artifacts.ts          |  1 -
 src/candidate-execution/builder.ts            |  1 -
 src/candidate-execution/claim-file-formats.ts |  9 ---
 src/candidate-execution/claim-file-store.ts   | 12 ----
 src/candidate-execution/claim-terminal.ts     | 21 +-----
 src/candidate-execution/claim.ts              | 35 ++--------
 src/candidate-execution/execute.ts            |  4 --
 .../executor-capture-evidence.ts              |  1 -
 src/candidate-execution/finalize.ts           |  4 --
 src/candidate-execution/outcome-evidence.ts   | 10 ---
 src/candidate-execution/prepare.ts            |  5 +-
 src/candidate-execution/prepared-state.ts     |  2 +-
 src/candidate-execution/profile.ts            |  1 -
 src/candidate-execution/workspace-archive.ts  | 35 +++++-----
 src/improvement/agentic-generator.ts          |  2 -
 src/improvement/profile-diff-proposer.test.ts |  7 +-
 src/improvement/profile-diff-proposer.ts      |  2 +-
 src/intelligence/delivery.test.ts             |  1 -
 src/intelligence/improvement-cycle.ts         |  4 --
 .../sandbox-approved-candidate.ts             |  1 -
 src/intelligence/with-intelligence.test.ts    |  2 -
 src/knowledge/improvement-job.ts              |  3 -
 tests/agentic-generator.test.ts               |  1 -
 tests/candidate-bundle-builder.test.ts        |  1 -
 tests/candidate-execution-claim.test.ts       | 13 ++--
 tests/candidate-execution-core.test.ts        |  5 --
 tests/candidate-execution-dispose.test.ts     |  2 +-
 tests/candidate-execution-execute.test.ts     |  3 -
 tests/candidate-execution-finalize.test.ts    |  1 -
 tests/candidate-execution-prepare.test.ts     |  9 +--
 .../candidate-execution-task-outcome.test.ts  |  1 -
 tests/helpers/candidate-execution-fixture.ts  |  5 --
 tests/knowledge-improvement-job.test.ts       |  1 -
 tests/sandbox-approved-candidate.test.ts      |  1 -
 36 files changed, 82 insertions(+), 212 deletions(-)

diff --git a/bench/pier_agents/candidate_contract.py b/bench/pier_agents/candidate_contract.py
index 58920583..bc773401 100644
--- a/bench/pier_agents/candidate_contract.py
+++ b/bench/pier_agents/candidate_contract.py
@@ -107,6 +107,17 @@ def _object(value: Any, label: str) -> dict[str, Any]:
     return value
 
 
+def _contract_object(
+    value: Any, label: str, *, kind: str | None = None
+) -> dict[str, Any]:
+    obj = _object(value, label)
+    if "schemaVersion" in obj or "version" in obj:
+        raise CandidateContractError(f"{label} contains obsolete version fields")
+    if kind is not None and obj.get("kind") != kind:
+        raise CandidateContractError(f"{label} has the wrong kind")
+    return obj
+
+
 def _string(value: Any, label: str) -> str:
     if not isinstance(value, str) or not value or "\0" in value:
         raise CandidateContractError(f"{label} must be a non-empty string without NUL")
@@ -236,18 +247,14 @@ def _embedded_bytes(value: Any, label: str) -> bytes:
 
 
 def _workspace_files(value: Any, label: str) -> tuple[WorkspaceFile, ...]:
-    snapshot = _object(value, label)
-    if (
-        snapshot.get("schemaVersion") != 1
-        or snapshot.get("kind") != "agent-candidate-workspace-snapshot"
-    ):
-        raise CandidateContractError(f"{label} has the wrong snapshot kind")
-    material = _object(snapshot.get("material"), f"{label}.material")
-    if (
-        material.get("schemaVersion") != 1
-        or material.get("kind") != "agent-candidate-workspace-manifest"
-    ):
-        raise CandidateContractError(f"{label}.material has the wrong manifest kind")
+    snapshot = _contract_object(
+        value, label, kind="agent-candidate-workspace-snapshot"
+    )
+    material = _contract_object(
+        snapshot.get("material"),
+        f"{label}.material",
+        kind="agent-candidate-workspace-manifest",
+    )
     values = material.get("files")
     if not isinstance(values, list):
         raise CandidateContractError(f"{label}.material.files must be an array")
@@ -359,26 +366,23 @@ def load_prepared_candidate_contract(
         raise CandidateContractError(
             "materialization receipt bytes do not match the expected digest"
         )
-    if plan.get("schemaVersion") != 1 or plan.get("kind") != _PLAN_KIND:
-        raise CandidateContractError("execution plan has the wrong schema or kind")
-    if receipt.get("schemaVersion") != 1 or receipt.get("kind") != _RECEIPT_KIND:
-        raise CandidateContractError(
-            "materialization receipt has the wrong schema or kind"
-        )
+    plan = _contract_object(plan, "execution plan", kind=_PLAN_KIND)
+    receipt = _contract_object(
+        receipt, "materialization receipt", kind=_RECEIPT_KIND
+    )
     if receipt.get("digestAlgorithm") != "rfc8785-sha256" or "digest" in receipt:
         raise CandidateContractError(
             "materialization receipt must be exact digest-free canonical material"
         )
 
-    execution_evidence = _object(receipt.get("executionPlan"), "receipt.executionPlan")
+    execution_evidence = _contract_object(
+        receipt.get("executionPlan"),
+        "receipt.executionPlan",
+        kind=_EXECUTION_EVIDENCE_KIND,
+    )
     plan_digest = _digest(
         execution_evidence.get("digest"), "receipt.executionPlan.digest"
     )
-    if (
-        execution_evidence.get("schemaVersion") != 1
-        or execution_evidence.get("kind") != _EXECUTION_EVIDENCE_KIND
-    ):
-        raise CandidateContractError("receipt execution evidence has the wrong kind")
     if (
         sha256_bytes(plan_raw) != plan_digest
         or execution_evidence.get("material") != plan
@@ -476,12 +480,11 @@ def load_prepared_candidate_contract(
     if receipt.get("candidateWorkspace") != candidate_snapshot:
         raise CandidateContractError("receipt and plan candidate workspaces differ")
 
-    profile_evidence = _object(receipt.get("profilePlan"), "receipt.profilePlan")
-    if (
-        profile_evidence.get("schemaVersion") != 1
-        or profile_evidence.get("kind") != _PROFILE_PLAN_KIND
-    ):
-        raise CandidateContractError("receipt profile evidence has the wrong kind")
+    profile_evidence = _contract_object(
+        receipt.get("profilePlan"),
+        "receipt.profilePlan",
+        kind=_PROFILE_PLAN_KIND,
+    )
     profile_digest = _digest(
         profile_evidence.get("digest"), "receipt.profilePlan.digest"
     )
@@ -494,7 +497,7 @@ def load_prepared_candidate_contract(
         profile_artifact_material = json.loads(profile_raw)
     except (UnicodeDecodeError, json.JSONDecodeError) as exc:
         raise CandidateContractError("profile artifact is not UTF-8 JSON") from exc
-    profile_material = _object(
+    profile_material = _contract_object(
         profile_evidence.get("material"), "receipt.profilePlan.material"
     )
     if profile_artifact_material != profile_material:
diff --git a/bench/pier_agents/tangle_candidate_test.py b/bench/pier_agents/tangle_candidate_test.py
index c6907e78..a759a3d6 100644
--- a/bench/pier_agents/tangle_candidate_test.py
+++ b/bench/pier_agents/tangle_candidate_test.py
@@ -162,7 +162,6 @@ def _embedded(raw):
 
 def _workspace(files):
     material = {
-        "schemaVersion": 1,
         "kind": "agent-candidate-workspace-manifest",
         "files": [
             {
@@ -176,7 +175,6 @@ def _workspace(files):
     }
     manifest = _canonical(material)
     return {
-        "schemaVersion": 1,
         "kind": "agent-candidate-workspace-snapshot",
         "digest": _sha(manifest),
         "material": material,
@@ -248,7 +246,6 @@ def _fixture(root: Path):
     task_snapshot = _workspace([("src/status.txt", 0o644, status)])
     candidate_snapshot = _workspace([("runner.py", 0o755, runner)])
     profile_material = {
-        "version": 1,
         "harness": "codex",
         "files": [
             {
@@ -265,7 +262,6 @@ def _fixture(root: Path):
     profile_digest = _sha(profile_raw)
     bundle_digest = f"sha256:{'b' * 64}"
     plan = {
-        "schemaVersion": 1,
         "kind": "agent-candidate-execution-plan-material",
         "bundleDigest": bundle_digest,
         "executionId": "pier-fixture-execution",
@@ -346,21 +342,18 @@ def _fixture(root: Path):
     plan_raw = _canonical(plan)
     plan_digest = _sha(plan_raw)
     execution_evidence = {
-        "schemaVersion": 1,
         "kind": "agent-candidate-execution-plan",
         "digest": plan_digest,
         "material": plan,
         "artifact": _embedded(plan_raw),
     }
     profile_evidence = {
-        "schemaVersion": 1,
         "kind": "agent-profile-workspace-plan",
         "digest": profile_digest,
         "material": profile_material,
         "artifact": _embedded(profile_raw),
     }
     receipt = {
-        "schemaVersion": 1,
         "kind": "agent-candidate-materialization",
         "digestAlgorithm": "rfc8785-sha256",
         "bundleDigest": bundle_digest,
@@ -404,7 +397,6 @@ def _rewrite_signed_plan(fixture):
     plan_digest = _sha(plan_raw)
     receipt = json.loads(fixture["receipt_path"].read_text())
     receipt["executionPlan"] = {
-        "schemaVersion": 1,
         "kind": "agent-candidate-execution-plan",
         "digest": plan_digest,
         "material": fixture["plan"],
@@ -782,6 +774,21 @@ def test_rejects_extra_utf8_file_delivery_fields(self):
                     fixture["receipt_digest"],
                 )
 
+    def test_rejects_obsolete_contract_version_fields(self):
+        with tempfile.TemporaryDirectory() as directory:
+            fixture = _fixture(Path(directory))
+            fixture["plan"]["schemaVersion"] = 1
+            _rewrite_signed_plan(fixture)
+
+            with self.assertRaisesRegex(
+                contract_module.CandidateContractError, "obsolete version fields"
+            ):
+                contract_module.load_prepared_candidate_contract(
+                    fixture["plan_path"],
+                    fixture["receipt_path"],
+                    fixture["receipt_digest"],
+                )
+
     def test_accepts_only_frozen_public_model_gateway_domains(self):
         with tempfile.TemporaryDirectory() as directory:
             fixture = _fixture(Path(directory))
diff --git a/src/candidate-execution/artifacts.ts b/src/candidate-execution/artifacts.ts
index 4cbe4446..60183d6f 100644
--- a/src/candidate-execution/artifacts.ts
+++ b/src/candidate-execution/artifacts.ts
@@ -117,7 +117,6 @@ export function candidateWorkspaceManifest(
   files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
 ): AgentCandidateWorkspaceManifestMaterial {
   return {
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-manifest',
     files: files
       .map((file) => ({
diff --git a/src/candidate-execution/builder.ts b/src/candidate-execution/builder.ts
index 0b4f4160..a9ed5a59 100644
--- a/src/candidate-execution/builder.ts
+++ b/src/candidate-execution/builder.ts
@@ -80,7 +80,6 @@ export function buildAgentCandidateBundle(
   const compiledProfile = compileCandidateProfile(input.profile)
   const profileDiffIds = compiledProfile.profileDiffIds
   const bundle: AgentCandidateBundleInput = {
-    schemaVersion: 2,
     kind: 'agent-candidate-bundle',
     digestAlgorithm: 'rfc8785-sha256',
     profile: compiledProfile.profile,
diff --git a/src/candidate-execution/claim-file-formats.ts b/src/candidate-execution/claim-file-formats.ts
index c6e4535c..098f0b02 100644
--- a/src/candidate-execution/claim-file-formats.ts
+++ b/src/candidate-execution/claim-file-formats.ts
@@ -8,11 +8,6 @@ import type { AgentCandidateExecutionClaim, AgentCandidateExecutionTerminalRecor
 import { immutableCandidateValue } from './digest'
 import { assertExactObjectKeys } from './exact-object'
 
-export const CLAIM_FORMAT_VERSION = 8
-export const PENDING_FORMAT_VERSION = 2
-export const TERMINAL_FORMAT_VERSION = 4
-export const PHASE_FORMAT_VERSION = 1
-
 export interface AgentCandidatePreparationEvidence {
   readonly executionPlan: AgentCandidateArtifactRef
   readonly materializationReceipt: AgentCandidateArtifactRef
@@ -44,19 +39,16 @@ export function sealCandidatePreparationEvidence(
 }
 
 export interface PersistedAgentCandidateExecutionClaim extends AgentCandidateExecutionClaim {
-  version: typeof CLAIM_FORMAT_VERSION
   phase: 'claimed'
   leaseDigest: Sha256Digest
 }
 
 export interface PersistedAgentCandidateExecutionPending {
-  version: typeof PENDING_FORMAT_VERSION
   kind: 'candidate-execution-pending-terminal'
   terminal: AgentCandidateExecutionTerminalRecord
 }
 
 export interface PersistedAgentCandidateExecutionPhase {
-  version: typeof PHASE_FORMAT_VERSION
   kind: 'candidate-execution-phase'
   executionId: string
   attempt: number
@@ -65,6 +57,5 @@ export interface PersistedAgentCandidateExecutionPhase {
 }
 
 export interface PersistedAgentCandidateExecutionTerminal {
-  version: typeof TERMINAL_FORMAT_VERSION
   terminal: AgentCandidateExecutionTerminalRecord
 }
diff --git a/src/candidate-execution/claim-file-store.ts b/src/candidate-execution/claim-file-store.ts
index d4d9aeba..15fe4900 100644
--- a/src/candidate-execution/claim-file-store.ts
+++ b/src/candidate-execution/claim-file-store.ts
@@ -21,12 +21,6 @@ import {
   type AgentCandidateRetryRejection,
   candidateClaimFileInternals,
 } from './claim'
-import {
-  CLAIM_FORMAT_VERSION,
-  PENDING_FORMAT_VERSION,
-  PHASE_FORMAT_VERSION,
-  TERMINAL_FORMAT_VERSION,
-} from './claim-file-formats'
 import {
   assertRecoveryMatchesStaged,
   assertTerminalAllowedInPhase,
@@ -100,7 +94,6 @@ export class FileAgentCandidateExecutionClaimStore implements AgentCandidateExec
 
     const lease = newLease(claim)
     const acquired = await writeRecordIfAbsent(this.directory, claimPath, {
-      version: CLAIM_FORMAT_VERSION,
       ...claim,
       phase: 'claimed',
       leaseDigest: leaseDigest(lease),
@@ -138,7 +131,6 @@ export class FileAgentCandidateExecutionClaimStore implements AgentCandidateExec
       this.directory,
       this.transitionPath(stored.claim, 1),
       {
-        version: PHASE_FORMAT_VERSION,
         kind: 'candidate-execution-phase',
         executionId: stored.claim.executionId,
         attempt: stored.claim.attempt,
@@ -174,7 +166,6 @@ export class FileAgentCandidateExecutionClaimStore implements AgentCandidateExec
       this.directory,
       this.transitionPath(stored.claim, transition),
       {
-        version: PENDING_FORMAT_VERSION,
         kind: 'candidate-execution-pending-terminal',
         terminal,
       },
@@ -212,7 +203,6 @@ export class FileAgentCandidateExecutionClaimStore implements AgentCandidateExec
       this.directory,
       terminalPath,
       {
-        version: TERMINAL_FORMAT_VERSION,
         terminal: staged,
       },
       this.ownerPublication(stored.claim),
@@ -245,7 +235,6 @@ export class FileAgentCandidateExecutionClaimStore implements AgentCandidateExec
         this.directory,
         this.transitionPath(record.claim, transition),
         {
-          version: PENDING_FORMAT_VERSION,
           kind: 'candidate-execution-pending-terminal',
           terminal: recovered,
         },
@@ -266,7 +255,6 @@ export class FileAgentCandidateExecutionClaimStore implements AgentCandidateExec
 
     const terminalPath = this.terminalPath(attempt)
     const didFinish = await writeRecordIfAbsent(this.directory, terminalPath, {
-      version: TERMINAL_FORMAT_VERSION,
       terminal: staged,
     })
     if (didFinish) return Object.freeze({ finished: true, terminal: staged })
diff --git a/src/candidate-execution/claim-terminal.ts b/src/candidate-execution/claim-terminal.ts
index 684a3f08..1a30735d 100644
--- a/src/candidate-execution/claim-terminal.ts
+++ b/src/candidate-execution/claim-terminal.ts
@@ -47,7 +47,6 @@ export function recoveredTerminalRecord(
 ): AgentCandidateExecutionTerminalRecord {
   const recovered = sealRecoveryEvidence(evidence, claim)
   return terminalRecord(claim, {
-    schemaVersion: 1,
     status: 'failed',
     failureClass:
       recovered.failureClass === 'pre-model-infrastructure' && phase !== 'claimed'
@@ -147,7 +146,6 @@ export function sealTerminalRecordValue(
           'executionPlanDigest',
           'preparationEvidence',
           'terminalDigest',
-          'schemaVersion',
           'status',
           'usage',
           'modelSettlement',
@@ -162,7 +160,6 @@ export function sealTerminalRecordValue(
           'executionPlanDigest',
           'preparationEvidence',
           'terminalDigest',
-          'schemaVersion',
           'status',
           'failureClass',
           'usage',
@@ -198,7 +195,6 @@ export function sealTerminalRecordValue(
   const result = sealTerminalResult(
     status === 'succeeded'
       ? {
-          schemaVersion: requireNumber(value.schemaVersion, label, 'schemaVersion') as 1,
           status,
           usage: requireObject(value.usage, label, 'usage') as unknown as AgentCandidateFixedSpend,
           modelSettlement: requireArtifactRef(value.modelSettlement, label, 'modelSettlement'),
@@ -207,7 +203,6 @@ export function sealTerminalRecordValue(
           runReceipt: requireArtifactRef(value.runReceipt, label, 'runReceipt'),
         }
       : {
-          schemaVersion: requireNumber(value.schemaVersion, label, 'schemaVersion') as 1,
           status,
           failureClass: requireFailureClass(value.failureClass, label),
           usage: requireObject(value.usage, label, 'usage') as unknown as AgentCandidateFixedSpend,
@@ -360,17 +355,8 @@ function sealTerminalResult(
   assertExactKeys(
     result,
     result.status === 'succeeded'
-      ? [
-          'schemaVersion',
-          'status',
-          'usage',
-          'modelSettlement',
-          'taskOutcome',
-          'benchmarkResult',
-          'runReceipt',
-        ]
+      ? ['status', 'usage', 'modelSettlement', 'taskOutcome', 'benchmarkResult', 'runReceipt']
       : [
-          'schemaVersion',
           'status',
           'failureClass',
           'usage',
@@ -379,14 +365,10 @@ function sealTerminalResult(
         ],
     'candidate execution terminal result',
   )
-  if (result.schemaVersion !== 1) {
-    throw new Error('candidate execution terminal schemaVersion must be 1')
-  }
   const usage = sealUsage(result.usage)
   const modelSettlement = sealArtifactRef(result.modelSettlement, 'modelSettlement')
   if (result.status === 'succeeded') {
     return Object.freeze({
-      schemaVersion: 1,
       status: 'succeeded',
       usage,
       modelSettlement,
@@ -400,7 +382,6 @@ function sealTerminalResult(
     throw new Error('pre-model infrastructure failure cannot contain model calls')
   }
   return Object.freeze({
-    schemaVersion: 1,
     status: 'failed',
     failureClass: result.failureClass,
     usage,
diff --git a/src/candidate-execution/claim.ts b/src/candidate-execution/claim.ts
index 5940620d..84cecfce 100644
--- a/src/candidate-execution/claim.ts
+++ b/src/candidate-execution/claim.ts
@@ -12,15 +12,11 @@ import {
 } from '@tangle-network/agent-interface'
 import {
   type AgentCandidatePreparationEvidence,
-  CLAIM_FORMAT_VERSION,
-  PENDING_FORMAT_VERSION,
   type PersistedAgentCandidateExecutionClaim,
   type PersistedAgentCandidateExecutionPending,
   type PersistedAgentCandidateExecutionPhase,
   type PersistedAgentCandidateExecutionTerminal,
-  PHASE_FORMAT_VERSION,
   sealCandidatePreparationEvidence,
-  TERMINAL_FORMAT_VERSION,
 } from './claim-file-formats'
 import {
   assertRecoveryMatchesStaged,
@@ -89,7 +85,6 @@ export type AgentCandidateExecutionFailureClass =
 /** Evaluator-owned terminal facts staged durably before the terminal CAS. */
 export type AgentCandidateExecutionTerminalResult =
   | {
-      readonly schemaVersion: 1
       readonly status: 'succeeded'
       readonly usage: AgentCandidateFixedSpend
       readonly modelSettlement: AgentCandidateArtifactRef
@@ -98,7 +93,6 @@ export type AgentCandidateExecutionTerminalResult =
       readonly runReceipt: AgentCandidateArtifactRef
     }
   | {
-      readonly schemaVersion: 1
       readonly status: 'failed'
       readonly failureClass: AgentCandidateExecutionFailureClass
       readonly usage: AgentCandidateFixedSpend
@@ -405,8 +399,8 @@ export class InMemoryAgentCandidateExecutionClaimStore
 }
 
 const SHA256_PATTERN = /^sha256:[a-f0-9]{64}$/
-const LEASE_TOKEN_PATTERN = /^candidate-execution-lease-v1\.[A-Za-z0-9_-]{43}$/
-const PREPARATION_ID_PATTERN = /^candidate-preparation-v1\.[A-Za-z0-9_-]{43}$/
+const LEASE_TOKEN_PATTERN = /^candidate-execution-lease\.[A-Za-z0-9_-]{43}$/
+const PREPARATION_ID_PATTERN = /^candidate-preparation\.[A-Za-z0-9_-]{43}$/
 
 function sealClaim(claim: AgentCandidateExecutionClaim): AgentCandidateExecutionClaim {
   assertExactKeys(
@@ -556,7 +550,7 @@ function newLease(claim: AgentCandidateExecutionClaim): AgentCandidateExecutionL
   return Object.freeze({
     executionId: claim.executionId,
     attempt: claim.attempt,
-    token: `candidate-execution-lease-v1.${randomBytes(32).toString('base64url')}`,
+    token: `candidate-execution-lease.${randomBytes(32).toString('base64url')}`,
     expiresAtMs: claim.leaseExpiresAtMs,
   })
 }
@@ -648,13 +642,9 @@ function rejectedRetry(
 async function readClaim(path: string): Promise<StoredClaim> {
   const parsed = await readJsonObject(path, 'claim')
   const record = parsed as Partial<PersistedAgentCandidateExecutionClaim>
-  if (record.version !== CLAIM_FORMAT_VERSION) {
-    throw new Error(`candidate execution claim at ${path} has unsupported version`)
-  }
   assertExactKeys(
     parsed,
     [
-      'version',
       'executionId',
       'attempt',
       'maxAttempts',
@@ -724,10 +714,7 @@ async function readClaimIfPresent(path: string): Promise<StoredClaim | undefined
 async function readTerminal(path: string): Promise<AgentCandidateExecutionTerminalRecord> {
   const parsed = await readJsonObject(path, 'terminal record')
   const record = parsed as Partial<PersistedAgentCandidateExecutionTerminal>
-  if (record.version !== TERMINAL_FORMAT_VERSION) {
-    throw new Error(`candidate execution terminal record at ${path} has unsupported version`)
-  }
-  assertExactKeys(parsed, ['version', 'terminal'], `candidate execution terminal record at ${path}`)
+  assertExactKeys(parsed, ['terminal'], `candidate execution terminal record at ${path}`)
   return sealTerminalRecordValue(
     requireObject(record.terminal, path, 'terminal'),
     `candidate execution terminal record at ${path}`,
@@ -762,12 +749,9 @@ async function readTransitionIfPresent(
   }
   if (parsed.kind === 'candidate-execution-phase') {
     const record = parsed as unknown as PersistedAgentCandidateExecutionPhase
-    if (record.version !== PHASE_FORMAT_VERSION) {
-      throw new Error(`candidate execution phase record at ${path} has unsupported version`)
-    }
     assertExactKeys(
       parsed,
-      ['version', 'kind', 'executionId', 'attempt', 'executionPlanDigest', 'phase'],
+      ['kind', 'executionId', 'attempt', 'executionPlanDigest', 'phase'],
       `candidate execution phase record at ${path}`,
     )
     if (
@@ -782,14 +766,7 @@ async function readTransitionIfPresent(
   }
   if (parsed.kind === 'candidate-execution-pending-terminal') {
     const record = parsed as unknown as PersistedAgentCandidateExecutionPending
-    if (record.version !== PENDING_FORMAT_VERSION) {
-      throw new Error(`candidate execution pending record at ${path} has unsupported version`)
-    }
-    assertExactKeys(
-      parsed,
-      ['version', 'kind', 'terminal'],
-      `candidate execution pending record at ${path}`,
-    )
+    assertExactKeys(parsed, ['kind', 'terminal'], `candidate execution pending record at ${path}`)
     const terminal = sealTerminalRecordValue(
       requireObject(record.terminal, path, 'terminal'),
       `candidate execution pending record at ${path}`,
diff --git a/src/candidate-execution/execute.ts b/src/candidate-execution/execute.ts
index 548bbed0..f05874eb 100644
--- a/src/candidate-execution/execute.ts
+++ b/src/candidate-execution/execute.ts
@@ -456,7 +456,6 @@ export async function executePreparedAgentCandidate(
   try {
     terminal = result.succeeded
       ? {
-          schemaVersion: 1,
           status: 'succeeded',
           usage: settlementResult.settlement.usage,
           modelSettlement: result.artifacts.modelSettlement,
@@ -465,7 +464,6 @@ export async function executePreparedAgentCandidate(
           runReceipt: result.artifacts.runReceipt,
         }
       : {
-          schemaVersion: 1,
           status: 'failed',
           failureClass,
           usage: settlementResult.settlement.usage,
@@ -780,7 +778,6 @@ async function failClaimedExecution(
         claimStore,
         lease,
         {
-          schemaVersion: 1,
           status: 'failed',
           failureClass,
           usage: settled.settlement.usage,
@@ -912,7 +909,6 @@ async function persistFailureEvidence(
   outputArtifacts: AgentCandidateOutputArtifactPort,
 ) {
   const bytes = canonicalCandidateBytes({
-    schemaVersion: 1,
     kind: 'agent-candidate-execution-failure',
     executionId: state.executionId,
     bundleDigest: state.bundle.digest,
diff --git a/src/candidate-execution/executor-capture-evidence.ts b/src/candidate-execution/executor-capture-evidence.ts
index d674180b..1e899c98 100644
--- a/src/candidate-execution/executor-capture-evidence.ts
+++ b/src/candidate-execution/executor-capture-evidence.ts
@@ -73,7 +73,6 @@ export async function persistAgentCandidateExecutorCapture(
       )
     : undefined
   const bytes = canonicalCandidateBytes({
-    schemaVersion: 1,
     kind: 'agent-candidate-executor-capture',
     executionPlanDigest: identity.executionPlanDigest,
     ...(executorEvidence ? { executorEvidence } : {}),
diff --git a/src/candidate-execution/finalize.ts b/src/candidate-execution/finalize.ts
index 8fbd3365..1a402ef0 100644
--- a/src/candidate-execution/finalize.ts
+++ b/src/candidate-execution/finalize.ts
@@ -113,7 +113,6 @@ export async function finalizeAgentCandidateRun(
 
     const redacted = redactProtectedValue(
       {
-        schemaVersion: 1,
         run: { ...run, redactionVersion: REDACTION_VERSION },
         spans: orderedSpans,
         events: orderedEvents,
@@ -145,7 +144,6 @@ export async function finalizeAgentCandidateRun(
       signal,
     })
     const trace = {
-      schemaVersion: 1 as const,
       artifact: traceArtifact,
       eventCount:
         1 +
@@ -156,7 +154,6 @@ export async function finalizeAgentCandidateRun(
       modelCallCount: modelSpans.length,
     }
     const document = canonicalCandidateDocument<AgentCandidateRunReceipt>({
-      schemaVersion: 3,
       kind: 'agent-candidate-run',
       digestAlgorithm: 'rfc8785-sha256',
       bundleDigest: state.bundle.digest,
@@ -266,7 +263,6 @@ async function memoryReceipt(
   const afterState = capture.memoryAfter.afterState
   const manifestBytes = canonicalCandidateBytes(afterState)
   const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: afterState,
diff --git a/src/candidate-execution/outcome-evidence.ts b/src/candidate-execution/outcome-evidence.ts
index 75e5be71..d7178686 100644
--- a/src/candidate-execution/outcome-evidence.ts
+++ b/src/candidate-execution/outcome-evidence.ts
@@ -85,7 +85,6 @@ export async function persistCandidateModelSettlementEvidence(
   outputArtifacts: AgentCandidateOutputArtifactPort,
 ): Promise<PersistedAgentCandidateModelSettlement> {
   const material = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-model-settlement-material' as const,
     executionPlanDigest: identity.executionPlanDigest,
     preparationId: settlement.value.preparationId,
@@ -117,7 +116,6 @@ export async function persistCandidateModelSettlementEvidence(
   })
   return immutableCandidateValue(
     agentCandidateModelSettlementEvidenceSchema.parse({
-      schemaVersion: 2,
       kind: 'agent-candidate-model-settlement',
       digest,
       material,
@@ -225,7 +223,6 @@ async function verifyWorkspaceTaskCapture(
   const manifestBytes = canonicalCandidateBytes(afterState)
   assertNoProtectedBytes(manifestBytes, protectedValues)
   const provisionalSnapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: afterState,
@@ -255,7 +252,6 @@ async function persistVerifiedWorkspaceTaskOutcome(
 ): Promise<VerifiedAgentCandidateTaskOutcome> {
   const { afterState, manifestBytes, patch, repository } = verified
   const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: afterState,
@@ -263,7 +259,6 @@ async function persistVerifiedWorkspaceTaskOutcome(
     archive: persisted.archive,
   })
   const material = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-task-outcome-material' as const,
     executionPlanDigest: state.executionPlan.value.digest,
     outcome: {
@@ -337,7 +332,6 @@ async function persistVerifiedOutputTaskOutcome(
     throw new Error('persisted task output changed the signed media constraints')
   }
   const material = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-task-outcome-material' as const,
     executionPlanDigest: state.executionPlan.value.digest,
     outcome: {
@@ -380,7 +374,6 @@ async function persistTaskOutcomeEvidence<Material extends AgentCandidateTaskOut
   })
   return immutableCandidateValue(
     agentCandidateTaskOutcomeEvidenceSchema.parse({
-      schemaVersion: 2,
       kind: 'agent-candidate-task-outcome',
       digest,
       material,
@@ -425,7 +418,6 @@ export async function persistCandidateBenchmarkResult(
   })
   const task = state.executionPlan.value.material.task
   const material = {
-    schemaVersion: 1 as const,
     kind: 'agent-candidate-benchmark-result-material' as const,
     executionPlanDigest: state.executionPlan.value.digest,
     taskOutcomeDigest: outcome.evidence.digest,
@@ -455,7 +447,6 @@ export async function persistCandidateBenchmarkResult(
   })
   return immutableCandidateValue(
     agentCandidateBenchmarkResultEvidenceSchema.parse({
-      schemaVersion: 1,
       kind: 'agent-candidate-benchmark-result',
       digest,
       material,
@@ -502,7 +493,6 @@ async function verifyMemoryCapture(
   assertNoProtectedBytes(manifestBytes, protectedValues)
   assertNoProtectedBytes(memory.archive, protectedValues)
   const snapshot = agentCandidateWorkspaceSnapshotEvidenceSchema.parse({
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256Bytes(manifestBytes),
     material: memory.afterState,
diff --git a/src/candidate-execution/prepare.ts b/src/candidate-execution/prepare.ts
index 46407fb1..d57026c0 100644
--- a/src/candidate-execution/prepare.ts
+++ b/src/candidate-execution/prepare.ts
@@ -183,7 +183,7 @@ export async function prepareAgentCandidateExecution(
 
   const container = await resolveContainer(candidate, task, ports)
   const resolvedModel = await resolveModel(candidate, task, ports)
-  const preparationId = `candidate-preparation-v1.${randomBytes(32).toString('base64url')}`
+  const preparationId = `candidate-preparation.${randomBytes(32).toString('base64url')}`
   const reservationExpiresAtMs = Date.now() + Math.max(MIN_RESERVATION_TTL_MS, reservationWindowMs)
   const modelReservation = await withinCandidateCleanupDeadline(
     () =>
@@ -231,7 +231,6 @@ export async function prepareAgentCandidateExecution(
     )
     const routes = modelRoutes(bundle.profile, task.model.requested)
     const executionMaterial: AgentCandidateExecutionPlanMaterial = {
-      schemaVersion: 2,
       kind: 'agent-candidate-execution-plan-material',
       bundleDigest: bundle.digest,
       executionId: task.executionId,
@@ -292,7 +291,6 @@ export async function prepareAgentCandidateExecution(
     }
     const executionPlan: AgentCandidateExecutionPlanEvidence =
       agentCandidateExecutionPlanEvidenceSchema.parse({
-        schemaVersion: 2,
         kind: 'agent-candidate-execution-plan',
         digest: executionDigest,
         material: executionMaterial,
@@ -302,7 +300,6 @@ export async function prepareAgentCandidateExecution(
     const entrypoint = candidateEntrypointReceipt(candidate)
     const materializationReceipt = canonicalCandidateDocument<AgentCandidateMaterializationReceipt>(
       {
-        schemaVersion: 2,
         kind: 'agent-candidate-materialization',
         digestAlgorithm: 'rfc8785-sha256',
         bundleDigest: bundle.digest,
diff --git a/src/candidate-execution/prepared-state.ts b/src/candidate-execution/prepared-state.ts
index b6ef6b22..5c096352 100644
--- a/src/candidate-execution/prepared-state.ts
+++ b/src/candidate-execution/prepared-state.ts
@@ -324,7 +324,7 @@ function assertPrivateCandidateIntegrity(state: PreparedCandidateState): void {
     throw new Error('prepared instruction no longer matches the signed execution plan')
   }
   if (
-    !/^candidate-preparation-v1\.[A-Za-z0-9_-]{43}$/.test(state.preparationId) ||
+    !/^candidate-preparation\.[A-Za-z0-9_-]{43}$/.test(state.preparationId) ||
     !Number.isSafeInteger(state.reservationExpiresAtMs) ||
     state.reservationExpiresAtMs <= 0 ||
     !Number.isSafeInteger(state.cleanupTimeoutMs) ||
diff --git a/src/candidate-execution/profile.ts b/src/candidate-execution/profile.ts
index 4abfbc6c..c3ca6da4 100644
--- a/src/candidate-execution/profile.ts
+++ b/src/candidate-execution/profile.ts
@@ -81,7 +81,6 @@ export function createAgentCandidateProfileActivation(
   }
   return parseAgentCandidateProfileActivation(
     canonicalCandidateDocument<AgentCandidateProfileActivation>({
-      schemaVersion: 1,
       kind: 'agent-candidate-profile-activation',
       profilePlan,
       files,
diff --git a/src/candidate-execution/workspace-archive.ts b/src/candidate-execution/workspace-archive.ts
index c02c21c6..a7597faf 100644
--- a/src/candidate-execution/workspace-archive.ts
+++ b/src/candidate-execution/workspace-archive.ts
@@ -73,7 +73,7 @@ const defaultLimits: AgentCandidateWorkspaceArchiveLimits = Object.freeze({
   maxRepositoryBundleBytes: 128 * 1024 * 1024,
 })
 
-interface WorkspaceArchiveRepositoryV1 {
+interface WorkspaceArchiveRepository {
   headCommit: string
   headTree: string
   bundle: {
@@ -83,8 +83,7 @@ interface WorkspaceArchiveRepositoryV1 {
   }
 }
 
-interface WorkspaceArchiveRepositoryMetadataV1 {
-  schemaVersion: 1
+interface WorkspaceArchiveRepositoryMetadata {
   kind: typeof archiveKind
   headCommit: string
   headTree: string
@@ -96,7 +95,7 @@ interface WorkspaceArchiveRepositoryMetadataV1 {
 
 interface DecodedWorkspaceArchive {
   files: Array<{ path: string; mode: number; bytes: Uint8Array }>
-  repository?: WorkspaceArchiveRepositoryV1
+  repository?: WorkspaceArchiveRepository
 }
 
 export interface CaptureAgentCandidateWorkspaceOptions {
@@ -158,7 +157,7 @@ export async function captureAgentCandidateWorkspaceFiles(
 
 async function captureWorkspaceFiles(
   files: readonly AgentCandidateExecutorWorkspaceFile[],
-  repository: WorkspaceArchiveRepositoryV1 | undefined,
+  repository: WorkspaceArchiveRepository | undefined,
   limits: AgentCandidateWorkspaceArchiveLimits,
   artifactPersistence: CaptureAgentCandidateWorkspaceOptions['artifactPersistence'],
 ): Promise<CapturedAgentCandidateWorkspace> {
@@ -177,7 +176,6 @@ async function captureWorkspaceFiles(
   const manifestArtifact = await captureWorkspaceArtifact('manifest', manifest, artifactPersistence)
   const archiveArtifact = await captureWorkspaceArtifact('archive', archive, artifactPersistence)
   const snapshot = deepFreezeCandidate({
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: canonicalCandidateDigest(material),
     material,
@@ -265,7 +263,7 @@ async function captureRepository(
   limits: AgentCandidateWorkspaceArchiveLimits,
 ): Promise<{
   files: Array<{ path: string; mode: number; bytes: Uint8Array }>
-  repository: WorkspaceArchiveRepositoryV1
+  repository: WorkspaceArchiveRepository
 }> {
   const stats = await lstat(root)
   if (!stats.isDirectory() || stats.isSymbolicLink() || (await realpath(root)) !== root) {
@@ -338,7 +336,7 @@ async function captureRepository(
 
 async function encodeWorkspaceArchive(
   files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
-  repository: WorkspaceArchiveRepositoryV1 | undefined,
+  repository: WorkspaceArchiveRepository | undefined,
   maxArchiveBytes: number,
 ): Promise<Uint8Array> {
   const archive = pack()
@@ -356,7 +354,7 @@ async function encodeWorkspaceArchive(
 
 async function verifyCanonicalWorkspaceArchive(
   files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
-  repository: WorkspaceArchiveRepositoryV1 | undefined,
+  repository: WorkspaceArchiveRepository | undefined,
   expected: Uint8Array,
   maxArchiveBytes: number,
 ): Promise<void> {
@@ -381,14 +379,13 @@ async function verifyCanonicalWorkspaceArchive(
 async function writeWorkspaceArchiveEntries(
   archive: Pack,
   files: ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }>,
-  repository: WorkspaceArchiveRepositoryV1 | undefined,
+  repository: WorkspaceArchiveRepository | undefined,
 ): Promise<void> {
   for (const file of files) {
     await writeTarEntry(archive, `${workspaceEntryPrefix}${file.path}`, file.mode, file.bytes)
   }
   if (!repository) return
   const metadata = canonicalCandidateBytes({
-    schemaVersion: 1,
     kind: archiveKind,
     headCommit: repository.headCommit,
     headTree: repository.headTree,
@@ -396,7 +393,7 @@ async function writeWorkspaceArchiveEntries(
       sha256: repository.bundle.sha256,
       byteLength: repository.bundle.byteLength,
     },
-  } satisfies WorkspaceArchiveRepositoryMetadataV1)
+  } satisfies WorkspaceArchiveRepositoryMetadata)
   await writeTarEntry(archive, repositoryMetadataEntry, 0o600, metadata)
   await writeTarEntry(archive, repositoryBundleEntry, 0o600, repository.bundle.bytes)
 }
@@ -415,7 +412,7 @@ async function parseWorkspaceArchive(
   let totalBytes = 0
   let retainedEntryBytes = 0
   let previousPath: string | undefined
-  let repositoryMetadata: WorkspaceArchiveRepositoryMetadataV1 | undefined
+  let repositoryMetadata: WorkspaceArchiveRepositoryMetadata | undefined
   let repositoryBundle: Uint8Array | undefined
   const decoded = (async () => {
     for await (const entry of parser) {
@@ -517,7 +514,7 @@ function assertRetainedArchiveSize(
 function parseRepositoryMetadata(
   bytes: Uint8Array,
   maxBundleBytes: number,
-): WorkspaceArchiveRepositoryMetadataV1 {
+): WorkspaceArchiveRepositoryMetadata {
   let value: unknown
   try {
     value = JSON.parse(new TextDecoder('utf-8', { fatal: true }).decode(bytes))
@@ -526,8 +523,7 @@ function parseRepositoryMetadata(
   }
   if (
     !isRecord(value) ||
-    Object.keys(value).sort().join(',') !== 'bundle,headCommit,headTree,kind,schemaVersion' ||
-    value.schemaVersion !== 1 ||
+    Object.keys(value).sort().join(',') !== 'bundle,headCommit,headTree,kind' ||
     value.kind !== archiveKind ||
     typeof value.headCommit !== 'string' ||
     typeof value.headTree !== 'string' ||
@@ -550,7 +546,6 @@ function parseRepositoryMetadata(
     throw new Error('candidate repository HEAD and tree use different object formats')
   }
   return {
-    schemaVersion: 1,
     kind: archiveKind,
     headCommit: value.headCommit,
     headTree: value.headTree,
@@ -562,9 +557,9 @@ function parseRepositoryMetadata(
 }
 
 function repositoryFromTar(
-  metadata: WorkspaceArchiveRepositoryMetadataV1,
+  metadata: WorkspaceArchiveRepositoryMetadata,
   bundle: Uint8Array,
-): WorkspaceArchiveRepositoryV1 {
+): WorkspaceArchiveRepository {
   verifyBytes(bundle, metadata.bundle.sha256, metadata.bundle.byteLength, 'candidate Git bundle')
   return {
     headCommit: metadata.headCommit,
@@ -699,7 +694,7 @@ async function writeWorkspaceFiles(
 }
 
 async function materializeRepository(
-  repository: WorkspaceArchiveRepositoryV1,
+  repository: WorkspaceArchiveRepository,
   destination: string,
 ): Promise<void> {
   const temporary = await mkdtemp(join(tmpdir(), 'agent-candidate-workspace-bundle-'))
diff --git a/src/improvement/agentic-generator.ts b/src/improvement/agentic-generator.ts
index 703774ed..4229b410 100644
--- a/src/improvement/agentic-generator.ts
+++ b/src/improvement/agentic-generator.ts
@@ -80,7 +80,6 @@ export interface VerifyResult {
 export type Verifier = (worktreePath: string) => Promise<VerifyResult> | VerifyResult
 
 export interface AgenticGeneratorShotReceipt {
-  readonly schemaVersion: 1
   readonly generation: number | null
   readonly candidateIndex: number | null
   /** One-based shot number within this candidate. */
@@ -503,7 +502,6 @@ function shotReceipt(input: {
   const result = input.result
   const costBasis = costBasisFor(input.costReceipt)
   return {
-    schemaVersion: 1,
     generation: input.generation ?? null,
     candidateIndex: input.candidateIndex ?? null,
     shot: input.shot + 1,
diff --git a/src/improvement/profile-diff-proposer.test.ts b/src/improvement/profile-diff-proposer.test.ts
index 504bfe21..6e26e77c 100644
--- a/src/improvement/profile-diff-proposer.test.ts
+++ b/src/improvement/profile-diff-proposer.test.ts
@@ -27,7 +27,6 @@ describe('profileDiffProposer', () => {
       proposeDiffs: () => [
         {
           diff: defineAgentProfileDiff({
-            schemaVersion: 1,
             kind: 'agent-profile-diff',
             id: 'github-review-skill',
             title: 'Pinned review skill',
@@ -100,7 +99,6 @@ describe('profileDiffProposer', () => {
       proposeDiffs: () => [
         {
           diff: {
-            schemaVersion: 1,
             kind: 'agent-profile-diff',
             set: { prompt: { systemPrompt: 'candidate' } },
             unknownField: true,
@@ -114,24 +112,21 @@ describe('profileDiffProposer', () => {
   it('drops no-op diffs and respects the shared population budget', async () => {
     const proposer = profileDiffProposer({
       proposeDiffs: () => [
-        { diff: defineAgentProfileDiff({ schemaVersion: 1, kind: 'agent-profile-diff' }) },
+        { diff: defineAgentProfileDiff({ kind: 'agent-profile-diff' }) },
         {
           diff: defineAgentProfileDiff({
-            schemaVersion: 1,
             kind: 'agent-profile-diff',
             set: { prompt: { systemPrompt: 'baseline' } },
           }),
         },
         {
           diff: defineAgentProfileDiff({
-            schemaVersion: 1,
             kind: 'agent-profile-diff',
             set: { prompt: { systemPrompt: 'first' } },
           }),
         },
         {
           diff: defineAgentProfileDiff({
-            schemaVersion: 1,
             kind: 'agent-profile-diff',
             set: { prompt: { systemPrompt: 'first' } },
           }),
diff --git a/src/improvement/profile-diff-proposer.ts b/src/improvement/profile-diff-proposer.ts
index 8eed304c..62efff43 100644
--- a/src/improvement/profile-diff-proposer.ts
+++ b/src/improvement/profile-diff-proposer.ts
@@ -55,7 +55,7 @@ export function profileDiffProposer<TFindings = unknown>(
       const candidates: ProposedCandidate[] = []
       const normalizedBaseline = applyExactAgentProfileDiff(
         profile,
-        { schemaVersion: 1, kind: 'agent-profile-diff' },
+        { kind: 'agent-profile-diff' },
         'profile diff baseline',
       )
       const seen = new Set<string>([canonicalJson(normalizedBaseline)])
diff --git a/src/intelligence/delivery.test.ts b/src/intelligence/delivery.test.ts
index f87f1798..a01f494c 100644
--- a/src/intelligence/delivery.test.ts
+++ b/src/intelligence/delivery.test.ts
@@ -83,7 +83,6 @@ describe('pullCertified', () => {
 
   it('deserializes the typed agentProfileDiffs the composed endpoint returns', async () => {
     const diff: AgentProfileDiff = {
-      schemaVersion: 1,
       kind: 'agent-profile-diff',
       set: { tools: { refund: true } },
     }
diff --git a/src/intelligence/improvement-cycle.ts b/src/intelligence/improvement-cycle.ts
index 2ba4ea24..f4817767 100644
--- a/src/intelligence/improvement-cycle.ts
+++ b/src/intelligence/improvement-cycle.ts
@@ -236,7 +236,6 @@ export function createAgentImprovementProposal(
   assertMeasuredCandidateBinding(evaluation, baselineProfile, candidateBundle)
   const changedSurfaces = deriveChangedSurfaces(baselineProfile, candidateProfile, candidateBundle)
   const withoutDigest = {
-    schemaVersion: 1 as const,
     kind: 'agent-improvement-proposal' as const,
     runId: options.runId,
     changedSurfaces,
@@ -267,7 +266,6 @@ export function reviewAgentImprovementProposal(
     }
   }
   const withoutDigest = {
-    schemaVersion: 1 as const,
     kind: 'agent-improvement-review' as const,
     proposalDigest: proposal.digest,
     candidateBundleDigest: proposal.candidateBundle.digest,
@@ -310,7 +308,6 @@ export async function executeApprovedAgentCandidate(
   const finalization = await executePreparedAgentCandidate(prepared, options.execution)
   if (!finalization.succeeded) return { finalization }
   const evidence = canonicalCandidateDocument<CandidateExecutionEvidence>({
-    schemaVersion: 1,
     kind: 'agent-candidate-execution-evidence',
     proposalDigest: proposal.digest,
     reviewDigest: review.digest,
@@ -728,7 +725,6 @@ export function createAgentImprovementMeasuredComparison<TScenario extends Scena
   }
 
   const comparison = immutableCandidateValue({
-    schemaVersion: 1 as const,
     kind: 'agent-improvement-measured-comparison' as const,
     benchmark,
     baselineProfileDigest: canonicalCandidateDigest(baselineProfile),
diff --git a/src/intelligence/sandbox-approved-candidate.ts b/src/intelligence/sandbox-approved-candidate.ts
index f4bc1e9e..90395e5d 100644
--- a/src/intelligence/sandbox-approved-candidate.ts
+++ b/src/intelligence/sandbox-approved-candidate.ts
@@ -267,7 +267,6 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
     }
     context.signal.throwIfAborted()
     const evidence = canonicalCandidateBytes({
-      schemaVersion: 1,
       kind: 'sandbox-agent-candidate-capture',
       sandboxId: sandbox.id,
       executionId: request.executionId,
diff --git a/src/intelligence/with-intelligence.test.ts b/src/intelligence/with-intelligence.test.ts
index 2a9f89f7..4c78d8ce 100644
--- a/src/intelligence/with-intelligence.test.ts
+++ b/src/intelligence/with-intelligence.test.ts
@@ -12,7 +12,6 @@ import { withIntelligence } from './with-intelligence'
 /** A valid, promoted profile diff — the previously-DROPPED typed artifact the
  *  composed endpoint returns alongside prompt/skill artifacts. */
 const DIFF: AgentProfileDiff = {
-  schemaVersion: 1,
   kind: 'agent-profile-diff',
   id: 'diff-1',
   title: 'certified refund tool',
@@ -247,7 +246,6 @@ describe('withIntelligence — SEND (a typed RunRecord to /v1/otlp)', () => {
               },
             ],
             candidateExecution: {
-              schemaVersion: 1,
               kind: 'agent-candidate-execution-evidence',
               proposalDigest: `sha256:${'1'.repeat(64)}`,
               reviewDigest: `sha256:${'2'.repeat(64)}`,
diff --git a/src/knowledge/improvement-job.ts b/src/knowledge/improvement-job.ts
index d387fd8c..13b2e677 100644
--- a/src/knowledge/improvement-job.ts
+++ b/src/knowledge/improvement-job.ts
@@ -275,7 +275,6 @@ async function freezeKnowledgeCandidate(
     })
     const evaluation = await captureKnowledgeEvidence(
       canonicalCandidateBytes({
-        schemaVersion: 1,
         kind: 'agent-knowledge-candidate-evaluation',
         candidate: candidateRef,
         metric: resolved.evaluation,
@@ -334,7 +333,6 @@ function interfaceKnowledgeCandidateRef(
   candidate: KnowledgeImprovementCandidateRef,
 ): AgentCandidateKnowledge['candidate'] {
   return {
-    schemaVersion: 1,
     kind: 'knowledge-improvement-candidate',
     runId: candidate.runId,
     candidateId: candidate.candidateId,
@@ -353,7 +351,6 @@ function agentKnowledgeCandidateRef(
   knowledge: AgentCandidateKnowledge,
 ): KnowledgeImprovementCandidateRef {
   return {
-    schemaVersion: 1,
     kind: 'knowledge-improvement-candidate',
     runId: knowledge.candidate.runId,
     candidateId: knowledge.candidate.candidateId,
diff --git a/tests/agentic-generator.test.ts b/tests/agentic-generator.test.ts
index b6253244..da42788c 100644
--- a/tests/agentic-generator.test.ts
+++ b/tests/agentic-generator.test.ts
@@ -192,7 +192,6 @@ describe('agenticGenerator — runs a harness in the worktree', () => {
     expect(out.applied).toBe(true)
     expect(receipts).toHaveLength(1)
     expect(receipts[0]).toMatchObject({
-      schemaVersion: 1,
       generation: 4,
       candidateIndex: 2,
       shot: 1,
diff --git a/tests/candidate-bundle-builder.test.ts b/tests/candidate-bundle-builder.test.ts
index fd3400a2..5cb4f82d 100644
--- a/tests/candidate-bundle-builder.test.ts
+++ b/tests/candidate-bundle-builder.test.ts
@@ -53,7 +53,6 @@ describe('public agent candidate bundle builder', () => {
       resources: { failOnError: true },
     }
     const diff: AgentProfileDiff = {
-      schemaVersion: 1,
       kind: 'agent-profile-diff',
       id: 'add-review-skill',
       source: { kind: 'optimizer', artifacts: ['trace:run-1'] },
diff --git a/tests/candidate-execution-claim.test.ts b/tests/candidate-execution-claim.test.ts
index f57cf006..5ad0e9ff 100644
--- a/tests/candidate-execution-claim.test.ts
+++ b/tests/candidate-execution-claim.test.ts
@@ -174,7 +174,6 @@ describe('candidate execution claim lifecycle', () => {
 
     expect(terminal.usage).toEqual(result.usage)
     expect(terminal).toMatchObject({
-      schemaVersion: 1,
       modelSettlement: result.modelSettlement,
       taskOutcome: result.taskOutcome,
       benchmarkResult: result.benchmarkResult,
@@ -521,7 +520,7 @@ describe('candidate execution claim lifecycle', () => {
     ).toMatchObject({ acquired: false, detail: 'retry-lineage-mismatch' })
   })
 
-  it('persists claim8, pending2, terminal4, phase, staged, and full usage across stores', async () => {
+  it('persists claims, transitions, terminals, and full usage across stores', async () => {
     const directory = await tempDirectory()
     const store = new FileAgentCandidateExecutionClaimStore({ directory })
     const acquired = await acquire(store, claim())
@@ -541,9 +540,7 @@ describe('candidate execution claim lifecycle', () => {
       staged: terminal,
       terminal,
     })
-    expect(files.find(({ name }) => name.endsWith('.claim.json'))?.text).toContain('"version":8')
-    expect(files.find(({ name }) => name.includes('transition-2'))?.text).toContain('"version":2')
-    expect(files.find(({ name }) => name.endsWith('.terminal.json'))?.text).toContain('"version":4')
+    for (const file of files) expect(JSON.parse(file.text)).not.toHaveProperty('version')
     expect(files.map(({ text }) => text).join('\n')).not.toContain(acquired.lease.token)
   })
 
@@ -700,7 +697,6 @@ function failed(
   overrides: Partial<Extract<AgentCandidateExecutionTerminalResult, { status: 'failed' }>> = {},
 ): Extract<AgentCandidateExecutionTerminalResult, { status: 'failed' }> {
   return {
-    schemaVersion: 1,
     status: 'failed',
     failureClass,
     usage: usage(modelCalls),
@@ -711,7 +707,6 @@ function failed(
 
 function succeeded(): Extract<AgentCandidateExecutionTerminalResult, { status: 'succeeded' }> {
   return {
-    schemaVersion: 1,
     status: 'succeeded',
     usage: usage(2),
     modelSettlement: artifact('1'),
@@ -754,7 +749,7 @@ function cleanupHandles(memory = false): AgentCandidateExecutionClaim['cleanup']
 }
 
 function preparationId(character: string): string {
-  return `candidate-preparation-v1.${character.repeat(43)}`
+  return `candidate-preparation.${character.repeat(43)}`
 }
 
 function recoveryEvidence(
@@ -793,7 +788,7 @@ function sha256(character: string): `sha256:${string}` {
 }
 
 function leaseToken(character: string): string {
-  return `candidate-execution-lease-v1.${character.repeat(43)}`
+  return `candidate-execution-lease.${character.repeat(43)}`
 }
 
 async function acquire(
diff --git a/tests/candidate-execution-core.test.ts b/tests/candidate-execution-core.test.ts
index e94803b2..38d37216 100644
--- a/tests/candidate-execution-core.test.ts
+++ b/tests/candidate-execution-core.test.ts
@@ -102,7 +102,6 @@ describe('candidate canonical bytes and artifacts', () => {
 
   it('requires workspace manifest bytes to equal canonical material, not only a locator', async () => {
     const material: AgentCandidateWorkspaceManifestMaterial = {
-      schemaVersion: 2,
       kind: 'agent-candidate-workspace-manifest',
       files: [],
     }
@@ -111,7 +110,6 @@ describe('candidate canonical bytes and artifacts', () => {
     await expect(
       verifyWorkspaceSnapshotArtifacts(
         {
-          schemaVersion: 2,
           kind: 'agent-candidate-workspace-snapshot',
           digest: manifest.sha256,
           material,
@@ -124,7 +122,6 @@ describe('candidate canonical bytes and artifacts', () => {
     await expect(
       verifyWorkspaceSnapshotArtifacts(
         {
-          schemaVersion: 2,
           kind: 'agent-candidate-workspace-snapshot',
           digest: manifest.sha256,
           material,
@@ -197,7 +194,6 @@ describe('materialized workspace identity', () => {
     writeFileSync(join(root, 'run.js'), 'ok\n', { mode: 0o755 })
     const bytes = Buffer.from('ok\n')
     const material: AgentCandidateWorkspaceManifestMaterial = {
-      schemaVersion: 2,
       kind: 'agent-candidate-workspace-manifest',
       files: [
         {
@@ -232,7 +228,6 @@ describe('materialized workspace identity', () => {
     chmodSync(join(root, 'source'), 0o644)
     symlinkSync('source', join(root, 'symlink'))
     const material: AgentCandidateWorkspaceManifestMaterial = {
-      schemaVersion: 2,
       kind: 'agent-candidate-workspace-manifest',
       files: [],
     }
diff --git a/tests/candidate-execution-dispose.test.ts b/tests/candidate-execution-dispose.test.ts
index 89f7fe46..7a73be9a 100644
--- a/tests/candidate-execution-dispose.test.ts
+++ b/tests/candidate-execution-dispose.test.ts
@@ -36,7 +36,7 @@ describe('prepared candidate disposal', () => {
       disposed: true,
     })
     expect(reasons).toHaveLength(1)
-    expect(reasons[0]).toMatch(/candidate-preparation-v1\..+:abandoned/)
+    expect(reasons[0]).toMatch(/candidate-preparation\..+:abandoned/)
     await expect(
       executePreparedAgentCandidate(prepared, {
         traceStore: {} as never,
diff --git a/tests/candidate-execution-execute.test.ts b/tests/candidate-execution-execute.test.ts
index 2f93e5f9..460c6f7d 100644
--- a/tests/candidate-execution-execute.test.ts
+++ b/tests/candidate-execution-execute.test.ts
@@ -194,7 +194,6 @@ describe('atomic prepared candidate execution', () => {
       succeeded: true,
       receipt: {
         value: {
-          schemaVersion: 3,
           executorCapture: expect.objectContaining({
             sha256: expect.stringMatching(/^sha256:/),
           }),
@@ -399,7 +398,6 @@ describe('atomic prepared candidate execution', () => {
     expect(result.receipt.value).toMatchObject({
       modelSettlement: {
         material: {
-          schemaVersion: 2,
           usage: {
             modelCalls: 1,
             inputTokens: 10,
@@ -1591,7 +1589,6 @@ describe('atomic prepared candidate execution', () => {
     }
     const memoryBytes = Buffer.from(secret, 'utf8')
     const afterState = {
-      schemaVersion: 2 as const,
       kind: 'agent-candidate-workspace-manifest' as const,
       files: [
         {
diff --git a/tests/candidate-execution-finalize.test.ts b/tests/candidate-execution-finalize.test.ts
index 819e5f78..d01bc94d 100644
--- a/tests/candidate-execution-finalize.test.ts
+++ b/tests/candidate-execution-finalize.test.ts
@@ -218,7 +218,6 @@ describe('protected candidate run finalization', () => {
     if (task.outcome.kind !== 'workspace') throw new Error('expected workspace task outcome')
     if (!task.repository) throw new Error('expected task repository identity')
     expect(result.receipt.value).toMatchObject({
-      schemaVersion: 3,
       taskOutcome: {
         artifact: result.artifacts.taskOutcome,
         material: {
diff --git a/tests/candidate-execution-prepare.test.ts b/tests/candidate-execution-prepare.test.ts
index 4f0ac943..57c6dd5f 100644
--- a/tests/candidate-execution-prepare.test.ts
+++ b/tests/candidate-execution-prepare.test.ts
@@ -64,7 +64,6 @@ function snapshot(
   files: Array<{ path: string; mode: 0o644 | 0o755 }>,
 ): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: files
       .map((file) => {
@@ -83,7 +82,6 @@ function snapshot(
   }
   const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
   return {
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: manifest.sha256,
     material,
@@ -97,7 +95,6 @@ function bundle(
   active?: { commit: string; tree: string; workspace: AgentCandidateWorkspaceSnapshotEvidence },
 ): AgentCandidateBundle {
   const value = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-bundle' as const,
     digestAlgorithm: 'rfc8785-sha256' as const,
     profile: {
@@ -165,13 +162,11 @@ function redigestBundle(
 
 function emptySnapshot(label: string): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: [],
   }
   const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
   return {
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: manifest.sha256,
     material,
@@ -565,7 +560,7 @@ describe('candidate execution preparation', () => {
       network: { mode: 'gateway-only', domains: ['router.tangle.tools'] },
     })
     mismatched.ports.models.settleGrant = async () => ({
-      preparationId: `candidate-preparation-v1.${'A'.repeat(43)}`,
+      preparationId: `candidate-preparation.${'A'.repeat(43)}`,
       grantDigest: sha('c'),
       closed: true,
       calls: [],
@@ -725,6 +720,6 @@ describe('candidate execution preparation', () => {
       ),
     ).rejects.toThrow(/not scoped/)
     expect(closed).toHaveLength(1)
-    expect(closed[0]).toMatch(/candidate-preparation-v1\..+:preparation-failed/)
+    expect(closed[0]).toMatch(/candidate-preparation\..+:preparation-failed/)
   })
 })
diff --git a/tests/candidate-execution-task-outcome.test.ts b/tests/candidate-execution-task-outcome.test.ts
index 1e8db32b..d370f4d0 100644
--- a/tests/candidate-execution-task-outcome.test.ts
+++ b/tests/candidate-execution-task-outcome.test.ts
@@ -50,7 +50,6 @@ function changedOutcome(root: string, baseTree: string) {
   const resultTree = git(root, ['write-tree'], undefined, environment)
   const resultBytes = Buffer.from('export const value = 2\n')
   const afterState = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: [
       {
diff --git a/tests/helpers/candidate-execution-fixture.ts b/tests/helpers/candidate-execution-fixture.ts
index 9e2eb320..8cb3ea5e 100644
--- a/tests/helpers/candidate-execution-fixture.ts
+++ b/tests/helpers/candidate-execution-fixture.ts
@@ -68,7 +68,6 @@ function snapshot(
   files: Array<{ path: string; mode: 0o644 | 0o755 }>,
 ): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: files
       .map((file) => {
@@ -87,7 +86,6 @@ function snapshot(
   }
   const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
   return {
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: manifest.sha256,
     material,
@@ -105,7 +103,6 @@ export function candidateBundle(
   },
 ): AgentCandidateBundle {
   const value = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-bundle' as const,
     digestAlgorithm: 'rfc8785-sha256' as const,
     profile: {
@@ -184,13 +181,11 @@ export function redigestCandidateBundle(
 
 export function emptyCandidateSnapshot(label: string): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
-    schemaVersion: 2 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files: [],
   }
   const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
   return {
-    schemaVersion: 2,
     kind: 'agent-candidate-workspace-snapshot',
     digest: manifest.sha256,
     material,
diff --git a/tests/knowledge-improvement-job.test.ts b/tests/knowledge-improvement-job.test.ts
index 2ec22661..d97e5e71 100644
--- a/tests/knowledge-improvement-job.test.ts
+++ b/tests/knowledge-improvement-job.test.ts
@@ -142,7 +142,6 @@ function measuredKnowledgeComparison(
     resamples: 2_000,
   }
   return {
-    schemaVersion: 1,
     kind: 'agent-improvement-measured-comparison',
     benchmark: { name: 'development', version: '1', splitDigest: candidateSha('f') },
     baselineProfileDigest: canonicalCandidateDigest(baselineProfile),
diff --git a/tests/sandbox-approved-candidate.test.ts b/tests/sandbox-approved-candidate.test.ts
index ea9e7866..3ea6bcb4 100644
--- a/tests/sandbox-approved-candidate.test.ts
+++ b/tests/sandbox-approved-candidate.test.ts
@@ -375,7 +375,6 @@ function measuredProfileResult(
   bundle: AgentCandidateBundle,
 ): AgentImprovementMeasuredComparison {
   return {
-    schemaVersion: 1,
     kind: 'agent-improvement-measured-comparison',
     benchmark: { name: 'development', version: '1', splitDigest: candidateSha('f') },
     baselineProfileDigest: canonicalCandidateDigest(baselineProfile),

From e81b7c1c8546b6d1b2d57e7a17b5f05c1137e886 Mon Sep 17 00:00:00 2001
From: Drew Stone <drewstone329@gmail.com>
Date: Wed, 15 Jul 2026 12:14:14 -0600
Subject: [PATCH 4/8] fix(candidate): materialize approved knowledge

---
 docs/api/candidate-execution.md               | 18 +++++
 docs/api/index.md                             | 60 ++++++++++++++-
 docs/api/intelligence.md                      | 44 +++++------
 docs/api/primitive-catalog.md                 | 10 ++-
 src/candidate-execution/index.ts              |  5 ++
 src/candidate-execution/knowledge.ts          | 25 +++++-
 src/candidate-execution/prepare.ts            | 29 ++++++-
 .../sandbox-approved-candidate.ts             | 76 +++++++++++++++++-
 tests/candidate-execution-prepare.test.ts     | 19 +++++
 tests/sandbox-approved-candidate.test.ts      | 77 ++++++++++++++++++-
 10 files changed, 330 insertions(+), 33 deletions(-)

diff --git a/docs/api/candidate-execution.md b/docs/api/candidate-execution.md
index ae8f3248..8210eba1 100644
--- a/docs/api/candidate-execution.md
+++ b/docs/api/candidate-execution.md
@@ -194,6 +194,24 @@ Re-exports [executePreparedAgentCandidate](index.md#executepreparedagentcandidat
 
 ***
 
+### CANDIDATE\_KNOWLEDGE\_RETRIEVAL\_CONFIG\_ENV
+
+Re-exports [CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV](index.md#candidate_knowledge_retrieval_config_env)
+
+***
+
+### CANDIDATE\_KNOWLEDGE\_ROOT\_ENV
+
+Re-exports [CANDIDATE_KNOWLEDGE_ROOT_ENV](index.md#candidate_knowledge_root_env)
+
+***
+
+### candidateKnowledgeExecutionPaths
+
+Re-exports [candidateKnowledgeExecutionPaths](index.md#candidateknowledgeexecutionpaths)
+
+***
+
 ### persistCandidateOutputArtifact
 
 Re-exports [persistCandidateOutputArtifact](index.md#persistcandidateoutputartifact)
diff --git a/docs/api/index.md b/docs/api/index.md
index 4191c5f4..236eed40 100644
--- a/docs/api/index.md
+++ b/docs/api/index.md
@@ -1872,7 +1872,7 @@ Maximum time for task verification, executable grading, and receipt construction
 
 ### PrepareAgentCandidateExecutionOptions
 
-Defined in: [candidate-execution/prepare.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L76)
+Defined in: [candidate-execution/prepare.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L81)
 
 #### Properties
 
@@ -1880,13 +1880,13 @@ Defined in: [candidate-execution/prepare.ts:76](https://github.com/tangle-networ
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/prepare.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L77)
+Defined in: [candidate-execution/prepare.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L82)
 
 ##### resultTimeoutMs?
 
 > `optional` **resultTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/prepare.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L79)
+Defined in: [candidate-execution/prepare.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L84)
 
 Maximum time for task verification, executable grading, and receipt construction.
 
@@ -11833,6 +11833,26 @@ MUST map this to `RunRecord.error` rather than recording silent
 
 ## Variables
 
+### CANDIDATE\_KNOWLEDGE\_ROOT\_ENV
+
+> `const` **CANDIDATE\_KNOWLEDGE\_ROOT\_ENV**: `"TANGLE_CANDIDATE_KNOWLEDGE_ROOT"` = `'TANGLE_CANDIDATE_KNOWLEDGE_ROOT'`
+
+Defined in: [candidate-execution/knowledge.ts:14](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/knowledge.ts#L14)
+
+Environment variable containing the materialized candidate knowledge root.
+
+***
+
+### CANDIDATE\_KNOWLEDGE\_RETRIEVAL\_CONFIG\_ENV
+
+> `const` **CANDIDATE\_KNOWLEDGE\_RETRIEVAL\_CONFIG\_ENV**: `"TANGLE_CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG"` = `'TANGLE_CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG'`
+
+Defined in: [candidate-execution/knowledge.ts:16](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/knowledge.ts#L16)
+
+Environment variable containing the materialized retrieval configuration path.
+
+***
+
 ### CANDIDATE\_TRACE\_TAGS
 
 > `const` **CANDIDATE\_TRACE\_TAGS**: `object`
@@ -12355,6 +12375,38 @@ Executes and finalizes one durably claimed candidate without exposing an unprove
 
 ***
 
+### candidateKnowledgeExecutionPaths()
+
+> **candidateKnowledgeExecutionPaths**(`taskRoot`, `hasRetrievalConfig`): `object`
+
+Defined in: [candidate-execution/knowledge.ts:20](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/knowledge.ts#L20)
+
+Deterministic, signed locations used by every candidate executor.
+
+#### Parameters
+
+##### taskRoot
+
+`string`
+
+##### hasRetrievalConfig
+
+`boolean`
+
+#### Returns
+
+`object`
+
+##### root
+
+> **root**: `string`
+
+##### retrievalConfig?
+
+> `optional` **retrievalConfig?**: `string`
+
+***
+
 ### persistCandidateOutputArtifact()
 
 > **persistCandidateOutputArtifact**(`port`, `input`): `Promise`\<`AgentCandidateArtifactRef`\>
@@ -12397,7 +12449,7 @@ Persist evaluator evidence, read it back, and bind the returned locator to the e
 
 > **prepareAgentCandidateExecution**(`candidate`, `task`, `ports`, `options?`): `Promise`\<[`PreparedAgentCandidateExecution`](#preparedagentcandidateexecution)\>
 
-Defined in: [candidate-execution/prepare.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L83)
+Defined in: [candidate-execution/prepare.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L88)
 
 Materializes a verified candidate into one immutable evaluator-owned execution plan.
 
diff --git a/docs/api/intelligence.md b/docs/api/intelligence.md
index d57a9f22..85b1535a 100644
--- a/docs/api/intelligence.md
+++ b/docs/api/intelligence.md
@@ -2530,7 +2530,7 @@ carrying an un-admitted binding kind is a hard error, not a soft drop).
 
 ### CreateSandboxApprovedCandidateExecutorOptions
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L60)
+Defined in: [intelligence/sandbox-approved-candidate.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L65)
 
 #### Properties
 
@@ -2538,43 +2538,43 @@ Defined in: [intelligence/sandbox-approved-candidate.ts:60](https://github.com/t
 
 > **client**: `SandboxClientPort`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L61)
+Defined in: [intelligence/sandbox-approved-candidate.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L66)
 
 ##### ports
 
 > **ports**: [`AgentCandidateExecutionPorts`](index.md#agentcandidateexecutionports)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L62)
+Defined in: [intelligence/sandbox-approved-candidate.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L67)
 
 ##### grader
 
 > **grader**: [`AgentCandidateBenchmarkGraderPort`](index.md#agentcandidatebenchmarkgraderport)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L63)
+Defined in: [intelligence/sandbox-approved-candidate.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L68)
 
 ##### outputArtifacts
 
 > **outputArtifacts**: [`AgentCandidateOutputArtifactPort`](index.md#agentcandidateoutputartifactport)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L64)
+Defined in: [intelligence/sandbox-approved-candidate.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L69)
 
 ##### traceStore
 
 > **traceStore**: `TraceStore`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L65)
+Defined in: [intelligence/sandbox-approved-candidate.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L70)
 
 ##### claimStore
 
 > **claimStore**: [`AgentCandidateExecutionClaimStore`](index.md#agentcandidateexecutionclaimstore)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L66)
+Defined in: [intelligence/sandbox-approved-candidate.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L71)
 
 ##### authorizeReview
 
 > **authorizeReview**: (`review`, `proposal`) => `boolean` \| `Promise`\<`boolean`\>
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L67)
+Defined in: [intelligence/sandbox-approved-candidate.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L72)
 
 ###### Parameters
 
@@ -2594,7 +2594,7 @@ Defined in: [intelligence/sandbox-approved-candidate.ts:67](https://github.com/t
 
 > `optional` **sandbox?**: `object`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L71)
+Defined in: [intelligence/sandbox-approved-candidate.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L76)
 
 ###### teamId?
 
@@ -2616,19 +2616,19 @@ Defined in: [intelligence/sandbox-approved-candidate.ts:71](https://github.com/t
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L77)
+Defined in: [intelligence/sandbox-approved-candidate.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L82)
 
 ##### resultTimeoutMs?
 
 > `optional` **resultTimeoutMs?**: `number`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L78)
+Defined in: [intelligence/sandbox-approved-candidate.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L83)
 
 ***
 
 ### SandboxApprovedCandidateExecution
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L81)
+Defined in: [intelligence/sandbox-approved-candidate.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L86)
 
 #### Properties
 
@@ -2636,31 +2636,31 @@ Defined in: [intelligence/sandbox-approved-candidate.ts:81](https://github.com/t
 
 > **proposal**: `AgentImprovementProposal`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L82)
+Defined in: [intelligence/sandbox-approved-candidate.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L87)
 
 ##### review
 
 > **review**: `AgentImprovementReview`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L83)
+Defined in: [intelligence/sandbox-approved-candidate.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L88)
 
 ##### task
 
 > **task**: [`AgentCandidateTaskExecution`](index.md#agentcandidatetaskexecution)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L84)
+Defined in: [intelligence/sandbox-approved-candidate.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L89)
 
 ##### preparation?
 
 > `optional` **preparation?**: [`PrepareAgentCandidateExecutionOptions`](index.md#prepareagentcandidateexecutionoptions)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L85)
+Defined in: [intelligence/sandbox-approved-candidate.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L90)
 
 ***
 
 ### SandboxApprovedCandidateExecutor
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L88)
+Defined in: [intelligence/sandbox-approved-candidate.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L93)
 
 #### Properties
 
@@ -2668,7 +2668,7 @@ Defined in: [intelligence/sandbox-approved-candidate.ts:88](https://github.com/t
 
 > `readonly` **executor**: [`AgentCandidateExecutorPort`](index.md#agentcandidateexecutorport)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L90)
+Defined in: [intelligence/sandbox-approved-candidate.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L95)
 
 The same port is usable by Runtime's expired-claim recovery path.
 
@@ -2678,7 +2678,7 @@ The same port is usable by Runtime's expired-claim recovery path.
 
 > **execute**(`input`): `Promise`\<`CandidateExecutionEvidence`\>
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L91)
+Defined in: [intelligence/sandbox-approved-candidate.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L96)
 
 ###### Parameters
 
@@ -3293,9 +3293,9 @@ The default tier when a client declares no effort. `'standard'` turns
 
 ### sandboxApprovedCandidateExecutionSupport
 
-> `const` **sandboxApprovedCandidateExecutionSupport**: `Readonly`\<\{ `outcomes`: readonly \[`"output"`\]; `outputMediaTypes`: readonly \[`"text/*"`, `"application/json"`, `"*+json"`\]; `code`: readonly \[`"disabled"`\]; `memory`: readonly \[`"disabled"`\]; `knowledge`: `false`; `profile`: `Readonly`\<\{ `mcpTransports`: readonly \[`"stdio"`\]; `remoteMcp`: `false`; `tools`: `false`; `permissions`: `false`; `modes`: `false`; `confidential`: `false`; \}\>; `isolation`: `Readonly`\<\{ `freshSandbox`: `true`; `exactProcess`: `true`; `egress`: readonly \[`"blocked"`, `"strict"`\]; \}\>; \}\>
+> `const` **sandboxApprovedCandidateExecutionSupport**: `Readonly`\<\{ `outcomes`: readonly \[`"output"`\]; `outputMediaTypes`: readonly \[`"text/*"`, `"application/json"`, `"*+json"`\]; `code`: readonly \[`"disabled"`\]; `memory`: readonly \[`"disabled"`\]; `knowledge`: `true`; `profile`: `Readonly`\<\{ `mcpTransports`: readonly \[`"stdio"`\]; `remoteMcp`: `false`; `tools`: `false`; `permissions`: `false`; `modes`: `false`; `confidential`: `false`; \}\>; `isolation`: `Readonly`\<\{ `freshSandbox`: `true`; `exactProcess`: `true`; `egress`: readonly \[`"blocked"`, `"strict"`\]; \}\>; \}\>
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L46)
+Defined in: [intelligence/sandbox-approved-candidate.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L51)
 
 Declares the exact candidate surfaces the sandbox executor can run.
 
@@ -3886,7 +3886,7 @@ Lower a plane `CertifiedProfile` straight into a `ResolvedSurface` via
 
 > **createSandboxApprovedCandidateExecutor**(`options`): [`SandboxApprovedCandidateExecutor`](#sandboxapprovedcandidateexecutor)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L95)
+Defined in: [intelligence/sandbox-approved-candidate.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L100)
 
 Compose approved-candidate execution directly onto fresh Tangle sandboxes.
 
diff --git a/docs/api/primitive-catalog.md b/docs/api/primitive-catalog.md
index 36916301..dc38bb52 100644
--- a/docs/api/primitive-catalog.md
+++ b/docs/api/primitive-catalog.md
@@ -15,7 +15,7 @@ Every subpath this package declares in `package.json` `exports`. Reach for these
 
 ### Root — task lifecycle, conversation, RSI verbs, observability
 
-Import from `@tangle-network/agent-runtime` — 340 exports.
+Import from `@tangle-network/agent-runtime` — 343 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -30,6 +30,7 @@ Import from `@tangle-network/agent-runtime` — 340 exports.
 | `buildLoopSpanNodes` | function | Sink-neutral core behind {@link buildLoopOtelSpans}: reconstruct the |
 | `buildRuntimeEventOtelSpans` | function | Convert normalized runtime events into lossless, redacted child spans. |
 | `candidateExecutionClaim` | function | Extract the complete durable claim from a prepared execution. |
+| `candidateKnowledgeExecutionPaths` | function | Deterministic, signed locations used by every candidate executor. |
 | `captureAgentCandidateWorkspace` | function | Capture one exact regular-file workspace for immutable candidate execution. |
 | `captureAgentCandidateWorkspaceFiles` | function | Capture detached files returned by a remote executor into the standard archive. |
 | `cleanModelId` | function | Trim a candidate model id; `undefined` for non-strings and blanks. |
@@ -114,6 +115,8 @@ Import from `@tangle-network/agent-runtime` — 340 exports.
 | `worktreeLoopRunner` | function | `code` mode on the GENERIC recursive path: author one `AgentProfile` per harness, run them as a |
 | `AGENT_CANDIDATE_EXECUTION_SUPPORT` | const | Surfaces admitted by Runtime's verifier before an environment adapter is selected. |
 | `AGENTIC_PROFILE_RESOURCE_ROOT` | const | Dedicated ephemeral root for generic author-profile files. Every declared |
+| `CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV` | const | Environment variable containing the materialized retrieval configuration path. |
+| `CANDIDATE_KNOWLEDGE_ROOT_ENV` | const | Environment variable containing the materialized candidate knowledge root. |
 | `CANDIDATE_TRACE_ENV` | const | Environment keys used to propagate immutable candidate trace identity. |
 | `CANDIDATE_TRACE_TAGS` | const | Protected trace tags that bind a run to one prepared candidate execution. |
 | `DEFAULT_MAX_DEPTH` | const | Hard cap on chained gateway hops; refused beyond this. Default keeps recursion bounded. |
@@ -945,13 +948,14 @@ Import from `@tangle-network/agent-runtime/primeintellect` — 27 exports.
 
 ### Candidate execution — immutable prepare, run, grade, and receipt
 
-Import from `@tangle-network/agent-runtime/candidate-execution` — 94 exports.
+Import from `@tangle-network/agent-runtime/candidate-execution` — 97 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
 | `applyExactAgentProfileDiff` | function | Apply one exact diff and reject any value that cannot be preserved canonically. |
 | `buildAgentCandidateBundle` | function | Compile one measured profile/code candidate into the immutable execution |
 | `candidateExecutionClaim` | function | Extract the complete durable claim from a prepared execution. |
+| `candidateKnowledgeExecutionPaths` | function | Deterministic, signed locations used by every candidate executor. |
 | `captureAgentCandidateWorkspace` | function | Capture one exact regular-file workspace for immutable candidate execution. |
 | `captureAgentCandidateWorkspaceFiles` | function | Capture detached files returned by a remote executor into the standard archive. |
 | `createAgentCandidateWorkspacePort` | function | Create the standard bounded materializer for candidate execution ports. |
@@ -966,6 +970,8 @@ Import from `@tangle-network/agent-runtime/candidate-execution` — 94 exports.
 | `sealAgentCandidateBundle` | function | Validate and content-address a candidate bundle before it crosses an approval boundary. |
 | `verifyAgentCandidateBundle` | function | Verifies every digest, resource, workspace, and Git object in a candidate bundle. |
 | `AGENT_CANDIDATE_EXECUTION_SUPPORT` | const | Surfaces admitted by Runtime's verifier before an environment adapter is selected. |
+| `CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV` | const | Environment variable containing the materialized retrieval configuration path. |
+| `CANDIDATE_KNOWLEDGE_ROOT_ENV` | const | Environment variable containing the materialized candidate knowledge root. |
 | `CANDIDATE_TRACE_ENV` | const | Environment keys used to propagate immutable candidate trace identity. |
 | `CANDIDATE_TRACE_TAGS` | const | Protected trace tags that bind a run to one prepared candidate execution. |
 | `FileAgentCandidateExecutionClaimStore` | class | Cross-process lifecycle implemented as fsynced, create-if-absent records. |
diff --git a/src/candidate-execution/index.ts b/src/candidate-execution/index.ts
index 16f7c469..7ff08ba1 100644
--- a/src/candidate-execution/index.ts
+++ b/src/candidate-execution/index.ts
@@ -41,6 +41,11 @@ export {
   type ExecutePreparedAgentCandidateOptions,
   executePreparedAgentCandidate,
 } from './execute'
+export {
+  CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV,
+  CANDIDATE_KNOWLEDGE_ROOT_ENV,
+  candidateKnowledgeExecutionPaths,
+} from './knowledge'
 export { persistCandidateOutputArtifact } from './output-artifacts'
 export {
   type PrepareAgentCandidateExecutionOptions,
diff --git a/src/candidate-execution/knowledge.ts b/src/candidate-execution/knowledge.ts
index 90ca254a..83ab3220 100644
--- a/src/candidate-execution/knowledge.ts
+++ b/src/candidate-execution/knowledge.ts
@@ -1,6 +1,6 @@
 import { mkdtemp, rm } from 'node:fs/promises'
 import { tmpdir } from 'node:os'
-import { join } from 'node:path'
+import { join, posix } from 'node:path'
 
 import { readMaterializedWorkspaceFiles } from './artifacts'
 import type {
@@ -10,6 +10,29 @@ import type {
 } from './types'
 import { verifiedArtifactBytes } from './verify'
 
+/** Environment variable containing the materialized candidate knowledge root. */
+export const CANDIDATE_KNOWLEDGE_ROOT_ENV = 'TANGLE_CANDIDATE_KNOWLEDGE_ROOT'
+/** Environment variable containing the materialized retrieval configuration path. */
+export const CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV =
+  'TANGLE_CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG'
+
+/** Deterministic, signed locations used by every candidate executor. */
+export function candidateKnowledgeExecutionPaths(
+  taskRoot: string,
+  hasRetrievalConfig: boolean,
+): { root: string; retrievalConfig?: string } {
+  if (!posix.isAbsolute(taskRoot) || posix.normalize(taskRoot) !== taskRoot) {
+    throw new Error('candidate knowledge requires a normalized absolute task root')
+  }
+  const parent = posix.join(taskRoot, '.tangle')
+  return {
+    root: posix.join(parent, 'knowledge'),
+    ...(hasRetrievalConfig
+      ? { retrievalConfig: posix.join(parent, 'knowledge-retrieval-config.json') }
+      : {}),
+  }
+}
+
 /** Verify and detach the exact file-backed knowledge admitted by the bundle. */
 export async function prepareAgentCandidateKnowledge(
   candidate: VerifiedAgentCandidate,
diff --git a/src/candidate-execution/prepare.ts b/src/candidate-execution/prepare.ts
index 6e6ded92..cf142437 100644
--- a/src/candidate-execution/prepare.ts
+++ b/src/candidate-execution/prepare.ts
@@ -51,7 +51,12 @@ import {
 } from './digest'
 import { candidateExecutionOwnerWindowMs } from './execution-window'
 import { verifyTaskCheckout } from './git-materialize'
-import { prepareAgentCandidateKnowledge } from './knowledge'
+import {
+  CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV,
+  CANDIDATE_KNOWLEDGE_ROOT_ENV,
+  candidateKnowledgeExecutionPaths,
+  prepareAgentCandidateKnowledge,
+} from './knowledge'
 import { sealAgentCandidateModelSettlement, usdToNanos } from './model-settlement'
 import { createPreparedCandidateExecution } from './prepared-state'
 import { candidateMaterializerHarness, createAgentCandidateProfileActivation } from './profile'
@@ -218,6 +223,12 @@ export async function prepareAgentCandidateExecution(
     )
     const memory = preparedMemory.value
     const knowledge = await prepareAgentCandidateKnowledge(candidate, ports)
+    const knowledgePaths = knowledge
+      ? candidateKnowledgeExecutionPaths(
+          task.executionRoots.taskRoot,
+          knowledge.retrievalConfig !== undefined,
+        )
+      : undefined
     const baseLaunch = buildLaunch(candidate, task, profileApplication.flags)
     const publicEnv = mergePublicEnvironment(
       bundle.execution.env ?? {},
@@ -230,6 +241,22 @@ export async function prepareAgentCandidateExecution(
             },
           }
         : {},
+      knowledgePaths
+        ? {
+            [CANDIDATE_KNOWLEDGE_ROOT_ENV]: {
+              kind: 'public',
+              value: knowledgePaths.root,
+            },
+            ...(knowledgePaths.retrievalConfig
+              ? {
+                  [CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV]: {
+                    kind: 'public' as const,
+                    value: knowledgePaths.retrievalConfig,
+                  },
+                }
+              : {}),
+          }
+        : {},
     )
     const routes = modelRoutes(bundle.profile, task.model.requested)
     const executionMaterial: AgentCandidateExecutionPlanMaterial = {
diff --git a/src/intelligence/sandbox-approved-candidate.ts b/src/intelligence/sandbox-approved-candidate.ts
index 90395e5d..b0fd4ecd 100644
--- a/src/intelligence/sandbox-approved-candidate.ts
+++ b/src/intelligence/sandbox-approved-candidate.ts
@@ -20,6 +20,11 @@ import type {
 
 import type { AgentCandidateExecutionClaimStore } from '../candidate-execution/claim'
 import { canonicalCandidateBytes } from '../candidate-execution/digest'
+import {
+  CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV,
+  CANDIDATE_KNOWLEDGE_ROOT_ENV,
+  candidateKnowledgeExecutionPaths,
+} from '../candidate-execution/knowledge'
 import type { PrepareAgentCandidateExecutionOptions } from '../candidate-execution/prepare'
 import type {
   AgentCandidateBenchmarkGraderPort,
@@ -48,7 +53,7 @@ export const sandboxApprovedCandidateExecutionSupport = Object.freeze({
   outputMediaTypes: Object.freeze(['text/*', 'application/json', '*+json'] as const),
   code: Object.freeze(['disabled'] as const),
   memory: Object.freeze(['disabled'] as const),
-  knowledge: false,
+  knowledge: true,
   profile: AGENT_CANDIDATE_EXECUTION_SUPPORT.profile,
   isolation: Object.freeze({
     freshSandbox: true,
@@ -403,6 +408,7 @@ async function materializeRequest(
   sandbox: SandboxInstance,
   request: AgentCandidateExecutorRequest,
 ): Promise<void> {
+  const knowledgePaths = assertKnowledgeExecutionBinding(request)
   for (const file of request.inputs.task.files) {
     await writeExactFile(sandbox, beneath(request.roots.taskRoot, file.path), file.bytes, file.mode)
   }
@@ -419,6 +425,19 @@ async function materializeRequest(
       file.mode,
     )
   }
+  if (request.knowledge && knowledgePaths) {
+    for (const file of request.knowledge.files) {
+      await writeExactFile(sandbox, beneath(knowledgePaths.root, file.path), file.bytes, file.mode)
+    }
+    if (request.knowledge.retrievalConfig && knowledgePaths.retrievalConfig) {
+      await writeExactFile(
+        sandbox,
+        knowledgePaths.retrievalConfig,
+        request.knowledge.retrievalConfig,
+        0o644,
+      )
+    }
+  }
   if (request.instruction.delivery.kind === 'utf8-file') {
     await writeExactFile(
       sandbox,
@@ -429,6 +448,61 @@ async function materializeRequest(
   }
 }
 
+function assertKnowledgeExecutionBinding(
+  request: AgentCandidateExecutorRequest,
+): ReturnType<typeof candidateKnowledgeExecutionPaths> | undefined {
+  const knowledge = request.knowledge
+  if (!knowledge) {
+    if (
+      request.launch.env[CANDIDATE_KNOWLEDGE_ROOT_ENV] !== undefined ||
+      request.launch.env[CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV] !== undefined
+    ) {
+      throw new Error('candidate launch declares knowledge paths without verified knowledge')
+    }
+    return undefined
+  }
+  const paths = candidateKnowledgeExecutionPaths(
+    request.roots.taskRoot,
+    knowledge.retrievalConfig !== undefined,
+  )
+  if (
+    request.launch.env[CANDIDATE_KNOWLEDGE_ROOT_ENV] !== paths.root ||
+    request.launch.env[CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV] !== paths.retrievalConfig
+  ) {
+    throw new Error('candidate knowledge paths do not match the signed launch environment')
+  }
+
+  const reserved = [paths.root, paths.retrievalConfig].filter(
+    (value): value is string => value !== undefined,
+  )
+  const profileRoot =
+    request.executionPlan.value.material.profile.targetWorkspace === 'task'
+      ? request.roots.taskRoot
+      : request.roots.candidateRoot
+  const otherFiles = [
+    ...request.inputs.task.files.map((file) => beneath(request.roots.taskRoot, file.path)),
+    ...(profileRoot
+      ? request.profileActivation.files.map((file) => beneath(profileRoot, file.path))
+      : []),
+    ...(request.instruction.delivery.kind === 'utf8-file'
+      ? [request.instruction.delivery.path]
+      : []),
+  ]
+  if (
+    otherFiles.some((path) =>
+      reserved.some(
+        (reservedPath) =>
+          path === reservedPath ||
+          path.startsWith(`${reservedPath}/`) ||
+          reservedPath.startsWith(`${path}/`),
+      ),
+    )
+  ) {
+    throw new Error('candidate knowledge paths overlap other execution inputs')
+  }
+  return paths
+}
+
 function exactLaunch(request: AgentCandidateExecutorRequest): {
   executable: string
   args: readonly string[]
diff --git a/tests/candidate-execution-prepare.test.ts b/tests/candidate-execution-prepare.test.ts
index 6dde1b86..cf19c9d7 100644
--- a/tests/candidate-execution-prepare.test.ts
+++ b/tests/candidate-execution-prepare.test.ts
@@ -16,6 +16,10 @@ import {
   canonicalCandidateDigest,
   embeddedCandidateArtifact,
 } from '../src/candidate-execution/digest'
+import {
+  CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV,
+  CANDIDATE_KNOWLEDGE_ROOT_ENV,
+} from '../src/candidate-execution/knowledge'
 import { prepareAgentCandidateExecution } from '../src/candidate-execution/prepare'
 import type {
   AgentCandidateExecutionPorts,
@@ -694,6 +698,21 @@ describe('candidate execution preparation', () => {
       files: [],
     })
     expect(Buffer.from(prepared.knowledge?.retrievalConfig ?? [])).toEqual(knowledgeBytes)
+    expect(prepared.launch.env).toMatchObject({
+      [CANDIDATE_KNOWLEDGE_ROOT_ENV]: '/workspace/task/.tangle/knowledge',
+      [CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV]:
+        '/workspace/task/.tangle/knowledge-retrieval-config.json',
+    })
+    expect(prepared.executionPlan.value.material.launch.env).toMatchObject({
+      [CANDIDATE_KNOWLEDGE_ROOT_ENV]: {
+        kind: 'public',
+        value: '/workspace/task/.tangle/knowledge',
+      },
+      [CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV]: {
+        kind: 'public',
+        value: '/workspace/task/.tangle/knowledge-retrieval-config.json',
+      },
+    })
   })
 
   it('rejects snapshot-only knowledge before artifact access', async () => {
diff --git a/tests/sandbox-approved-candidate.test.ts b/tests/sandbox-approved-candidate.test.ts
index 3ea6bcb4..955aa2af 100644
--- a/tests/sandbox-approved-candidate.test.ts
+++ b/tests/sandbox-approved-candidate.test.ts
@@ -1,3 +1,6 @@
+import { chmodSync, mkdirSync, writeFileSync } from 'node:fs'
+import { dirname, join } from 'node:path'
+
 import { InMemoryTraceStore } from '@tangle-network/agent-eval'
 import type {
   AgentCandidateBundle,
@@ -8,7 +11,15 @@ import type { Process, ProcessStatus, SandboxInstance } from '@tangle-network/sa
 import { afterEach, describe, expect, it, vi } from 'vitest'
 
 import { InMemoryAgentCandidateExecutionClaimStore } from '../src/candidate-execution/claim'
-import { canonicalCandidateDigest } from '../src/candidate-execution/digest'
+import {
+  canonicalCandidateBytes,
+  canonicalCandidateDigest,
+  embeddedCandidateArtifact,
+} from '../src/candidate-execution/digest'
+import {
+  CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV,
+  CANDIDATE_KNOWLEDGE_ROOT_ENV,
+} from '../src/candidate-execution/knowledge'
 import type { AgentCandidateExecutorRequest } from '../src/candidate-execution/types'
 import {
   createAgentImprovementProposal,
@@ -23,6 +34,7 @@ import {
   cleanupCandidateFixtures,
   createCandidateOutputExecutionFixture,
   createCandidateOutputFixture,
+  redigestCandidateBundle,
 } from './helpers/candidate-execution-fixture'
 
 afterEach(cleanupCandidateFixtures)
@@ -30,6 +42,52 @@ afterEach(cleanupCandidateFixtures)
 describe('Sandbox approved candidate executor', () => {
   it('runs one exact bounded-output task in a fresh strict-egress sandbox', async () => {
     const fixture = createCandidateOutputExecutionFixture('application/json', 1_024)
+    const knowledgeBytes = Buffer.from('# Exact runbook\nUse the verified endpoint.\n', 'utf8')
+    const knowledgeMaterial = {
+      kind: 'agent-candidate-workspace-manifest' as const,
+      files: [
+        {
+          path: 'runbook.md',
+          mode: 0o644,
+          sha256: embeddedCandidateArtifact(knowledgeBytes).sha256,
+          byteLength: knowledgeBytes.byteLength,
+        },
+      ],
+    }
+    const knowledgeManifest = embeddedCandidateArtifact(canonicalCandidateBytes(knowledgeMaterial))
+    const knowledgeSnapshot = {
+      kind: 'agent-candidate-workspace-snapshot' as const,
+      digest: knowledgeManifest.sha256,
+      material: knowledgeMaterial,
+      manifest: knowledgeManifest,
+      archive: embeddedCandidateArtifact(Buffer.from('knowledge-archive', 'utf8')),
+    }
+    const retrievalConfigBytes = Buffer.from('{"topK":4}\n', 'utf8')
+    fixture.bundle = redigestCandidateBundle(fixture.bundle, {
+      knowledge: {
+        candidate: {
+          kind: 'knowledge-improvement-candidate',
+          runId: 'knowledge-run-1',
+          candidateId: 'knowledge-candidate-1',
+          goalHash: candidateSha('1'),
+          baseHash: candidateSha('2'),
+          candidateHash: candidateSha('3'),
+          evidenceHash: candidateSha('4'),
+          promotionPlanHash: candidateSha('5'),
+        },
+        snapshot: knowledgeSnapshot,
+        retrievalConfig: embeddedCandidateArtifact(retrievalConfigBytes),
+        evaluation: embeddedCandidateArtifact(Buffer.from('{"score":1}\n', 'utf8')),
+      },
+    })
+    const materialize = fixture.ports.workspaces.materialize
+    fixture.ports.workspaces.materialize = async (input) => {
+      if (input.role !== 'knowledge') return materialize(input)
+      const path = join(input.destination, 'runbook.md')
+      mkdirSync(dirname(path), { recursive: true })
+      writeFileSync(path, knowledgeBytes)
+      chmodSync(path, 0o644)
+    }
     const baselineProfile = { name: 'baseline' }
     const proposal = createAgentImprovementProposal({
       runId: 'sandbox-proposal-1',
@@ -135,6 +193,11 @@ describe('Sandbox approved candidate executor', () => {
     })
     expect(launch.env).not.toHaveProperty('PATH')
     expect(launch.env).toHaveProperty('MODEL_GATEWAY_TOKEN', 'protected')
+    expect(launch.env).toMatchObject({
+      [CANDIDATE_KNOWLEDGE_ROOT_ENV]: '/workspace/task/.tangle/knowledge',
+      [CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV]:
+        '/workspace/task/.tangle/knowledge-retrieval-config.json',
+    })
     expect(writes.length).toBeGreaterThan(0)
     expect(deleted).toBe(true)
     expect(evidence).toMatchObject({
@@ -167,6 +230,16 @@ describe('Sandbox approved candidate executor', () => {
       expect(written?.mode).toBe(file.mode)
       expect(Buffer.from(written?.content ?? '', 'base64').toString('utf8')).toBe(file.content)
     }
+    const writtenKnowledge = writes.find(
+      (entry) => entry.path === '/workspace/task/.tangle/knowledge/runbook.md',
+    )
+    expect(Buffer.from(writtenKnowledge?.content ?? '', 'base64')).toEqual(knowledgeBytes)
+    const writtenRetrievalConfig = writes.find(
+      (entry) => entry.path === '/workspace/task/.tangle/knowledge-retrieval-config.json',
+    )
+    expect(Buffer.from(writtenRetrievalConfig?.content ?? '', 'base64')).toEqual(
+      retrievalConfigBytes,
+    )
     await expect(
       adapter.executor.stop(
         {
@@ -189,7 +262,7 @@ describe('Sandbox approved candidate executor', () => {
       outputMediaTypes: ['text/*', 'application/json', '*+json'],
       code: ['disabled'],
       memory: ['disabled'],
-      knowledge: false,
+      knowledge: true,
       profile: {
         mcpTransports: ['stdio'],
         remoteMcp: false,

From 72ebca0d23766f35911c22bd5e22e94c15970468 Mon Sep 17 00:00:00 2001
From: Drew Stone <drewstone329@gmail.com>
Date: Wed, 15 Jul 2026 13:30:54 -0600
Subject: [PATCH 5/8] fix(package): verify compatible runtime and bench
 releases

---
 .github/workflows/ci.yml                   |  18 +-
 .github/workflows/publish.yml              |  35 +-
 bench/package.json                         |  20 +-
 bench/pier_agents/tangle_candidate_test.py |  13 +-
 bench/pnpm-lock.yaml                       | 958 ---------------------
 bench/scripts/verify-packed-consumer.mjs   |  67 +-
 bench/scripts/verify-pier-agent.mts        |  11 +-
 package.json                               |   4 +-
 pnpm-lock.yaml                             |  25 +
 pnpm-workspace.yaml                        |   2 +
 scripts/verify-package-exports.mjs         |  54 +-
 11 files changed, 167 insertions(+), 1040 deletions(-)
 delete mode 100644 bench/pnpm-lock.yaml
 create mode 100644 pnpm-workspace.yaml

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index b90e790e..48e1b5dc 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -42,9 +42,6 @@ jobs:
 
   agent-bench:
     runs-on: ubuntu-latest
-    defaults:
-      run:
-        working-directory: bench
     steps:
       - uses: actions/checkout@v4
 
@@ -54,16 +51,13 @@ jobs:
         with:
           node-version: 22
           cache: pnpm
-          cache-dependency-path: bench/pnpm-lock.yaml
+          cache-dependency-path: pnpm-lock.yaml
 
-      - name: Install published dependencies
+      - name: Install package set
         run: pnpm install --frozen-lockfile
 
-      - name: Typecheck public entrypoint against installed packages
-        run: pnpm run typecheck:public
-
-      - name: Test deterministic package paths
-        run: pnpm test
+      - name: Build current agent-runtime
+        run: pnpm run build
 
-      - name: Verify packed package in a clean consumer
-        run: pnpm run verify:package
+      - name: Verify agent-bench against current agent-runtime
+        run: pnpm run verify:bench
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 416bccf8..59061e70 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -40,6 +40,9 @@ jobs:
       - name: Verify packed package exports
         run: pnpm run verify:package
 
+      - name: Verify agent-bench against this release
+        run: pnpm run verify:bench
+
       - name: Verify tag/version lock
         run: |
           NPM_VERSION=$(node -p "require('./package.json').version")
@@ -98,9 +101,6 @@ jobs:
     # tag/version locking stays intact and the root package is never co-published.
     if: startsWith(github.ref, 'refs/tags/agent-bench-v')
     runs-on: ubuntu-latest
-    defaults:
-      run:
-        working-directory: bench
     steps:
       - uses: actions/checkout@v4
 
@@ -110,23 +110,17 @@ jobs:
         with:
           node-version: 22
           cache: pnpm
-          cache-dependency-path: bench/pnpm-lock.yaml
+          cache-dependency-path: pnpm-lock.yaml
 
       - name: Install deps
         run: pnpm install --frozen-lockfile
 
-      - name: Typecheck public entrypoint against installed packages
-        run: pnpm run typecheck:public
-
-      - name: Test deterministic package paths
-        run: pnpm test
-
-      - name: Verify packed package in a clean consumer
-        run: pnpm run verify:package
+      - name: Verify packed agent-bench against published dependencies
+        run: pnpm run verify:bench:published
 
       - name: Verify tag/version lock
         run: |
-          NPM_VERSION=$(node -p "require('./package.json').version")
+          NPM_VERSION=$(node -p "require('./bench/package.json').version")
           TAG_VERSION="${GITHUB_REF#refs/tags/agent-bench-v}"
           if [ "$TAG_VERSION" != "$NPM_VERSION" ]; then
             echo "::error::Tag/version mismatch: tag=$TAG_VERSION package=$NPM_VERSION."
@@ -141,9 +135,6 @@ jobs:
     permissions:
       contents: read
       id-token: write
-    defaults:
-      run:
-        working-directory: bench
     steps:
       - uses: actions/checkout@v4
 
@@ -153,7 +144,7 @@ jobs:
         with:
           node-version: 22
           cache: pnpm
-          cache-dependency-path: bench/pnpm-lock.yaml
+          cache-dependency-path: pnpm-lock.yaml
 
       - run: pnpm install --frozen-lockfile
 
@@ -162,10 +153,14 @@ jobs:
           # Requires the npmjs Trusted Publisher on @tangle-network/agent-bench:
           # org tangle-network, repo agent-runtime, workflow publish.yml.
           npm install -g npm@11
-          NAME=$(node -p "require('./package.json').name")
-          VERSION=$(node -p "require('./package.json').version")
+          NAME=$(node -p "require('./bench/package.json').name")
+          VERSION=$(node -p "require('./bench/package.json').version")
           if npm view "$NAME@$VERSION" version >/dev/null 2>&1; then
             echo "$NAME@$VERSION already on registry; skipping publish"
           else
-            npm publish --provenance --access public
+            mkdir -p "$RUNNER_TEMP/agent-bench-package"
+            pnpm --dir bench pack --pack-destination "$RUNNER_TEMP/agent-bench-package"
+            package="$(find "$RUNNER_TEMP/agent-bench-package" -maxdepth 1 -name '*.tgz' -print -quit)"
+            test -n "$package"
+            npm publish "$package" --provenance --access public
           fi
diff --git a/bench/package.json b/bench/package.json
index c29a8aeb..2f83e4fa 100644
--- a/bench/package.json
+++ b/bench/package.json
@@ -24,23 +24,19 @@
     "test": "node scripts/run-package-tests.mjs",
     "typecheck:public": "tsc -p tsconfig.public.json",
     "verify:package": "node scripts/verify-packed-consumer.mjs",
+    "verify:package:local-runtime": "node scripts/verify-packed-consumer.mjs --local-runtime",
     "verify:pier": "tsx scripts/verify-pier-pair.mts"
   },
   "dependencies": {
-    "@tangle-network/agent-eval": "^0.117.1",
-    "@tangle-network/agent-interface": "^0.25.0",
-    "@tangle-network/agent-runtime": "0.94.9",
-    "@tangle-network/sandbox": "^0.10.3"
+    "@tangle-network/agent-eval": "0.121.0",
+    "@tangle-network/agent-interface": "0.28.0",
+    "@tangle-network/agent-runtime": "workspace:*",
+    "@tangle-network/sandbox": "^0.10.5"
   },
   "devDependencies": {
-    "@types/node": "^25.0.0",
-    "tsx": "^4.19.0",
-    "typescript": "^6.0.3"
-  },
-  "pnpm": {
-    "overrides": {
-      "@tangle-network/agent-eval": "0.117.1"
-    }
+    "@types/node": "^25.9.3",
+    "tsx": "^4.22.4",
+    "typescript": "^5.7.0"
   },
   "files": [
     "src",
diff --git a/bench/pier_agents/tangle_candidate_test.py b/bench/pier_agents/tangle_candidate_test.py
index a759a3d6..cc77be39 100644
--- a/bench/pier_agents/tangle_candidate_test.py
+++ b/bench/pier_agents/tangle_candidate_test.py
@@ -4,6 +4,7 @@
 import importlib
 import json
 import os
+import shlex
 import shutil
 import subprocess
 import sys
@@ -965,7 +966,7 @@ def test_enforces_exact_signed_timeout(self):
         with tempfile.TemporaryDirectory() as directory:
             fixture = _fixture(Path(directory))
             fixture["plan"]["launch"]["env"] = {
-                "FIXTURE_SLEEP": {"kind": "public", "value": "0.25"}
+                "FIXTURE_SLEEP": {"kind": "public", "value": "2"}
             }
             fixture["plan"]["limits"]["timeoutMs"] = 10
             _rewrite_signed_plan(fixture)
@@ -976,9 +977,17 @@ def test_enforces_exact_signed_timeout(self):
             started = time.monotonic()
             with self.assertRaises(asyncio.TimeoutError):
                 asyncio.run(agent.run("make the task ready", environment, context))
-            self.assertLess(time.monotonic() - started, 0.2)
+            wall_elapsed = time.monotonic() - started
+            candidate_command = next(
+                command
+                for command in environment.commands
+                if "/run-" in command and "/timeout-" in command
+            )
+            self.assertEqual(shlex.split(candidate_command)[4], "0.010")
             self.assertEqual(context.metadata["termination"]["kind"], "timeout")
             self.assertEqual(context.metadata["termination"]["timeoutMs"], 10)
+            self.assertLess(context.metadata["observedElapsedMs"], 250)
+            self.assertLess(wall_elapsed, 1)
 
     def test_profile_cleanup_does_not_follow_candidate_links(self):
         for attack in ("symlink", "hardlink"):
diff --git a/bench/pnpm-lock.yaml b/bench/pnpm-lock.yaml
deleted file mode 100644
index 94d8d185..00000000
--- a/bench/pnpm-lock.yaml
+++ /dev/null
@@ -1,958 +0,0 @@
-lockfileVersion: '9.0'
-
-settings:
-  autoInstallPeers: true
-  excludeLinksFromLockfile: false
-
-overrides:
-  '@tangle-network/agent-eval': 0.117.1
-
-importers:
-
-  .:
-    dependencies:
-      '@tangle-network/agent-eval':
-        specifier: 0.117.1
-        version: 0.117.1(typescript@6.0.3)
-      '@tangle-network/agent-interface':
-        specifier: ^0.25.0
-        version: 0.25.0
-      '@tangle-network/agent-runtime':
-        specifier: 0.94.9
-        version: 0.94.9(@tangle-network/agent-eval@0.117.1(typescript@6.0.3))(@tangle-network/agent-interface@0.25.0)(@tangle-network/sandbox@0.10.3(viem@2.52.0(typescript@6.0.3)(zod@4.4.3)))(typescript@6.0.3)
-      '@tangle-network/sandbox':
-        specifier: ^0.10.3
-        version: 0.10.3(viem@2.52.0(typescript@6.0.3)(zod@4.4.3))
-    devDependencies:
-      '@types/node':
-        specifier: ^25.0.0
-        version: 25.9.5
-      tsx:
-        specifier: ^4.19.0
-        version: 4.22.4
-      typescript:
-        specifier: ^6.0.3
-        version: 6.0.3
-
-packages:
-
-  '@adraffy/ens-normalize@1.11.1':
-    resolution: {integrity: sha512-nhCBV3quEgesuf7c7KYfperqSS14T8bYuvJ8PcLJp6znkZpFc0AuW4qBtr8eKVyPPe/8RSr7sglCWPU5eaxwKQ==}
-
-  '@asteasolutions/zod-to-openapi@8.5.0':
-    resolution: {integrity: sha512-SABbKiObg5dLRiTFnqiW1WWwGcg1BJfmHtT2asIBnBHg6Smy/Ms2KHc650+JI4Hw7lSkdiNebEGXpwoxfben8Q==}
-    peerDependencies:
-      zod: ^4.0.0
-
-  '@ax-llm/ax@19.0.45':
-    resolution: {integrity: sha512-zC/OqUutTV0omOuAxBgeCihmxP3e1eUcbA78FCaMF3pPXZ3/FgQDZf0R04LH/lVIxm5BH4/qDdNNYcIVGwii8Q==}
-    hasBin: true
-    peerDependencies:
-      zod: ^3.24.0 || ^4.0.0
-    peerDependenciesMeta:
-      zod:
-        optional: true
-
-  '@esbuild/aix-ppc64@0.28.0':
-    resolution: {integrity: sha512-lhRUCeuOyJQURhTxl4WkpFTjIsbDayJHih5kZC1giwE+MhIzAb7mEsQMqMf18rHLsrb5qI1tafG20mLxEWcWlA==}
-    engines: {node: '>=18'}
-    cpu: [ppc64]
-    os: [aix]
-
-  '@esbuild/android-arm64@0.28.0':
-    resolution: {integrity: sha512-+WzIXQOSaGs33tLEgYPYe/yQHf0WTU0X42Jca3y8NWMbUVhp7rUnw+vAsRC/QiDrdD31IszMrZy+qwPOPjd+rw==}
-    engines: {node: '>=18'}
-    cpu: [arm64]
-    os: [android]
-
-  '@esbuild/android-arm@0.28.0':
-    resolution: {integrity: sha512-wqh0ByljabXLKHeWXYLqoJ5jKC4XBaw6Hk08OfMrCRd2nP2ZQ5eleDZC41XHyCNgktBGYMbqnrJKq/K/lzPMSQ==}
-    engines: {node: '>=18'}
-    cpu: [arm]
-    os: [android]
-
-  '@esbuild/android-x64@0.28.0':
-    resolution: {integrity: sha512-+VJggoaKhk2VNNqVL7f6S189UzShHC/mR9EE8rDdSkdpN0KflSwWY/gWjDrNxxisg8Fp1ZCD9jLMo4m0OUfeUA==}
-    engines: {node: '>=18'}
-    cpu: [x64]
-    os: [android]
-
-  '@esbuild/darwin-arm64@0.28.0':
-    resolution: {integrity: sha512-0T+A9WZm+bZ84nZBtk1ckYsOvyA3x7e2Acj1KdVfV4/2tdG4fzUp91YHx+GArWLtwqp77pBXVCPn2We7Letr0Q==}
-    engines: {node: '>=18'}
-    cpu: [arm64]
-    os: [darwin]
-
-  '@esbuild/darwin-x64@0.28.0':
-    resolution: {integrity: sha512-fyzLm/DLDl/84OCfp2f/XQ4flmORsjU7VKt8HLjvIXChJoFFOIL6pLJPH4Yhd1n1gGFF9mPwtlN5Wf82DZs+LQ==}
-    engines: {node: '>=18'}
-    cpu: [x64]
-    os: [darwin]
-
-  '@esbuild/freebsd-arm64@0.28.0':
-    resolution: {integrity: sha512-l9GeW5UZBT9k9brBYI+0WDffcRxgHQD8ShN2Ur4xWq/NFzUKm3k5lsH4PdaRgb2w7mI9u61nr2gI2mLI27Nh3Q==}
-    engines: {node: '>=18'}
-    cpu: [arm64]
-    os: [freebsd]
-
-  '@esbuild/freebsd-x64@0.28.0':
-    resolution: {integrity: sha512-BXoQai/A0wPO6Es3yFJ7APCiKGc1tdAEOgeTNy3SsB491S3aHn4S4r3e976eUnPdU+NbdtmBuLncYir2tMU9Nw==}
-    engines: {node: '>=18'}
-    cpu: [x64]
-    os: [freebsd]
-
-  '@esbuild/linux-arm64@0.28.0':
-    resolution: {integrity: sha512-RVyzfb3FWsGA55n6WY0MEIEPURL1FcbhFE6BffZEMEekfCzCIMtB5yyDcFnVbTnwk+CLAgTujmV/Lgvih56W+A==}
-    engines: {node: '>=18'}
-    cpu: [arm64]
-    os: [linux]
-
-  '@esbuild/linux-arm@0.28.0':
-    resolution: {integrity: sha512-CjaaREJagqJp7iTaNQjjidaNbCKYcd4IDkzbwwxtSvjI7NZm79qiHc8HqciMddQ6CKvJT6aBd8lO9kN/ZudLlw==}
-    engines: {node: '>=18'}
-    cpu: [arm]
-    os: [linux]
-
-  '@esbuild/linux-ia32@0.28.0':
-    resolution: {integrity: sha512-KBnSTt1kxl9x70q+ydterVdl+Cn0H18ngRMRCEQfrbqdUuntQQ0LoMZv47uB97NljZFzY6HcfqEZ2SAyIUTQBQ==}
-    engines: {node: '>=18'}
-    cpu: [ia32]
-    os: [linux]
-
-  '@esbuild/linux-loong64@0.28.0':
-    resolution: {integrity: sha512-zpSlUce1mnxzgBADvxKXX5sl8aYQHo2ezvMNI8I0lbblJtp8V4odlm3Yzlj7gPyt3T8ReksE6bK+pT3WD+aJRg==}
-    engines: {node: '>=18'}
-    cpu: [loong64]
-    os: [linux]
-
-  '@esbuild/linux-mips64el@0.28.0':
-    resolution: {integrity: sha512-2jIfP6mmjkdmeTlsX/9vmdmhBmKADrWqN7zcdtHIeNSCH1SqIoNI63cYsjQR8J+wGa4Y5izRcSHSm8K3QWmk3w==}
-    engines: {node: '>=18'}
-    cpu: [mips64el]
-    os: [linux]
-
-  '@esbuild/linux-ppc64@0.28.0':
-    resolution: {integrity: sha512-bc0FE9wWeC0WBm49IQMPSPILRocGTQt3j5KPCA8os6VprfuJ7KD+5PzESSrJ6GmPIPJK965ZJHTUlSA6GNYEhg==}
-    engines: {node: '>=18'}
-    cpu: [ppc64]
-    os: [linux]
-
-  '@esbuild/linux-riscv64@0.28.0':
-    resolution: {integrity: sha512-SQPZOwoTTT/HXFXQJG/vBX8sOFagGqvZyXcgLA3NhIqcBv1BJU1d46c0rGcrij2B56Z2rNiSLaZOYW5cUk7yLQ==}
-    engines: {node: '>=18'}
-    cpu: [riscv64]
-    os: [linux]
-
-  '@esbuild/linux-s390x@0.28.0':
-    resolution: {integrity: sha512-SCfR0HN8CEEjnYnySJTd2cw0k9OHB/YFzt5zgJEwa+wL/T/raGWYMBqwDNAC6dqFKmJYZoQBRfHjgwLHGSrn3Q==}
-    engines: {node: '>=18'}
-    cpu: [s390x]
-    os: [linux]
-
-  '@esbuild/linux-x64@0.28.0':
-    resolution: {integrity: sha512-us0dSb9iFxIi8srnpl931Nvs65it/Jd2a2K3qs7fz2WfGPHqzfzZTfec7oxZJRNPXPnNYZtanmRc4AL/JwVzHQ==}
-    engines: {node: '>=18'}
-    cpu: [x64]
-    os: [linux]
-
-  '@esbuild/netbsd-arm64@0.28.0':
-    resolution: {integrity: sha512-CR/RYotgtCKwtftMwJlUU7xCVNg3lMYZ0RzTmAHSfLCXw3NtZtNpswLEj/Kkf6kEL3Gw+BpOekRX0BYCtklhUw==}
-    engines: {node: '>=18'}
-    cpu: [arm64]
-    os: [netbsd]
-
-  '@esbuild/netbsd-x64@0.28.0':
-    resolution: {integrity: sha512-nU1yhmYutL+fQ71Kxnhg8uEOdC0pwEW9entHykTgEbna2pw2dkbFSMeqjjyHZoCmt8SBkOSvV+yNmm94aUrrqw==}
-    engines: {node: '>=18'}
-    cpu: [x64]
-    os: [netbsd]
-
-  '@esbuild/openbsd-arm64@0.28.0':
-    resolution: {integrity: sha512-cXb5vApOsRsxsEl4mcZ1XY3D4DzcoMxR/nnc4IyqYs0rTI8ZKmW6kyyg+11Z8yvgMfAEldKzP7AdP64HnSC/6g==}
-    engines: {node: '>=18'}
-    cpu: [arm64]
-    os: [openbsd]
-
-  '@esbuild/openbsd-x64@0.28.0':
-    resolution: {integrity: sha512-8wZM2qqtv9UP3mzy7HiGYNH/zjTA355mpeuA+859TyR+e+Tc08IHYpLJuMsfpDJwoLo1ikIJI8jC3GFjnRClzA==}
-    engines: {node: '>=18'}
-    cpu: [x64]
-    os: [openbsd]
-
-  '@esbuild/openharmony-arm64@0.28.0':
-    resolution: {integrity: sha512-FLGfyizszcef5C3YtoyQDACyg95+dndv79i2EekILBofh5wpCa1KuBqOWKrEHZg3zrL3t5ouE5jgr94vA+Wb2w==}
-    engines: {node: '>=18'}
-    cpu: [arm64]
-    os: [openharmony]
-
-  '@esbuild/sunos-x64@0.28.0':
-    resolution: {integrity: sha512-1ZgjUoEdHZZl/YlV76TSCz9Hqj9h9YmMGAgAPYd+q4SicWNX3G5GCyx9uhQWSLcbvPW8Ni7lj4gDa1T40akdlw==}
-    engines: {node: '>=18'}
-    cpu: [x64]
-    os: [sunos]
-
-  '@esbuild/win32-arm64@0.28.0':
-    resolution: {integrity: sha512-Q9StnDmQ/enxnpxCCLSg0oo4+34B9TdXpuyPeTedN/6+iXBJ4J+zwfQI28u/Jl40nOYAxGoNi7mFP40RUtkmUA==}
-    engines: {node: '>=18'}
-    cpu: [arm64]
-    os: [win32]
-
-  '@esbuild/win32-ia32@0.28.0':
-    resolution: {integrity: sha512-zF3ag/gfiCe6U2iczcRzSYJKH1DCI+ByzSENHlM2FcDbEeo5Zd2C86Aq0tKUYAJJ1obRP84ymxIAksZUcdztHA==}
-    engines: {node: '>=18'}
-    cpu: [ia32]
-    os: [win32]
-
-  '@esbuild/win32-x64@0.28.0':
-    resolution: {integrity: sha512-pEl1bO9mfAmIC+tW5btTmrKaujg3zGtUmWNdCw/xs70FBjwAL3o9OEKNHvNmnyylD6ubxUERiEhdsL0xBQ9efw==}
-    engines: {node: '>=18'}
-    cpu: [x64]
-    os: [win32]
-
-  '@hono/node-server@2.0.4':
-    resolution: {integrity: sha512-Ut3y0dMMPWy6bZ2kVfx25EOVbZlm15dhF4mOsezMlhpNHy+4MkU1qN9Y6lnruYi4wPmFzimGX2X7LF/FwHli4A==}
-    engines: {node: '>=20'}
-    peerDependencies:
-      hono: ^4
-
-  '@noble/ciphers@1.3.0':
-    resolution: {integrity: sha512-2I0gnIVPtfnMw9ee9h1dJG7tp81+8Ob3OJb3Mv37rx5L40/b0i7djjCVvGOVqc9AEIQyvyu1i6ypKdFw8R8gQw==}
-    engines: {node: ^14.21.3 || >=16}
-
-  '@noble/curves@1.9.1':
-    resolution: {integrity: sha512-k11yZxZg+t+gWvBbIswW0yoJlu8cHOC7dhunwOzoWH/mXGBiYyR4YY6hAEK/3EUs4UpB8la1RfdRpeGsFHkWsA==}
-    engines: {node: ^14.21.3 || >=16}
-
-  '@noble/curves@2.2.0':
-    resolution: {integrity: sha512-T/BoHgFXirb0ENSPBquzX0rcjXeM6Lo892a2jlYJkqk83LqZx0l1Of7DzlKJ6jkpvMrkHSnAcgb5JegL8SeIkQ==}
-    engines: {node: '>= 20.19.0'}
-
-  '@noble/hashes@1.8.0':
-    resolution: {integrity: sha512-jCs9ldd7NwzpgXDIf6P3+NrHh9/sD6CQdxHyjQI+h/6rDNo88ypBxxz45UDuZHz9r3tNz7N/VInSVoVdtXEI4A==}
-    engines: {node: ^14.21.3 || >=16}
-
-  '@noble/hashes@2.2.0':
-    resolution: {integrity: sha512-IYqDGiTXab6FniAgnSdZwgWbomxpy9FtYvLKs7wCUs2a8RkITG+DFGO1DM9cr+E3/RgADRpFjrKVaJ1z6sjtEg==}
-    engines: {node: '>= 20.19.0'}
-
-  '@opentelemetry/api@1.9.1':
-    resolution: {integrity: sha512-gLyJlPHPZYdAk1JENA9LeHejZe1Ti77/pTeFm/nMXmQH/HFZlcS/O2XJB+L8fkbrNSqhdtlvjBVjxwUYanNH5Q==}
-    engines: {node: '>=8.0.0'}
-
-  '@scure/base@1.2.6':
-    resolution: {integrity: sha512-g/nm5FgUa//MCj1gV09zTJTaM6KBAHqLN907YVQqf7zC49+DcO4B1so4ZX07Ef10Twr6nuqYEH9GEggFXA4Fmg==}
-
-  '@scure/base@2.2.0':
-    resolution: {integrity: sha512-b8XEupJibegiXV+tDUseI8oLQc8ei3d/4Jkb2RpbHh3MfE054ov3uIz2dhFkB3FI8iwYkEh0gGCApkrYggkPNg==}
-
-  '@scure/bip32@1.7.0':
-    resolution: {integrity: sha512-E4FFX/N3f4B80AKWp5dP6ow+flD1LQZo/w8UnLGYZO674jS6YnYeepycOOksv+vLPSpgN35wgKgy+ybfTb2SMw==}
-
-  '@scure/bip32@2.2.0':
-    resolution: {integrity: sha512-zFr7t2F+a9+5tB7QbarF2HQNYrgjCNaoLAupZdKkrFMYMozJf5zqH2WJCQibMzm1qQ0QogrxVGO3qXfQDYMaQg==}
-
-  '@scure/bip39@1.6.0':
-    resolution: {integrity: sha512-+lF0BbLiJNwVlev4eKelw1WWLaiKXw7sSl8T6FvBlWkdX+94aGJ4o8XjUdlyhTCjd8c+B3KT3JfS8P0bLRNU6A==}
-
-  '@scure/bip39@2.2.0':
-    resolution: {integrity: sha512-T/Bj/YvYMNkIPq6EENO6/rcs2e7qTNuyoUXf0KBFDmp0ZDu0H2X4Lq6yC3i0c8PcWkov5EbW+yQZZbdMmk154A==}
-
-  '@tangle-network/agent-core@0.3.4':
-    resolution: {integrity: sha512-Hvz3ABRouNtBmRvGqPxifAO2yuILneJMylWH5jW/jeS2F03RvqkGYuXyGXWWLqosYbb3hVAvSEe4Ykm2FMGEDQ==}
-
-  '@tangle-network/agent-core@0.4.9':
-    resolution: {integrity: sha512-BFU4WV12Y6D28PX+o8LIVkIMD+cXwrL1uyV182l12Bn9zEu7VHt2oVMEEzbsN8L+X0xM9baEfMCHTFKFNJv7Aw==}
-
-  '@tangle-network/agent-eval@0.117.1':
-    resolution: {integrity: sha512-mtBdGhGIC+nrPZbseLQo0Iako1LR66ZeIr2UjVBJcNj/lHYfib2jbhdWy3RmbQcHR0Cxdqqydn7pLJ5dSXVYtQ==}
-    engines: {node: '>=20'}
-    hasBin: true
-
-  '@tangle-network/agent-interface@0.13.0':
-    resolution: {integrity: sha512-CeTPGRLoXqpt0h+BCyFgZPkfU1zyRpWmqfD+85i/uk+uvbqxkfI+JprfKVf3tBsQuCgJPSjPt5qjdW8n3h2BVg==}
-
-  '@tangle-network/agent-interface@0.14.0':
-    resolution: {integrity: sha512-9CyGhIpl90E7v4MTm3b1ti3Bp7BfPigk2Nafgi21Lg0U+QxlNB656F2JmVpUuSbOo9aGZPtg5nXu5EBTlV5a1g==}
-
-  '@tangle-network/agent-interface@0.21.0':
-    resolution: {integrity: sha512-jDxhVJgxymrvU1RLWxWKueuaWQIpBAfrW8BuVTB5m2Y4eFMLo1SawDBEMDLdZN4/Gf34xrFrsRk+PAj9brGKMQ==}
-
-  '@tangle-network/agent-interface@0.22.0':
-    resolution: {integrity: sha512-7fsJhNdvTmOB1X9E2owl06jzyrqaN+jMkOPVKbK7dvNqQvf627PowCtL/edhUqEEv+K0FWtkwG3R3xqjX7VNZg==}
-
-  '@tangle-network/agent-interface@0.24.0':
-    resolution: {integrity: sha512-iaHWNTYne49cBYXwb72NjGDw5bjT2KVJlfawXygbvkqnfUsQG57BS++Yjf/XYvwJJdDeEaGs+Xvy8SWYkvu0qA==}
-
-  '@tangle-network/agent-interface@0.25.0':
-    resolution: {integrity: sha512-bMKjyjk8bk2651HleiTEUOTPgNE2bWxCEl9VBVUyub/UkP87Y09byXOxeDEsNGc4TIT19Y5Uo0b+k1kRWXkMVQ==}
-
-  '@tangle-network/agent-knowledge@1.12.1':
-    resolution: {integrity: sha512-j5riyIvz5C+ZBELTtNVtmUbcKAMpgTHMNmDQ12MhOIPJv5Y7AeX2oCWuF6c8aPqg1hOrsLqgqXSr1zkV0ePsrA==}
-    engines: {node: '>=20'}
-    hasBin: true
-
-  '@tangle-network/agent-profile-materialize@0.3.2':
-    resolution: {integrity: sha512-jCj1Hc/brPQ5eS7or9A+Klv0FAZcGtyFr7kx2cKfk+wCj0/4Di3k1pMjayrmRNgI/7Oc47gt6/t+qvdtZxBkAw==}
-
-  '@tangle-network/agent-runtime@0.94.9':
-    resolution: {integrity: sha512-p0mKK4DbzmrVVQPhNTH+Si7mpIV5+/0avhGVNzJcq2YdcH0YWYW2+OUy86mhXkxv/vVNkdR70Sco91hZuoC6ig==}
-    engines: {node: '>=20'}
-    hasBin: true
-    peerDependencies:
-      '@tangle-network/agent-eval': 0.117.1
-      '@tangle-network/agent-interface': '>=0.25.0 <0.26.0'
-      '@tangle-network/sandbox': '>=0.8.0 <1.0.0'
-      playwright: ^1.40.0
-    peerDependenciesMeta:
-      '@tangle-network/sandbox':
-        optional: true
-      playwright:
-        optional: true
-
-  '@tangle-network/sandbox@0.10.3':
-    resolution: {integrity: sha512-3nZnIaXc/vCH3Lb6SCZA3cYEJT4+ThGXLPkGo5z2kh434n3eJryk/PdnSXSwPoA6xuTQToJAPbGGIdMEHFU/Vw==}
-    peerDependencies:
-      '@mastra/core': ^1.36.0
-      '@modelcontextprotocol/sdk': ^1.29.0
-      ai: ^6.0.175
-      openai: ^6.36.0
-      viem: ^2.0.0
-    peerDependenciesMeta:
-      '@mastra/core':
-        optional: true
-      '@modelcontextprotocol/sdk':
-        optional: true
-      ai:
-        optional: true
-      openai:
-        optional: true
-      viem:
-        optional: true
-
-  '@tangle-network/sandbox@0.9.7':
-    resolution: {integrity: sha512-9pCwJ5MlF7RUpp0AQKQDFyR0yu+E0udEhWkqhrlb/RuoJxlt72zVPuzO4FnMb1MZTkfjStmomC3k5xQyqi1YSA==}
-    peerDependencies:
-      '@mastra/core': ^1.36.0
-      '@modelcontextprotocol/sdk': ^1.29.0
-      ai: ^6.0.175
-      openai: ^6.36.0
-      viem: ^2.0.0
-    peerDependenciesMeta:
-      '@mastra/core':
-        optional: true
-      '@modelcontextprotocol/sdk':
-        optional: true
-      ai:
-        optional: true
-      openai:
-        optional: true
-      viem:
-        optional: true
-
-  '@tangle-network/tcloud-attestation@0.1.1':
-    resolution: {integrity: sha512-+TAF9s5t1jOWGyGHvKhIWe2FYmG7puVaxmmg0Et67ylAjGa7GqUAvISXGjG/6dzld7A170V0kQHK0WVdh2Wh0Q==}
-    engines: {node: '>=18'}
-
-  '@tangle-network/tcloud@0.4.14':
-    resolution: {integrity: sha512-jWYt//cGdLBDOv0luLH6xAGS4gbuOt8uHIkaCWwDDpQ1zp0FUPATHIrA3RMuF0qtQq9Vq00IhLrmCnHdHBP+dg==}
-    engines: {node: '>=18'}
-    hasBin: true
-
-  '@types/node@25.9.5':
-    resolution: {integrity: sha512-OScDchr2fwuUmWdf4kZ9h7PcJiYDVInhJizG/biAq3cAvqwYktuy/TYGGdZNMtNTFUP7rnb0NU4TUdm82kt4Rg==}
-
-  abitype@1.2.3:
-    resolution: {integrity: sha512-Ofer5QUnuUdTFsBRwARMoWKOH1ND5ehwYhJ3OJ/BQO+StkwQjHw0XyVh4vDttzHB7QOFhPHa/o413PJ82gU/Tg==}
-    peerDependencies:
-      typescript: '>=5.0.4'
-      zod: ^3.22.0 || ^4.0.0
-    peerDependenciesMeta:
-      typescript:
-        optional: true
-      zod:
-        optional: true
-
-  b4a@1.8.1:
-    resolution: {integrity: sha512-aiqre1Nr0B/6DgE2N5vwTc+2/oQZ4Wh1t4NznYY4E00y8LCt6NqdRv81so00oo27D8MVKTpUa/MwUUtBLXCoDw==}
-    peerDependencies:
-      react-native-b4a: '*'
-    peerDependenciesMeta:
-      react-native-b4a:
-        optional: true
-
-  bare-events@2.9.1:
-    resolution: {integrity: sha512-Z0oHEHAFDZkffN8Qc39zNZjQlMDkPJRyyyZieU1VH7u8c5S+qHZ2S8ixdKIAxEjfHO7FJxXmJWgteOghVanIsg==}
-    peerDependencies:
-      bare-abort-controller: '*'
-    peerDependenciesMeta:
-      bare-abort-controller:
-        optional: true
-
-  bare-fs@4.7.4:
-    resolution: {integrity: sha512-y1kC+ffIx/tPLdTE693uNjHfzTfr+ravR5tvWlMXe25nELbkqV400S71qHDwbkAQ1FVEZobB1NFRzFbCCcyBCQ==}
-    engines: {bare: '>=1.16.0'}
-    peerDependencies:
-      bare-buffer: '*'
-    peerDependenciesMeta:
-      bare-buffer:
-        optional: true
-
-  bare-path@3.1.1:
-    resolution: {integrity: sha512-JprUlveX3QjApC1cTpsUOiscADftCGVWkzitbHsRqv84hzYwYHw2mbluddsq5TvI8mH/8Ov1f4BiMAdcB0oYnQ==}
-
-  bare-stream@2.13.3:
-    resolution: {integrity: sha512-Kc+brLqvEqGkjyfiwJmImAOqLZL7OsoLKuavx+hJjgVV3nLTOjloJyPMFxjUPerGGHrNH0fLU06jjykMLWrERQ==}
-    peerDependencies:
-      bare-abort-controller: '*'
-      bare-buffer: '*'
-      bare-events: '*'
-    peerDependenciesMeta:
-      bare-abort-controller:
-        optional: true
-      bare-buffer:
-        optional: true
-      bare-events:
-        optional: true
-
-  bare-url@2.4.5:
-    resolution: {integrity: sha512-K+y9xF1tN+CdPu4qWwr0QiK1Al07eFPGYK5M2pDXcmHdMdgC/tT/bpmMe1hrmRHaidKLkXrC+cRNYf3XVDUhSQ==}
-
-  commander@14.0.3:
-    resolution: {integrity: sha512-H+y0Jo/T1RZ9qPP4Eh1pkcQcLRglraJaSLoyOtHxu6AapkjWVCy2Sit1QQ4x3Dng8qDlSsZEet7g5Pq06MvTgw==}
-    engines: {node: '>=20'}
-
-  dayjs@1.11.21:
-    resolution: {integrity: sha512-98IT+HOahAisibz/yjKbzuOBwYcjJ7BCLPzARyHiyEBmRz4fatF+KPJszEHXsGYjUG234aH/cOjW1wwTbKUZlA==}
-
-  esbuild@0.28.0:
-    resolution: {integrity: sha512-sNR9MHpXSUV/XB4zmsFKN+QgVG82Cc7+/aaxJ8Adi8hyOac+EXptIp45QBPaVyX3N70664wRbTcLTOemCAnyqw==}
-    engines: {node: '>=18'}
-    hasBin: true
-
-  eventemitter3@5.0.1:
-    resolution: {integrity: sha512-GWkBvjiSZK87ELrYOSESUYeVIc9mvLLf/nXalMOS5dYrgZq9o5OVkbZAVM06CVxYsCwH9BDZFPlQTlPA1j4ahA==}
-
-  events-universal@1.0.1:
-    resolution: {integrity: sha512-LUd5euvbMLpwOF8m6ivPCbhQeSiYVNb8Vs0fQ8QjXo0JTkEHpz8pxdQf0gStltaPpw0Cca8b39KxvK9cfKRiAw==}
-
-  fast-fifo@1.3.2:
-    resolution: {integrity: sha512-/d9sfos4yxzpwkDkuN7k2SqFKtYNmCTzgfEpz82x34IM9/zc8KGxQoXg1liNC/izpRM/MBdt44Nmx41ZWqk+FQ==}
-
-  fsevents@2.3.3:
-    resolution: {integrity: sha512-5xoDfX+fL7faATnagmWPpbFtwh/R77WmMMqqHGS65C3vvB0YHrgF+B1YmZ3441tMj5n63k0212XNoJwzlhffQw==}
-    engines: {node: ^8.16.0 || ^10.6.0 || >=11.0.0}
-    os: [darwin]
-
-  hono@4.12.30:
-    resolution: {integrity: sha512-emn+JoJjrN9YTpRDS5it/UI2SO9BAE37T6I3d963RxcZ81G9A4pr2SZTEiiaiKbzx+NKRg5BZ89fCL7gCJCUog==}
-    engines: {node: '>=16.9.0'}
-
-  isows@1.0.7:
-    resolution: {integrity: sha512-I1fSfDCZL5P0v33sVqeTDSpcstAg/N+wF5HS033mogOVIp4B+oHC7oOCsA3axAbBSGTJ8QubbNmnIRN/h8U7hg==}
-    peerDependencies:
-      ws: '*'
-
-  openapi3-ts@4.6.0:
-    resolution: {integrity: sha512-a4sfn6L2sIShhtzJqmjGrARvxAW/3F2BJDdyRVvNF9VhAsZSh5hSyI3a9TNvmzBxXmq66nY5LNT5bQcBxYAZZg==}
-
-  ox@0.14.27:
-    resolution: {integrity: sha512-+xhLHo/f+f4BH121/1Pomm/1vgBBda1wYiFpTvjSo8o5OcEj76Pf1hGPJiepoYMTQoTm2SKdSBvWkFWk5l07PA==}
-    peerDependencies:
-      typescript: '>=5.4.0'
-    peerDependenciesMeta:
-      typescript:
-        optional: true
-
-  streamx@2.28.0:
-    resolution: {integrity: sha512-1Yowhzjf0ivGMrTIkY9hav5TxobO9qIVqUE41fiCGMGgc3CLlf4MY+9AHmZqBWgDTue0fY9zWjYFVyf6Diuobw==}
-
-  tar-stream@3.2.0:
-    resolution: {integrity: sha512-ojzvCvVaNp6aOTFmG7jaRD0meowIAuPc3cMMhSgKiVWws1GyHbGd/xvnyuRKcKlMpt3qvxx6r0hreCNITP9hIg==}
-
-  teex@1.0.1:
-    resolution: {integrity: sha512-eYE6iEI62Ni1H8oIa7KlDU6uQBtqr4Eajni3wX7rpfXD8ysFx8z0+dri+KWEPWpBsxXfxu58x/0jvTVT1ekOSg==}
-
-  text-decoder@1.2.7:
-    resolution: {integrity: sha512-vlLytXkeP4xvEq2otHeJfSQIRyWxo/oZGEbXrtEEF9Hnmrdly59sUbzZ/QgyWuLYHctCHxFF4tRQZNQ9k60ExQ==}
-
-  tsx@4.22.4:
-    resolution: {integrity: sha512-X8EX+XV4QR5xCsrgxaED954zTDfY8KqlDtskKEL0cHhyS/P8b4IFOvGDQpsC9Q1XnLq915wEfwwY/zzskCtmhg==}
-    engines: {node: '>=18.0.0'}
-    hasBin: true
-
-  typescript@6.0.3:
-    resolution: {integrity: sha512-y2TvuxSZPDyQakkFRPZHKFm+KKVqIisdg9/CZwm9ftvKXLP8NRWj38/ODjNbr43SsoXqNuAisEf1GdCxqWcdBw==}
-    engines: {node: '>=14.17'}
-    hasBin: true
-
-  undici-types@7.24.6:
-    resolution: {integrity: sha512-WRNW+sJgj5OBN4/0JpHFqtqzhpbnV0GuB+OozA9gCL7a993SmU+1JBZCzLNxYsbMfIeDL+lTsphD5jN5N+n0zg==}
-
-  viem@2.52.0:
-    resolution: {integrity: sha512-py2QPYe9e1f4DmPJCsXF7zHmyZ0PkJrBxdQZ5dvNXvzy3UzWkUn7dNfC0TMeNm6Qv1tKw3b6qXXExpx6L0oMbw==}
-    peerDependencies:
-      typescript: '>=5.0.4'
-    peerDependenciesMeta:
-      typescript:
-        optional: true
-
-  ws@8.20.1:
-    resolution: {integrity: sha512-It4dO0K5v//JtTXuPkfEOaI3uUN87iYPnqo/ZzqCoG3g8uhA66QUMs/SrM0YK7/NAu+r4LMh/9dq2A7k+rHs+w==}
-    engines: {node: '>=10.0.0'}
-    peerDependencies:
-      bufferutil: ^4.0.1
-      utf-8-validate: '>=5.0.2'
-    peerDependenciesMeta:
-      bufferutil:
-        optional: true
-      utf-8-validate:
-        optional: true
-
-  yaml@2.9.0:
-    resolution: {integrity: sha512-2AvhNX3mb8zd6Zy7INTtSpl1F15HW6Wnqj0srWlkKLcpYl/gMIMJiyuGq2KeI2YFxUPjdlB+3Lc10seMLtL4cA==}
-    engines: {node: '>= 14.6'}
-    hasBin: true
-
-  zod@4.4.3:
-    resolution: {integrity: sha512-ytENFjIJFl2UwYglde2jchW2Hwm4GJFLDiSXWdTrJQBIN9Fcyp7n4DhxJEiWNAJMV1/BqWfW/kkg71UDcHJyTQ==}
-
-snapshots:
-
-  '@adraffy/ens-normalize@1.11.1': {}
-
-  '@asteasolutions/zod-to-openapi@8.5.0(zod@4.4.3)':
-    dependencies:
-      openapi3-ts: 4.6.0
-      zod: 4.4.3
-
-  '@ax-llm/ax@19.0.45(zod@4.4.3)':
-    dependencies:
-      '@opentelemetry/api': 1.9.1
-      dayjs: 1.11.21
-    optionalDependencies:
-      zod: 4.4.3
-
-  '@esbuild/aix-ppc64@0.28.0':
-    optional: true
-
-  '@esbuild/android-arm64@0.28.0':
-    optional: true
-
-  '@esbuild/android-arm@0.28.0':
-    optional: true
-
-  '@esbuild/android-x64@0.28.0':
-    optional: true
-
-  '@esbuild/darwin-arm64@0.28.0':
-    optional: true
-
-  '@esbuild/darwin-x64@0.28.0':
-    optional: true
-
-  '@esbuild/freebsd-arm64@0.28.0':
-    optional: true
-
-  '@esbuild/freebsd-x64@0.28.0':
-    optional: true
-
-  '@esbuild/linux-arm64@0.28.0':
-    optional: true
-
-  '@esbuild/linux-arm@0.28.0':
-    optional: true
-
-  '@esbuild/linux-ia32@0.28.0':
-    optional: true
-
-  '@esbuild/linux-loong64@0.28.0':
-    optional: true
-
-  '@esbuild/linux-mips64el@0.28.0':
-    optional: true
-
-  '@esbuild/linux-ppc64@0.28.0':
-    optional: true
-
-  '@esbuild/linux-riscv64@0.28.0':
-    optional: true
-
-  '@esbuild/linux-s390x@0.28.0':
-    optional: true
-
-  '@esbuild/linux-x64@0.28.0':
-    optional: true
-
-  '@esbuild/netbsd-arm64@0.28.0':
-    optional: true
-
-  '@esbuild/netbsd-x64@0.28.0':
-    optional: true
-
-  '@esbuild/openbsd-arm64@0.28.0':
-    optional: true
-
-  '@esbuild/openbsd-x64@0.28.0':
-    optional: true
-
-  '@esbuild/openharmony-arm64@0.28.0':
-    optional: true
-
-  '@esbuild/sunos-x64@0.28.0':
-    optional: true
-
-  '@esbuild/win32-arm64@0.28.0':
-    optional: true
-
-  '@esbuild/win32-ia32@0.28.0':
-    optional: true
-
-  '@esbuild/win32-x64@0.28.0':
-    optional: true
-
-  '@hono/node-server@2.0.4(hono@4.12.30)':
-    dependencies:
-      hono: 4.12.30
-
-  '@noble/ciphers@1.3.0': {}
-
-  '@noble/curves@1.9.1':
-    dependencies:
-      '@noble/hashes': 1.8.0
-
-  '@noble/curves@2.2.0':
-    dependencies:
-      '@noble/hashes': 2.2.0
-
-  '@noble/hashes@1.8.0': {}
-
-  '@noble/hashes@2.2.0': {}
-
-  '@opentelemetry/api@1.9.1': {}
-
-  '@scure/base@1.2.6': {}
-
-  '@scure/base@2.2.0': {}
-
-  '@scure/bip32@1.7.0':
-    dependencies:
-      '@noble/curves': 1.9.1
-      '@noble/hashes': 1.8.0
-      '@scure/base': 1.2.6
-
-  '@scure/bip32@2.2.0':
-    dependencies:
-      '@noble/curves': 2.2.0
-      '@noble/hashes': 2.2.0
-      '@scure/base': 2.2.0
-
-  '@scure/bip39@1.6.0':
-    dependencies:
-      '@noble/hashes': 1.8.0
-      '@scure/base': 1.2.6
-
-  '@scure/bip39@2.2.0':
-    dependencies:
-      '@noble/hashes': 2.2.0
-      '@scure/base': 2.2.0
-
-  '@tangle-network/agent-core@0.3.4':
-    dependencies:
-      '@tangle-network/agent-interface': 0.14.0
-      zod: 4.4.3
-
-  '@tangle-network/agent-core@0.4.9':
-    dependencies:
-      '@tangle-network/agent-interface': 0.24.0
-      zod: 4.4.3
-
-  '@tangle-network/agent-eval@0.117.1(typescript@6.0.3)':
-    dependencies:
-      '@asteasolutions/zod-to-openapi': 8.5.0(zod@4.4.3)
-      '@ax-llm/ax': 19.0.45(zod@4.4.3)
-      '@hono/node-server': 2.0.4(hono@4.12.30)
-      '@tangle-network/agent-interface': 0.22.0
-      '@tangle-network/tcloud': 0.4.14(typescript@6.0.3)(zod@4.4.3)
-      hono: 4.12.30
-      zod: 4.4.3
-    transitivePeerDependencies:
-      - '@mastra/core'
-      - '@modelcontextprotocol/sdk'
-      - ai
-      - bufferutil
-      - openai
-      - typescript
-      - utf-8-validate
-
-  '@tangle-network/agent-interface@0.13.0':
-    dependencies:
-      zod: 4.4.3
-
-  '@tangle-network/agent-interface@0.14.0':
-    dependencies:
-      zod: 4.4.3
-
-  '@tangle-network/agent-interface@0.21.0':
-    dependencies:
-      zod: 4.4.3
-
-  '@tangle-network/agent-interface@0.22.0':
-    dependencies:
-      zod: 4.4.3
-
-  '@tangle-network/agent-interface@0.24.0':
-    dependencies:
-      zod: 4.4.3
-
-  '@tangle-network/agent-interface@0.25.0':
-    dependencies:
-      zod: 4.4.3
-
-  '@tangle-network/agent-knowledge@1.12.1(typescript@6.0.3)':
-    dependencies:
-      '@tangle-network/agent-eval': 0.117.1(typescript@6.0.3)
-      zod: 4.4.3
-    transitivePeerDependencies:
-      - '@mastra/core'
-      - '@modelcontextprotocol/sdk'
-      - ai
-      - bufferutil
-      - openai
-      - typescript
-      - utf-8-validate
-
-  '@tangle-network/agent-profile-materialize@0.3.2':
-    dependencies:
-      '@tangle-network/agent-interface': 0.25.0
-
-  '@tangle-network/agent-runtime@0.94.9(@tangle-network/agent-eval@0.117.1(typescript@6.0.3))(@tangle-network/agent-interface@0.25.0)(@tangle-network/sandbox@0.10.3(viem@2.52.0(typescript@6.0.3)(zod@4.4.3)))(typescript@6.0.3)':
-    dependencies:
-      '@tangle-network/agent-eval': 0.117.1(typescript@6.0.3)
-      '@tangle-network/agent-interface': 0.25.0
-      '@tangle-network/agent-knowledge': 1.12.1(typescript@6.0.3)
-      '@tangle-network/agent-profile-materialize': 0.3.2
-      tar-stream: 3.2.0
-    optionalDependencies:
-      '@tangle-network/sandbox': 0.10.3(viem@2.52.0(typescript@6.0.3)(zod@4.4.3))
-    transitivePeerDependencies:
-      - '@mastra/core'
-      - '@modelcontextprotocol/sdk'
-      - ai
-      - bare-abort-controller
-      - bare-buffer
-      - bufferutil
-      - openai
-      - react-native-b4a
-      - typescript
-      - utf-8-validate
-
-  '@tangle-network/sandbox@0.10.3(viem@2.52.0(typescript@6.0.3)(zod@4.4.3))':
-    dependencies:
-      '@tangle-network/agent-core': 0.4.9
-      '@tangle-network/agent-interface': 0.21.0
-    optionalDependencies:
-      viem: 2.52.0(typescript@6.0.3)(zod@4.4.3)
-
-  '@tangle-network/sandbox@0.9.7(viem@2.52.0(typescript@6.0.3)(zod@4.4.3))':
-    dependencies:
-      '@tangle-network/agent-core': 0.3.4
-      '@tangle-network/agent-interface': 0.13.0
-    optionalDependencies:
-      viem: 2.52.0(typescript@6.0.3)(zod@4.4.3)
-
-  '@tangle-network/tcloud-attestation@0.1.1': {}
-
-  '@tangle-network/tcloud@0.4.14(typescript@6.0.3)(zod@4.4.3)':
-    dependencies:
-      '@scure/bip32': 2.2.0
-      '@scure/bip39': 2.2.0
-      '@tangle-network/sandbox': 0.9.7(viem@2.52.0(typescript@6.0.3)(zod@4.4.3))
-      '@tangle-network/tcloud-attestation': 0.1.1
-      commander: 14.0.3
-      viem: 2.52.0(typescript@6.0.3)(zod@4.4.3)
-    transitivePeerDependencies:
-      - '@mastra/core'
-      - '@modelcontextprotocol/sdk'
-      - ai
-      - bufferutil
-      - openai
-      - typescript
-      - utf-8-validate
-      - zod
-
-  '@types/node@25.9.5':
-    dependencies:
-      undici-types: 7.24.6
-
-  abitype@1.2.3(typescript@6.0.3)(zod@4.4.3):
-    optionalDependencies:
-      typescript: 6.0.3
-      zod: 4.4.3
-
-  b4a@1.8.1: {}
-
-  bare-events@2.9.1: {}
-
-  bare-fs@4.7.4:
-    dependencies:
-      bare-events: 2.9.1
-      bare-path: 3.1.1
-      bare-stream: 2.13.3(bare-events@2.9.1)
-      bare-url: 2.4.5
-      fast-fifo: 1.3.2
-    transitivePeerDependencies:
-      - bare-abort-controller
-      - react-native-b4a
-
-  bare-path@3.1.1: {}
-
-  bare-stream@2.13.3(bare-events@2.9.1):
-    dependencies:
-      b4a: 1.8.1
-      streamx: 2.28.0
-      teex: 1.0.1
-    optionalDependencies:
-      bare-events: 2.9.1
-    transitivePeerDependencies:
-      - react-native-b4a
-
-  bare-url@2.4.5:
-    dependencies:
-      bare-path: 3.1.1
-
-  commander@14.0.3: {}
-
-  dayjs@1.11.21: {}
-
-  esbuild@0.28.0:
-    optionalDependencies:
-      '@esbuild/aix-ppc64': 0.28.0
-      '@esbuild/android-arm': 0.28.0
-      '@esbuild/android-arm64': 0.28.0
-      '@esbuild/android-x64': 0.28.0
-      '@esbuild/darwin-arm64': 0.28.0
-      '@esbuild/darwin-x64': 0.28.0
-      '@esbuild/freebsd-arm64': 0.28.0
-      '@esbuild/freebsd-x64': 0.28.0
-      '@esbuild/linux-arm': 0.28.0
-      '@esbuild/linux-arm64': 0.28.0
-      '@esbuild/linux-ia32': 0.28.0
-      '@esbuild/linux-loong64': 0.28.0
-      '@esbuild/linux-mips64el': 0.28.0
-      '@esbuild/linux-ppc64': 0.28.0
-      '@esbuild/linux-riscv64': 0.28.0
-      '@esbuild/linux-s390x': 0.28.0
-      '@esbuild/linux-x64': 0.28.0
-      '@esbuild/netbsd-arm64': 0.28.0
-      '@esbuild/netbsd-x64': 0.28.0
-      '@esbuild/openbsd-arm64': 0.28.0
-      '@esbuild/openbsd-x64': 0.28.0
-      '@esbuild/openharmony-arm64': 0.28.0
-      '@esbuild/sunos-x64': 0.28.0
-      '@esbuild/win32-arm64': 0.28.0
-      '@esbuild/win32-ia32': 0.28.0
-      '@esbuild/win32-x64': 0.28.0
-
-  eventemitter3@5.0.1: {}
-
-  events-universal@1.0.1:
-    dependencies:
-      bare-events: 2.9.1
-    transitivePeerDependencies:
-      - bare-abort-controller
-
-  fast-fifo@1.3.2: {}
-
-  fsevents@2.3.3:
-    optional: true
-
-  hono@4.12.30: {}
-
-  isows@1.0.7(ws@8.20.1):
-    dependencies:
-      ws: 8.20.1
-
-  openapi3-ts@4.6.0:
-    dependencies:
-      yaml: 2.9.0
-
-  ox@0.14.27(typescript@6.0.3)(zod@4.4.3):
-    dependencies:
-      '@adraffy/ens-normalize': 1.11.1
-      '@noble/ciphers': 1.3.0
-      '@noble/curves': 1.9.1
-      '@noble/hashes': 1.8.0
-      '@scure/bip32': 1.7.0
-      '@scure/bip39': 1.6.0
-      abitype: 1.2.3(typescript@6.0.3)(zod@4.4.3)
-      eventemitter3: 5.0.1
-    optionalDependencies:
-      typescript: 6.0.3
-    transitivePeerDependencies:
-      - zod
-
-  streamx@2.28.0:
-    dependencies:
-      events-universal: 1.0.1
-      fast-fifo: 1.3.2
-      text-decoder: 1.2.7
-    transitivePeerDependencies:
-      - bare-abort-controller
-      - react-native-b4a
-
-  tar-stream@3.2.0:
-    dependencies:
-      b4a: 1.8.1
-      bare-fs: 4.7.4
-      fast-fifo: 1.3.2
-      streamx: 2.28.0
-    transitivePeerDependencies:
-      - bare-abort-controller
-      - bare-buffer
-      - react-native-b4a
-
-  teex@1.0.1:
-    dependencies:
-      streamx: 2.28.0
-    transitivePeerDependencies:
-      - bare-abort-controller
-      - react-native-b4a
-
-  text-decoder@1.2.7:
-    dependencies:
-      b4a: 1.8.1
-    transitivePeerDependencies:
-      - react-native-b4a
-
-  tsx@4.22.4:
-    dependencies:
-      esbuild: 0.28.0
-    optionalDependencies:
-      fsevents: 2.3.3
-
-  typescript@6.0.3: {}
-
-  undici-types@7.24.6: {}
-
-  viem@2.52.0(typescript@6.0.3)(zod@4.4.3):
-    dependencies:
-      '@noble/curves': 1.9.1
-      '@noble/hashes': 1.8.0
-      '@scure/bip32': 1.7.0
-      '@scure/bip39': 1.6.0
-      abitype: 1.2.3(typescript@6.0.3)(zod@4.4.3)
-      isows: 1.0.7(ws@8.20.1)
-      ox: 0.14.27(typescript@6.0.3)(zod@4.4.3)
-      ws: 8.20.1
-    optionalDependencies:
-      typescript: 6.0.3
-    transitivePeerDependencies:
-      - bufferutil
-      - utf-8-validate
-      - zod
-
-  ws@8.20.1: {}
-
-  yaml@2.9.0: {}
-
-  zod@4.4.3: {}
diff --git a/bench/scripts/verify-packed-consumer.mjs b/bench/scripts/verify-packed-consumer.mjs
index 209261f2..02463758 100644
--- a/bench/scripts/verify-packed-consumer.mjs
+++ b/bench/scripts/verify-packed-consumer.mjs
@@ -1,5 +1,5 @@
 import { execFile } from 'node:child_process'
-import { mkdtemp, mkdir, readFile, rm, writeFile } from 'node:fs/promises'
+import { mkdtemp, mkdir, readFile, readdir, rm, writeFile } from 'node:fs/promises'
 import { tmpdir } from 'node:os'
 import path from 'node:path'
 import { fileURLToPath } from 'node:url'
@@ -7,10 +7,14 @@ import { promisify } from 'node:util'
 
 const execFileAsync = promisify(execFile)
 const benchDir = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..')
+const repoRoot = path.resolve(benchDir, '..')
 const scratch = await mkdtemp(path.join(tmpdir(), 'agent-bench-consumer-'))
-const runtimePackage = process.env.AGENT_RUNTIME_PACKAGE
-  ? path.resolve(process.env.AGENT_RUNTIME_PACKAGE)
-  : undefined
+const args = new Set(process.argv.slice(2))
+const useLocalRuntime = args.delete('--local-runtime')
+if (args.size > 0) throw new Error(`unknown arguments: ${[...args].join(', ')}`)
+
+const TYPESCRIPT_5 = '5.9.3'
+const TYPESCRIPT_6 = '6.0.3'
 
 async function run(command, args, cwd, env = process.env) {
   try {
@@ -29,15 +33,38 @@ async function run(command, args, cwd, env = process.env) {
   }
 }
 
+async function resolveRuntimePackage(packDir) {
+  if (process.env.AGENT_RUNTIME_PACKAGE) {
+    return path.resolve(process.env.AGENT_RUNTIME_PACKAGE)
+  }
+  if (!useLocalRuntime) return undefined
+  const manifest = JSON.parse(await readFile(path.join(repoRoot, 'package.json'), 'utf8'))
+  if (manifest.name !== '@tangle-network/agent-runtime') {
+    throw new Error('--local-runtime requires an agent-runtime source workspace')
+  }
+  await run('pnpm', ['pack', '--pack-destination', packDir], repoRoot)
+  const tarballs = (await readdir(packDir)).filter((name) => name.endsWith('.tgz'))
+  if (tarballs.length !== 1) {
+    throw new Error(`expected one packed agent-runtime tarball, found ${tarballs.length}`)
+  }
+  return path.join(packDir, tarballs[0])
+}
+
 try {
   const packDir = path.join(scratch, 'pack')
+  const runtimePackDir = path.join(scratch, 'runtime-pack')
   const consumerDir = path.join(scratch, 'consumer')
   await mkdir(packDir)
+  await mkdir(runtimePackDir)
   await mkdir(consumerDir)
 
-  const packed = await run('npm', ['pack', '--json', '--pack-destination', packDir], benchDir)
-  const [{ filename }] = JSON.parse(packed.stdout)
-  const tarball = path.join(packDir, filename)
+  await run('pnpm', ['pack', '--pack-destination', packDir], benchDir)
+  const packedFiles = (await readdir(packDir)).filter((name) => name.endsWith('.tgz'))
+  if (packedFiles.length !== 1) {
+    throw new Error(`expected one packed agent-bench tarball, found ${packedFiles.length}`)
+  }
+  const tarball = path.join(packDir, packedFiles[0])
+  const runtimePackage = await resolveRuntimePackage(runtimePackDir)
   const manifest = JSON.parse(await readFile(path.join(benchDir, 'package.json'), 'utf8'))
   const devDependencies = manifest.devDependencies
   if (
@@ -73,7 +100,7 @@ try {
         },
         devDependencies: {
           '@types/node': devDependencies['@types/node'],
-          typescript: devDependencies.typescript,
+          typescript: TYPESCRIPT_5,
           tsx: devDependencies.tsx,
         },
       },
@@ -128,6 +155,28 @@ for name in sorted(expected):
     consumerDir,
   )
   await run('npm', ['exec', '--', 'tsc', '-p', 'tsconfig.json'], consumerDir)
+  const typescript5 = await run('npm', ['exec', '--', 'tsc', '--version'], consumerDir)
+  if (typescript5.stdout.trim() !== `Version ${TYPESCRIPT_5}`) {
+    throw new Error(`expected TypeScript ${TYPESCRIPT_5}, received ${typescript5.stdout.trim()}`)
+  }
+  await run(
+    'npm',
+    [
+      'install',
+      '--ignore-scripts',
+      '--no-audit',
+      '--no-fund',
+      '--no-save',
+      '--package-lock=false',
+      `typescript@${TYPESCRIPT_6}`,
+    ],
+    consumerDir,
+  )
+  await run('npm', ['exec', '--', 'tsc', '-p', 'tsconfig.json'], consumerDir)
+  const typescript6 = await run('npm', ['exec', '--', 'tsc', '--version'], consumerDir)
+  if (typescript6.stdout.trim() !== `Version ${TYPESCRIPT_6}`) {
+    throw new Error(`expected TypeScript ${TYPESCRIPT_6}, received ${typescript6.stdout.trim()}`)
+  }
   await run('npm', ['exec', '--', 'tsx', 'index.ts'], consumerDir)
   const installedPackage = path.join(consumerDir, 'node_modules', '@tangle-network', 'agent-bench')
   const prepared = await run(
@@ -168,7 +217,7 @@ for name in sorted(expected):
     })
   }
   console.log(
-    `packed consumer verified: ${manifest.name}@${manifest.version} with @tangle-network/agent-runtime@${runtimeManifest.version}; prepared ${prepareProof.executionPlanDigest}`,
+    `packed consumer verified: ${manifest.name}@${manifest.version} with @tangle-network/agent-runtime@${runtimeManifest.version}, TypeScript ${TYPESCRIPT_5} and ${TYPESCRIPT_6}; prepared ${prepareProof.executionPlanDigest}`,
   )
 } finally {
   await rm(scratch, { recursive: true, force: true })
diff --git a/bench/scripts/verify-pier-agent.mts b/bench/scripts/verify-pier-agent.mts
index 4d34acc4..f6a61a1d 100644
--- a/bench/scripts/verify-pier-agent.mts
+++ b/bench/scripts/verify-pier-agent.mts
@@ -31,6 +31,7 @@ import {
   type AgentCandidateOutputArtifactPort,
   type AgentCandidateTaskExecution,
   type ResolvedAgentCandidateContainer,
+  sealAgentCandidateBundle,
   verifyAgentCandidateBundle,
 } from '@tangle-network/agent-runtime'
 
@@ -113,13 +114,11 @@ function workspaceSnapshot(
     })
     .sort((left, right) => left.path.localeCompare(right.path))
   const material = {
-    schemaVersion: 1 as const,
     kind: 'agent-candidate-workspace-manifest' as const,
     files,
-  }
+  } satisfies AgentCandidateWorkspaceSnapshotEvidence['material']
   const manifest = canonicalBytes(material)
   return {
-    schemaVersion: 1,
     kind: 'agent-candidate-workspace-snapshot',
     digest: sha256(manifest),
     material,
@@ -329,7 +328,6 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
   const taskWorkspace = workspaceSnapshot(taskRoot, ['src/status.txt'])
   const candidateWorkspace = workspaceSnapshot(candidateRoot, ['runner.py'])
   const bundleWithoutDigest = {
-    schemaVersion: 1 as const,
     kind: 'agent-candidate-bundle' as const,
     digestAlgorithm: 'rfc8785-sha256' as const,
     profile: {
@@ -388,10 +386,7 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
       },
     },
   }
-  const bundle = {
-    ...bundleWithoutDigest,
-    digest: sha256(canonicalBytes(bundleWithoutDigest)),
-  } as AgentCandidateBundle
+  const bundle: AgentCandidateBundle = sealAgentCandidateBundle(bundleWithoutDigest)
   const container: ResolvedAgentCandidateContainer = {
     source: 'evaluator-task-container',
     image: fixtureImage,
diff --git a/package.json b/package.json
index a76a246b..fe05c16b 100644
--- a/package.json
+++ b/package.json
@@ -100,13 +100,15 @@
   "scripts": {
     "build": "tsup",
     "dev": "tsup --watch",
-    "prepare": "(git rev-parse --git-dir > /dev/null 2>&1 && git config core.hooksPath .githooks) || true; tsup",
+    "prepublishOnly": "pnpm run build",
     "test": "vitest run",
     "test:watch": "vitest",
     "lint": "biome check src tests examples",
     "lint:fix": "biome check --write src tests examples",
     "typecheck": "tsc --noEmit && pnpm run typecheck:examples",
     "typecheck:examples": "tsc --noEmit -p tsconfig.examples.json",
+    "verify:bench": "pnpm --filter @tangle-network/agent-bench run typecheck:public && pnpm --filter @tangle-network/agent-bench test && pnpm --filter @tangle-network/agent-bench run verify:package:local-runtime",
+    "verify:bench:published": "pnpm --filter @tangle-network/agent-bench run typecheck:public && pnpm --filter @tangle-network/agent-bench test && pnpm --filter @tangle-network/agent-bench run verify:package",
     "verify:package": "node scripts/verify-package-exports.mjs",
     "verify:primeintellect": "pnpm build && node scripts/verify-primeintellect-v1.mjs",
     "verify:primeintellect:live": "pnpm build && node scripts/verify-primeintellect-live.mjs",
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index eeca138d..c960dccb 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -58,6 +58,31 @@ importers:
         specifier: ^3.0.0
         version: 3.2.4(@types/node@25.9.3)(tsx@4.22.4)(yaml@2.9.0)
 
+  bench:
+    dependencies:
+      '@tangle-network/agent-eval':
+        specifier: 0.121.0
+        version: 0.121.0(typescript@5.9.3)
+      '@tangle-network/agent-interface':
+        specifier: 0.28.0
+        version: 0.28.0
+      '@tangle-network/agent-runtime':
+        specifier: workspace:*
+        version: link:..
+      '@tangle-network/sandbox':
+        specifier: ^0.10.5
+        version: 0.10.5(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))
+    devDependencies:
+      '@types/node':
+        specifier: ^25.9.3
+        version: 25.9.3
+      tsx:
+        specifier: ^4.22.4
+        version: 4.22.4
+      typescript:
+        specifier: ^5.7.0
+        version: 5.9.3
+
 packages:
 
   '@adraffy/ens-normalize@1.11.1':
diff --git a/pnpm-workspace.yaml b/pnpm-workspace.yaml
new file mode 100644
index 00000000..49250f39
--- /dev/null
+++ b/pnpm-workspace.yaml
@@ -0,0 +1,2 @@
+packages:
+  - bench
diff --git a/scripts/verify-package-exports.mjs b/scripts/verify-package-exports.mjs
index aa93f78b..37083527 100644
--- a/scripts/verify-package-exports.mjs
+++ b/scripts/verify-package-exports.mjs
@@ -1,5 +1,5 @@
 import { spawnSync } from 'node:child_process'
-import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from 'node:fs'
+import { mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from 'node:fs'
 import { tmpdir } from 'node:os'
 import { dirname, join, resolve } from 'node:path'
 import { fileURLToPath } from 'node:url'
@@ -72,22 +72,15 @@ try {
     '@tangle-network',
     'agent-knowledge',
   )
-  if (!existsSync(join(knowledgePackageDir, 'package.json'))) {
-    throw new Error('packed consumer requires an installed @tangle-network/agent-knowledge package')
-  }
-  const knowledgePackDir = join(tempRoot, 'knowledge')
-  mkdirSync(knowledgePackDir, { recursive: true })
-  run('pnpm', ['pack', '--pack-destination', knowledgePackDir], knowledgePackageDir)
-  const knowledgeTarballs = run(
-    'find',
-    [knowledgePackDir, '-maxdepth', '1', '-name', '*.tgz', '-print'],
-    repoRoot,
+  const knowledgePackageJson = JSON.parse(
+    readFileSync(join(knowledgePackageDir, 'package.json'), 'utf8'),
   )
-    .trim()
-    .split('\n')
-    .filter(Boolean)
-  if (knowledgeTarballs.length !== 1) {
-    throw new Error(`expected exactly one packed knowledge tarball, found ${knowledgeTarballs.length}`)
+  if (
+    knowledgePackageJson.name !== '@tangle-network/agent-knowledge' ||
+    typeof knowledgePackageJson.version !== 'string' ||
+    knowledgePackageJson.version.length === 0
+  ) {
+    throw new Error('packed consumer requires an installed @tangle-network/agent-knowledge release')
   }
   const peerPackages = [
     '@tangle-network/agent-eval',
@@ -122,7 +115,7 @@ try {
   )
   writeFileSync(
     join(appDir, 'pnpm-workspace.yaml'),
-    `overrides:\n  '@tangle-network/agent-knowledge': file:${knowledgeTarballs[0]}\n`,
+    `overrides:\n  '@tangle-network/agent-knowledge': ${knowledgePackageJson.version}\n`,
   )
   writeFileSync(
     join(appDir, 'tsconfig.json'),
@@ -203,7 +196,7 @@ try {
       void outcome
     `,
   )
-  run('pnpm', ['install', '--ignore-scripts', '--config.auto-install-peers=false'], appDir)
+  run('pnpm', ['install', '--config.auto-install-peers=false'], appDir)
   run('pnpm', ['exec', 'tsc', '-p', 'tsconfig.json'], appDir)
 
   run(
@@ -344,6 +337,31 @@ try {
     ],
     appDir,
   )
+
+  const installedPackageDir = join(
+    appDir,
+    'node_modules',
+    '@tangle-network',
+    'agent-runtime',
+  )
+  const repackDir = join(tempRoot, 'repack')
+  mkdirSync(repackDir, { recursive: true })
+  run(
+    'npm',
+    ['pack', '--ignore-scripts=false', '--pack-destination', repackDir],
+    installedPackageDir,
+  )
+  const repackedTarballs = run(
+    'find',
+    [repackDir, '-maxdepth', '1', '-name', '*.tgz', '-print'],
+    repoRoot,
+  )
+    .trim()
+    .split('\n')
+    .filter(Boolean)
+  if (repackedTarballs.length !== 1) {
+    throw new Error(`expected exactly one repacked runtime tarball, found ${repackedTarballs.length}`)
+  }
 } finally {
   rmSync(tempRoot, { recursive: true, force: true })
 }

From f4aa97955ff58152baedbff92b9ce4280de204a7 Mon Sep 17 00:00:00 2001
From: Drew Stone <drewstone329@gmail.com>
Date: Wed, 15 Jul 2026 13:30:59 -0600
Subject: [PATCH 6/8] feat(intelligence): bind product metadata to measured
 comparisons

---
 src/intelligence/improvement-cycle.ts |  2 ++
 tests/improvement-cycle.test.ts       | 12 ++++++++++++
 2 files changed, 14 insertions(+)

diff --git a/src/intelligence/improvement-cycle.ts b/src/intelligence/improvement-cycle.ts
index f4817767..320ec909 100644
--- a/src/intelligence/improvement-cycle.ts
+++ b/src/intelligence/improvement-cycle.ts
@@ -121,6 +121,7 @@ export interface CreateAgentImprovementMeasuredComparisonOptions<
   measuredSurface: ImproveSurface
   baselineProfile: AgentProfile
   candidateBundle: AgentCandidateBundle
+  metadata?: AgentImprovementMeasuredComparison['metadata']
 }
 
 export interface ReviewAgentImprovementInput {
@@ -779,6 +780,7 @@ export function createAgentImprovementMeasuredComparison<TScenario extends Scena
       durationMs: result.durationMs,
       totalCostUsd: result.totalCostUsd,
     },
+    ...(options.metadata ? { metadata: options.metadata } : {}),
   })
   return agentImprovementMeasuredComparisonSchema.parse(comparison)
 }
diff --git a/tests/improvement-cycle.test.ts b/tests/improvement-cycle.test.ts
index c680ed81..dcced276 100644
--- a/tests/improvement-cycle.test.ts
+++ b/tests/improvement-cycle.test.ts
@@ -418,6 +418,18 @@ describe('agent improvement lifecycle', () => {
       measuredSurface: 'prompt',
       baselineProfile: profile,
       candidateBundle: proposed.proposal.candidateBundle,
+      metadata: {
+        executionTask: {
+          itemId: 'scenario-1',
+          instructionDigest: `sha256:${'c'.repeat(64)}`,
+        },
+      },
+    })
+    expect(multiJudgeComparison.metadata).toEqual({
+      executionTask: {
+        itemId: 'scenario-1',
+        instructionDigest: `sha256:${'c'.repeat(64)}`,
+      },
     })
     expect(multiJudgeComparison.objectives).toEqual(
       expect.arrayContaining([

From ef4c0b5e7c245f1d82e9fb5a4e6a7249f6a5f014 Mon Sep 17 00:00:00 2001
From: Drew Stone <drewstone329@gmail.com>
Date: Wed, 15 Jul 2026 13:48:14 -0600
Subject: [PATCH 7/8] fix(intelligence): stabilize measured comparisons

---
 src/intelligence/improvement-cycle.ts |  7 +--
 tests/improvement-cycle.test.ts       | 70 +++++++++++++++++++++++++++
 vitest.config.ts                      | 15 ++++--
 3 files changed, 84 insertions(+), 8 deletions(-)

diff --git a/src/intelligence/improvement-cycle.ts b/src/intelligence/improvement-cycle.ts
index 320ec909..7dc4374f 100644
--- a/src/intelligence/improvement-cycle.ts
+++ b/src/intelligence/improvement-cycle.ts
@@ -970,15 +970,16 @@ function measuredObjective(
   })
   const baselineMean = measuredMean(baseline)
   const candidateMean = measuredMean(candidate)
+  const delta = interval.mean
   const estimate = {
     availability: 'measured',
     baseline: baselineMean,
     candidate: candidateMean,
-    delta: candidateMean - baselineMean,
+    delta,
     confidenceInterval: {
       level: interval.confidence,
-      lower: interval.low,
-      upper: interval.high,
+      lower: Math.min(interval.low, delta),
+      upper: Math.max(interval.high, delta),
       method: 'paired-bootstrap',
       statistic: 'mean',
       resamples: interval.resamples,
diff --git a/tests/improvement-cycle.test.ts b/tests/improvement-cycle.test.ts
index dcced276..dc00e79e 100644
--- a/tests/improvement-cycle.test.ts
+++ b/tests/improvement-cycle.test.ts
@@ -448,6 +448,76 @@ describe('agent improvement lifecycle', () => {
       ]),
     )
 
+    const constantScoreResult = {
+      ...proposed.improvement.raw,
+      baseline: {
+        ...proposed.improvement.raw.baseline,
+        compositeMean: 0.5,
+        perScenario: Object.fromEntries(scenarios.map(({ id }) => [id, 0.5])),
+      },
+      winner: {
+        ...proposed.improvement.raw.winner,
+        compositeMean: 0.8,
+        perScenario: Object.fromEntries(scenarios.map(({ id }) => [id, 0.8])),
+      },
+      lift: 0.8 - 0.5,
+      provenance: {
+        ...proposed.improvement.raw.provenance,
+        baselineHoldoutComposite: 0.5,
+        winnerHoldoutComposite: 0.8,
+        heldOutLift: 0.8 - 0.5,
+      },
+      raw: {
+        ...proposed.improvement.raw.raw,
+        baselineOnHoldout: {
+          ...proposed.improvement.raw.raw.baselineOnHoldout,
+          cells: proposed.improvement.raw.raw.baselineOnHoldout.cells.map((cell) => ({
+            ...cell,
+            judgeScores: Object.fromEntries(
+              Object.entries(cell.judgeScores).map(([name, score]) => [
+                name,
+                {
+                  ...score,
+                  composite: 0.5,
+                  dimensions: Object.fromEntries(
+                    Object.keys(score.dimensions).map((dimension) => [dimension, 0.5]),
+                  ),
+                },
+              ]),
+            ),
+          })),
+        },
+        winnerOnHoldout: {
+          ...proposed.improvement.raw.raw.winnerOnHoldout,
+          cells: proposed.improvement.raw.raw.winnerOnHoldout.cells.map((cell) => ({
+            ...cell,
+            judgeScores: Object.fromEntries(
+              Object.entries(cell.judgeScores).map(([name, score]) => [
+                name,
+                {
+                  ...score,
+                  composite: 0.8,
+                  dimensions: Object.fromEntries(
+                    Object.keys(score.dimensions).map((dimension) => [dimension, 0.8]),
+                  ),
+                },
+              ]),
+            ),
+          })),
+        },
+      },
+    }
+    const constantScoreComparison = createAgentImprovementMeasuredComparison({
+      result: constantScoreResult,
+      measuredSurface: 'prompt',
+      baselineProfile: profile,
+      candidateBundle: proposed.proposal.candidateBundle,
+    })
+    expect(constantScoreComparison.overall.confidenceInterval).toMatchObject({
+      lower: constantScoreComparison.overall.delta,
+      upper: constantScoreComparison.overall.delta,
+    })
+
     const compoundProfile = {
       ...proposed.improvement.profile,
       tools: { inspect_repository: true },
diff --git a/vitest.config.ts b/vitest.config.ts
index af6d0d43..f5f6454c 100644
--- a/vitest.config.ts
+++ b/vitest.config.ts
@@ -3,11 +3,16 @@ import { defineConfig } from 'vitest/config'
 
 export default defineConfig({
   resolve: {
-    // The package's own public subpath, resolvable from dynamically imported authored
-    // modules under test (the repo does not self-link in node_modules).
-    alias: {
-      '@tangle-network/agent-runtime/loops': resolve(__dirname, 'src/runtime/index.ts'),
-    },
+    alias: [
+      {
+        find: /^@tangle-network\/agent-runtime$/,
+        replacement: resolve(__dirname, 'src/index.ts'),
+      },
+      {
+        find: /^@tangle-network\/agent-runtime\/loops$/,
+        replacement: resolve(__dirname, 'src/runtime/index.ts'),
+      },
+    ],
   },
   test: {
     exclude: ['**/node_modules/**', 'dist/**', 'bench/**', '**/.claude/worktrees/**'],

From 0ca7af672a375ba2c2a655fce155d5f2e06fed93 Mon Sep 17 00:00:00 2001
From: Drew Stone <drewstone329@gmail.com>
Date: Fri, 17 Jul 2026 02:42:51 -0600
Subject: [PATCH 8/8] fix(candidate): bind exact experiments before activation

---
 bench/package.json                            |    6 +-
 bench/scripts/verify-pier-agent.mts           |   82 +-
 bench/src/pier-agent.test.mts                 |   25 +-
 bench/src/pier-agent.ts                       |    4 +-
 docs/api/candidate-execution.md               |   18 +-
 docs/api/index.md                             |  686 ++++-----
 docs/api/intelligence.md                      | 1128 +++++++++-----
 docs/api/primitive-catalog.md                 |   51 +-
 docs/canonical-api.md                         |    7 +-
 package.json                                  |   16 +-
 pnpm-lock.yaml                                |   96 +-
 scripts/verify-package-exports.mjs            |   62 +-
 src/candidate-execution/benchmark-grader.ts   |   51 +-
 src/candidate-execution/builder.ts            |   37 +-
 src/candidate-execution/claim-plan.ts         |   12 +-
 src/candidate-execution/execute.ts            |   63 +-
 .../executor-capture-evidence.ts              |    2 +-
 src/candidate-execution/finalize.ts           |    6 +
 src/candidate-execution/index.ts              |    2 +
 src/candidate-execution/outcome-evidence.ts   |   20 +-
 src/candidate-execution/prepare.ts            |  235 +--
 src/candidate-execution/prepared-state.ts     |   93 +-
 src/candidate-execution/profile.ts            |    7 +-
 src/candidate-execution/recover.ts            |   97 +-
 src/candidate-execution/types.ts              |   51 +-
 src/candidate-execution/workspace-task.ts     |   14 -
 src/intelligence/improvement-cycle.ts         | 1333 +++++++----------
 src/intelligence/index.ts                     |   41 +-
 .../sandbox-approved-candidate.ts             |  170 ++-
 src/intelligence/with-intelligence.test.ts    |   22 +-
 src/knowledge/improvement-job.ts              |   94 +-
 src/mcp/worktree-harness.ts                   |   13 +-
 tests/candidate-bundle-builder.test.ts        |   35 +-
 tests/candidate-execution-claim.test.ts       |   14 +-
 tests/candidate-execution-execute.test.ts     |   98 +-
 tests/candidate-execution-finalize.test.ts    |   11 +-
 ...ndidate-execution-outcome-evidence.test.ts |   54 +-
 tests/candidate-execution-prepare.test.ts     |  407 +----
 tests/candidate-execution-recover.test.ts     |   35 +-
 .../candidate-execution-task-outcome.test.ts  |    8 +-
 tests/helpers/candidate-execution-fixture.ts  |  248 ++-
 tests/helpers/candidate-experiment-fixture.ts |  199 +++
 tests/improvement-cycle.test.ts               | 1021 +++----------
 tests/knowledge-improvement-job.test.ts       |  190 +--
 tests/mcp/worktree-harness.test.ts            |   88 +-
 tests/sandbox-approved-candidate.test.ts      |  499 +++---
 46 files changed, 3641 insertions(+), 3810 deletions(-)
 delete mode 100644 src/candidate-execution/workspace-task.ts
 create mode 100644 tests/helpers/candidate-experiment-fixture.ts

diff --git a/bench/package.json b/bench/package.json
index 2f83e4fa..e3f72697 100644
--- a/bench/package.json
+++ b/bench/package.json
@@ -28,10 +28,10 @@
     "verify:pier": "tsx scripts/verify-pier-pair.mts"
   },
   "dependencies": {
-    "@tangle-network/agent-eval": "0.121.0",
-    "@tangle-network/agent-interface": "0.28.0",
+    "@tangle-network/agent-eval": "0.122.1",
+    "@tangle-network/agent-interface": "0.30.0",
     "@tangle-network/agent-runtime": "workspace:*",
-    "@tangle-network/sandbox": "^0.10.5"
+    "@tangle-network/sandbox": "^0.11.1"
   },
   "devDependencies": {
     "@types/node": "^25.9.3",
diff --git a/bench/scripts/verify-pier-agent.mts b/bench/scripts/verify-pier-agent.mts
index f6a61a1d..fff7ca27 100644
--- a/bench/scripts/verify-pier-agent.mts
+++ b/bench/scripts/verify-pier-agent.mts
@@ -16,6 +16,10 @@ import path from 'node:path'
 import { fileURLToPath } from 'node:url'
 
 import { canonicalJson, InMemoryTraceStore } from '@tangle-network/agent-eval'
+import {
+  sealCandidateBenchmarkSuite,
+  sealCandidateBenchmarkTask,
+} from '@tangle-network/agent-eval/contract'
 import type {
   AgentCandidateArtifactRef,
   AgentCandidateBundle,
@@ -84,6 +88,10 @@ function canonicalBytes(value: unknown): Buffer {
   return Buffer.from(canonicalJson(value), 'utf8')
 }
 
+function sealCanonical<T extends object>(material: T): T & { digest: Sha256Digest } {
+  return { ...material, digest: sha256(canonicalBytes(material)) }
+}
+
 function embedded(bytes: Uint8Array) {
   return {
     encoding: 'base64' as const,
@@ -371,20 +379,6 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
       },
     },
     memory: { mode: 'disabled' as const },
-    lineage: {
-      source: 'optimizer' as const,
-      parentDigests: [`sha256:${'a'.repeat(64)}` as Sha256Digest],
-      runIds: [`pier-no-model-proposer-${proofArm}`],
-      benchmark: {
-        name: 'pier-runtime-proof',
-        version: '1',
-        splitDigest: `sha256:${'b'.repeat(64)}` as Sha256Digest,
-      },
-      spend: {
-        proposal: { costUsd: 0, inputTokens: 0, outputTokens: 0, modelCalls: 0 },
-        evaluation: { costUsd: 0, inputTokens: 0, outputTokens: 0, modelCalls: 0 },
-      },
-    },
   }
   const bundle: AgentCandidateBundle = sealAgentCandidateBundle(bundleWithoutDigest)
   const container: ResolvedAgentCandidateContainer = {
@@ -402,12 +396,19 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
     artifact: graderArtifact,
   })
   const executionId = `pier-no-model-${proofArm}-${contextDigest}`
-  const task: AgentCandidateTaskExecution = {
-    executionId,
-    benchmark: 'pier-runtime-proof',
-    benchmarkVersion: '1',
-    taskId: `agent-bench/pier-candidate-no-model-${proofArm}`,
-    splitDigest: `sha256:${'b'.repeat(64)}`,
+  const benchmarkTask = sealCandidateBenchmarkTask({
+    kind: 'agent-candidate-benchmark-task',
+    digestAlgorithm: 'rfc8785-sha256',
+    benchmark: {
+      name: 'pier-runtime-proof',
+      version: '1',
+      splitDigest: `sha256:${'b'.repeat(64)}`,
+    },
+    scenario: {
+      id: `agent-bench/pier-candidate-no-model-${proofArm}`,
+      kind: 'coding',
+      scenarioDigest: sha256(Buffer.from(`pier-runtime-proof:${proofArm}`, 'utf8')),
+    },
     instruction,
     repository: {
       identity: 'github.com/tangle-network/agent-bench-pier-fixture',
@@ -416,15 +417,20 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
       baseTree: repositoryState.tree,
     },
     outcome: { kind: 'workspace' },
-    attempt: { number: 1, maxAttempts: 1, retryPolicy: 'none' },
-    model: { requested: modelRequest, reasoningEffort: 'xhigh' },
+    attempt: { maxAttempts: 1, retryPolicy: 'none' },
+    model: {
+      requested: modelRequest,
+      provider: 'openai',
+      model: 'gpt-5.4',
+      snapshot: 'gpt-5.4-no-model-proof',
+      reasoningEffort: 'xhigh',
+    },
     grader: {
       name: grader.name,
       version: grader.version,
+      format: 'tangle-grader',
       artifact: grader.artifact,
     },
-    executionRoots: { taskRoot: '/app', candidateRoot: '/opt/tangle-candidate' },
-    stagingRoots: { taskRoot, candidateRoot, profileRoot },
     workspace: taskWorkspace,
     evaluatorTaskContainer: container,
     limits: {
@@ -435,6 +441,30 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
       maxOutputTokens: 0,
       maxCostUsd: 0,
     },
+  })
+  const benchmark = sealCandidateBenchmarkSuite({
+    tasks: [benchmarkTask],
+    reps: 1,
+    seeds: [42],
+  })
+  const task: AgentCandidateTaskExecution = {
+    executionId,
+    runCell: sealCanonical({
+      kind: 'agent-candidate-run-cell' as const,
+      experimentDigest: sha256(Buffer.from('pier-runtime-proof-experiment', 'utf8')),
+      arm: 'candidate' as const,
+      bundleDigest: bundle.digest,
+      suiteDigest: benchmark.suite.digest,
+      taskDigest: benchmarkTask.digest,
+      taskIndex: 0,
+      repetition: 0,
+      seed: 42,
+      attempt: 1,
+    }),
+    benchmarkSuite: benchmark.suite,
+    task: benchmarkTask,
+    executionRoots: { taskRoot: '/app', candidateRoot: '/opt/tangle-candidate' },
+    stagingRoots: { taskRoot, candidateRoot, profileRoot },
   }
   const workspaces = createAgentCandidateWorkspacePort()
   const ports: AgentCandidateExecutionPorts = {
@@ -507,7 +537,7 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
         prepared: true,
         disposed: true,
         executionPlanDigest: prepared.executionPlan.value.digest,
-        graderDigest: task.grader.artifact.sha256,
+        graderDigest: benchmarkTask.grader.artifact.sha256,
       })}\n`,
     )
   } else {
@@ -596,7 +626,7 @@ ${proofArm === 'success' ? "(task / 'src/status.txt').write_text('ready\\nowner=
           const endedAt = Date.now()
           await traceStore.appendRun({
             runId: request.trace.runId,
-            scenarioId: task.taskId,
+            scenarioId: benchmarkTask.scenario.id,
             startedAt: endedAt - observedElapsedMs,
             endedAt,
             status: 'completed',
diff --git a/bench/src/pier-agent.test.mts b/bench/src/pier-agent.test.mts
index 1232bf89..f664030b 100644
--- a/bench/src/pier-agent.test.mts
+++ b/bench/src/pier-agent.test.mts
@@ -23,22 +23,22 @@ const digest = (bytes: Uint8Array): `sha256:${string}` =>
 function prepared(root: string): PreparedAgentCandidateExecution {
   const bundleDigest = `sha256:${'b'.repeat(64)}` as const
   const executionId = 'pier-test-execution'
+  const task = {
+    outcome: {
+      kind: 'workspace',
+      repository: {
+        identity: 'fixture/repository',
+        rootIdentity: 'fixture/repository',
+        baseCommit: '1'.repeat(40),
+        baseTree: '2'.repeat(40),
+      },
+    },
+  }
   const material = {
     schemaVersion: 1,
     kind: 'agent-candidate-execution-plan-material',
-    bundleDigest,
+    runCell: { bundleDigest },
     executionId,
-    task: {
-      outcome: {
-        kind: 'workspace',
-        repository: {
-          identity: 'fixture/repository',
-          rootIdentity: 'fixture/repository',
-          baseCommit: '1'.repeat(40),
-          baseTree: '2'.repeat(40),
-        },
-      },
-    },
     limits: {
       timeoutMs: 60_000,
       maxSteps: 8,
@@ -62,6 +62,7 @@ function prepared(root: string): PreparedAgentCandidateExecution {
   const profileBytes = Buffer.from('fixture profile\n')
   return {
     bundle: { digest: bundleDigest },
+    benchmark: { task },
     executionId,
     roots: {
       execution: { taskRoot: '/app' },
diff --git a/bench/src/pier-agent.ts b/bench/src/pier-agent.ts
index 2389a576..0d1c1495 100644
--- a/bench/src/pier-agent.ts
+++ b/bench/src/pier-agent.ts
@@ -431,7 +431,7 @@ export function protectedCaptureFromPierResult(
   const metadata = record(agentResult.metadata, 'Pier agent_result.metadata')
   const expected = {
     executionId: request.executionId,
-    bundleDigest: request.executionPlan.value.material.bundleDigest,
+    bundleDigest: request.executionPlan.value.material.runCell.bundleDigest,
     executionPlanDigest: request.executionPlan.value.digest,
     materializationReceiptDigest: request.materializationReceipt.digest,
   }
@@ -609,7 +609,7 @@ export async function executePreparedPierCandidate(
       }
       if (!result) return {}
       officialResults.set(request.executionId, result)
-      const task = options.prepared.executionPlan.value.material.task
+      const task = options.prepared.benchmark.task
       const outcome = task.outcome
       if (outcome.kind !== 'workspace') {
         throw new Error('Pier candidate execution requires a workspace task outcome')
diff --git a/docs/api/candidate-execution.md b/docs/api/candidate-execution.md
index 8210eba1..ae948607 100644
--- a/docs/api/candidate-execution.md
+++ b/docs/api/candidate-execution.md
@@ -152,6 +152,12 @@ Re-exports [InMemoryAgentCandidateExecutionClaimStore](index.md#inmemoryagentcan
 
 ***
 
+### AgentCandidatePreparationEvidence
+
+Re-exports [AgentCandidatePreparationEvidence](index.md#agentcandidatepreparationevidence)
+
+***
+
 ### FileAgentCandidateExecutionClaimStore
 
 Re-exports [FileAgentCandidateExecutionClaimStore](index.md#fileagentcandidateexecutionclaimstore)
@@ -308,12 +314,6 @@ Re-exports [AgentCandidateArtifactPort](index.md#agentcandidateartifactport)
 
 ***
 
-### AgentCandidateBenchmarkGraderIdentity
-
-Re-exports [AgentCandidateBenchmarkGraderIdentity](index.md#agentcandidatebenchmarkgraderidentity)
-
-***
-
 ### AgentCandidateBenchmarkGraderPort
 
 Re-exports [AgentCandidateBenchmarkGraderPort](index.md#agentcandidatebenchmarkgraderport)
@@ -350,6 +350,12 @@ Re-exports [AgentCandidateExecutorPort](index.md#agentcandidateexecutorport)
 
 ***
 
+### AgentCandidateExecutorProfileFile
+
+Re-exports [AgentCandidateExecutorProfileFile](index.md#agentcandidateexecutorprofilefile)
+
+***
+
 ### AgentCandidateExecutorRequest
 
 Re-exports [AgentCandidateExecutorRequest](index.md#agentcandidateexecutorrequest)
diff --git a/docs/api/index.md b/docs/api/index.md
index 28c48b22..ba86e36a 100644
--- a/docs/api/index.md
+++ b/docs/api/index.md
@@ -1305,19 +1305,37 @@ Defined in: [candidate-execution/builder.ts:60](https://github.com/tangle-networ
 
 Defined in: [candidate-execution/builder.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/builder.ts#L61)
 
+##### knowledge?
+
+> `optional` **knowledge?**: `AgentCandidateKnowledge`
+
+Defined in: [candidate-execution/builder.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/builder.ts#L62)
+
 ##### memory
 
 > **memory**: `AgentCandidateMemoryPolicy`
 
-Defined in: [candidate-execution/builder.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/builder.ts#L62)
+Defined in: [candidate-execution/builder.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/builder.ts#L63)
 
-##### lineage
+***
+
+### AgentCandidatePreparationEvidence
 
-> **lineage**: `Omit`\<`AgentCandidateLineage`, `"profileDiffIds"`\>
+Defined in: [candidate-execution/claim-file-formats.ts:11](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim-file-formats.ts#L11)
 
-Defined in: [candidate-execution/builder.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/builder.ts#L64)
+#### Properties
+
+##### executionPlan
 
-`profileDiffIds` is derived from `profile`; callers cannot contradict it.
+> `readonly` **executionPlan**: `AgentCandidateArtifactRef`
+
+Defined in: [candidate-execution/claim-file-formats.ts:12](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim-file-formats.ts#L12)
+
+##### materializationReceipt
+
+> `readonly` **materializationReceipt**: `AgentCandidateArtifactRef`
+
+Defined in: [candidate-execution/claim-file-formats.ts:13](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim-file-formats.ts#L13)
 
 ***
 
@@ -1449,7 +1467,7 @@ Defined in: [candidate-execution/claim.ts:57](https://github.com/tangle-network/
 
 ##### preparationEvidence
 
-> `readonly` **preparationEvidence**: `AgentCandidatePreparationEvidence`
+> `readonly` **preparationEvidence**: [`AgentCandidatePreparationEvidence`](#agentcandidatepreparationevidence)
 
 Defined in: [candidate-execution/claim.ts:59](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/claim.ts#L59)
 
@@ -1816,7 +1834,7 @@ Defined in: [candidate-execution/dispose.ts:11](https://github.com/tangle-networ
 
 ### ExecutePreparedAgentCandidateOptions
 
-Defined in: [candidate-execution/execute.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L63)
+Defined in: [candidate-execution/execute.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L66)
 
 #### Properties
 
@@ -1824,31 +1842,31 @@ Defined in: [candidate-execution/execute.ts:63](https://github.com/tangle-networ
 
 > **executor**: [`AgentCandidateExecutorPort`](#agentcandidateexecutorport)
 
-Defined in: [candidate-execution/execute.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L64)
+Defined in: [candidate-execution/execute.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L67)
 
 ##### grader
 
 > **grader**: [`AgentCandidateBenchmarkGraderPort`](#agentcandidatebenchmarkgraderport)
 
-Defined in: [candidate-execution/execute.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L65)
+Defined in: [candidate-execution/execute.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L68)
 
 ##### outputArtifacts
 
 > **outputArtifacts**: [`AgentCandidateOutputArtifactPort`](#agentcandidateoutputartifactport)
 
-Defined in: [candidate-execution/execute.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L66)
+Defined in: [candidate-execution/execute.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L69)
 
 ##### traceStore
 
 > **traceStore**: `TraceStore`
 
-Defined in: [candidate-execution/execute.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L67)
+Defined in: [candidate-execution/execute.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L70)
 
 ##### claimStore
 
 > **claimStore**: [`AgentCandidateExecutionClaimStore`](#agentcandidateexecutionclaimstore)
 
-Defined in: [candidate-execution/execute.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L69)
+Defined in: [candidate-execution/execute.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L72)
 
 Long-lived evaluator-owned store shared by every process that can run this benchmark.
 
@@ -1856,7 +1874,7 @@ Long-lived evaluator-owned store shared by every process that can run this bench
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/execute.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L71)
+Defined in: [candidate-execution/execute.ts:74](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L74)
 
 Maximum time to prove process death and revoke protected access after a run ends.
 
@@ -1864,7 +1882,7 @@ Maximum time to prove process death and revoke protected access after a run ends
 
 > `optional` **resultTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/execute.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L73)
+Defined in: [candidate-execution/execute.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L76)
 
 Maximum time for task verification, executable grading, and receipt construction.
 
@@ -1872,7 +1890,7 @@ Maximum time for task verification, executable grading, and receipt construction
 
 ### PrepareAgentCandidateExecutionOptions
 
-Defined in: [candidate-execution/prepare.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L81)
+Defined in: [candidate-execution/prepare.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L85)
 
 #### Properties
 
@@ -1880,13 +1898,13 @@ Defined in: [candidate-execution/prepare.ts:81](https://github.com/tangle-networ
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/prepare.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L82)
+Defined in: [candidate-execution/prepare.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L86)
 
 ##### resultTimeoutMs?
 
 > `optional` **resultTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/prepare.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L84)
+Defined in: [candidate-execution/prepare.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L88)
 
 Maximum time for task verification, executable grading, and receipt construction.
 
@@ -2077,7 +2095,7 @@ Exact environment names the activation endpoint must return, no more or fewer.
 
 ### RecoverExpiredAgentCandidateOptions
 
-Defined in: [candidate-execution/recover.ts:38](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L38)
+Defined in: [candidate-execution/recover.ts:35](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L35)
 
 #### Properties
 
@@ -2085,49 +2103,49 @@ Defined in: [candidate-execution/recover.ts:38](https://github.com/tangle-networ
 
 > **attempt**: [`AgentCandidateExecutionAttemptRef`](#agentcandidateexecutionattemptref)
 
-Defined in: [candidate-execution/recover.ts:39](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L39)
+Defined in: [candidate-execution/recover.ts:36](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L36)
 
 ##### claimStore
 
 > **claimStore**: [`AgentCandidateExecutionClaimStore`](#agentcandidateexecutionclaimstore)
 
-Defined in: [candidate-execution/recover.ts:40](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L40)
+Defined in: [candidate-execution/recover.ts:37](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L37)
 
 ##### executor
 
 > **executor**: [`AgentCandidateExecutorPort`](#agentcandidateexecutorport)
 
-Defined in: [candidate-execution/recover.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L41)
+Defined in: [candidate-execution/recover.ts:38](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L38)
 
 ##### traceStore
 
 > **traceStore**: `TraceStore`
 
-Defined in: [candidate-execution/recover.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L42)
+Defined in: [candidate-execution/recover.ts:39](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L39)
 
 ##### ports
 
 > **ports**: `Pick`\<[`AgentCandidateExecutionPorts`](#agentcandidateexecutionports), `"models"` \| `"memory"`\>
 
-Defined in: [candidate-execution/recover.ts:43](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L43)
+Defined in: [candidate-execution/recover.ts:40](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L40)
 
 ##### outputArtifacts
 
 > **outputArtifacts**: [`AgentCandidateOutputArtifactPort`](#agentcandidateoutputartifactport)
 
-Defined in: [candidate-execution/recover.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L44)
+Defined in: [candidate-execution/recover.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L41)
 
 ##### cleanupTimeoutMs?
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [candidate-execution/recover.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L45)
+Defined in: [candidate-execution/recover.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L42)
 
 ##### now?
 
 > `optional` **now?**: () => `number`
 
-Defined in: [candidate-execution/recover.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L47)
+Defined in: [candidate-execution/recover.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L44)
 
 Evaluator clock; must be the same clock used by the claim store.
 
@@ -2169,7 +2187,7 @@ Defined in: [candidate-execution/types.ts:42](https://github.com/tangle-network/
 
 ### AgentCandidateOutputArtifactPort
 
-Defined in: [candidate-execution/types.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L68)
+Defined in: [candidate-execution/types.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L69)
 
 Durable content-addressed evidence store controlled only by the evaluator.
 
@@ -2203,7 +2221,7 @@ Defined in: [candidate-execution/types.ts:42](https://github.com/tangle-network/
 
 > **put**(`input`): `Promise`\<`AgentCandidateArtifactRef`\>
 
-Defined in: [candidate-execution/types.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L70)
+Defined in: [candidate-execution/types.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L71)
 
 Must be idempotent for identical bytes and return only a durable S3/IPFS locator.
 
@@ -2237,7 +2255,7 @@ Abort must prevent durable publication when it happens before resolution.
 
 ### AgentCandidateRepositoryPort
 
-Defined in: [candidate-execution/types.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L80)
+Defined in: [candidate-execution/types.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L81)
 
 Resolves a declared GitHub repository to an already-present local Git object store.
 
@@ -2247,7 +2265,7 @@ Resolves a declared GitHub repository to an already-present local Git object sto
 
 > **resolve**(`repository`): `Promise`\<`string`\>
 
-Defined in: [candidate-execution/types.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L81)
+Defined in: [candidate-execution/types.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L82)
 
 ###### Parameters
 
@@ -2263,7 +2281,7 @@ Defined in: [candidate-execution/types.ts:81](https://github.com/tangle-network/
 
 ### AgentCandidateVerificationPorts
 
-Defined in: [candidate-execution/types.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L84)
+Defined in: [candidate-execution/types.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L85)
 
 #### Extended by
 
@@ -2275,19 +2293,19 @@ Defined in: [candidate-execution/types.ts:84](https://github.com/tangle-network/
 
 > **artifacts**: [`AgentCandidateArtifactPort`](#agentcandidateartifactport)
 
-Defined in: [candidate-execution/types.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L85)
+Defined in: [candidate-execution/types.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L86)
 
 ##### repositories
 
 > **repositories**: [`AgentCandidateRepositoryPort`](#agentcandidaterepositoryport)
 
-Defined in: [candidate-execution/types.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L86)
+Defined in: [candidate-execution/types.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L87)
 
 ***
 
 ### AgentCandidateWorkspacePort
 
-Defined in: [candidate-execution/types.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L96)
+Defined in: [candidate-execution/types.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L97)
 
 Materializes an already-verified workspace archive.
 
@@ -2301,7 +2319,7 @@ any archive encoding, or no-op when the exact workspace is already present.
 
 > **materialize**(`input`): `Promise`\<`void`\>
 
-Defined in: [candidate-execution/types.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L97)
+Defined in: [candidate-execution/types.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L98)
 
 ###### Parameters
 
@@ -2331,7 +2349,7 @@ Defined in: [candidate-execution/types.ts:97](https://github.com/tangle-network/
 
 ### ResolvedAgentCandidateContainer
 
-Defined in: [candidate-execution/types.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L105)
+Defined in: [candidate-execution/types.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L106)
 
 #### Properties
 
@@ -2339,37 +2357,37 @@ Defined in: [candidate-execution/types.ts:105](https://github.com/tangle-network
 
 > **source**: `"pinned-container"` \| `"evaluator-task-container"`
 
-Defined in: [candidate-execution/types.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L106)
+Defined in: [candidate-execution/types.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L107)
 
 ##### image
 
 > **image**: `string`
 
-Defined in: [candidate-execution/types.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L107)
+Defined in: [candidate-execution/types.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L108)
 
 ##### indexDigest
 
 > **indexDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L108)
+Defined in: [candidate-execution/types.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L109)
 
 ##### manifestDigest
 
 > **manifestDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L109)
+Defined in: [candidate-execution/types.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L110)
 
 ##### platform
 
 > **platform**: `AgentCandidateOciPlatform`
 
-Defined in: [candidate-execution/types.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L110)
+Defined in: [candidate-execution/types.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L111)
 
 ***
 
 ### AgentCandidateContainerPort
 
-Defined in: [candidate-execution/types.ts:113](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L113)
+Defined in: [candidate-execution/types.ts:114](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L114)
 
 #### Methods
 
@@ -2377,7 +2395,7 @@ Defined in: [candidate-execution/types.ts:113](https://github.com/tangle-network
 
 > **resolve**(`input`): `Promise`\<[`ResolvedAgentCandidateContainer`](#resolvedagentcandidatecontainer)\>
 
-Defined in: [candidate-execution/types.ts:114](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L114)
+Defined in: [candidate-execution/types.ts:115](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L115)
 
 ###### Parameters
 
@@ -2399,7 +2417,7 @@ Defined in: [candidate-execution/types.ts:114](https://github.com/tangle-network
 
 ### AgentCandidateModelPort
 
-Defined in: [candidate-execution/types.ts:120](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L120)
+Defined in: [candidate-execution/types.ts:121](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L121)
 
 #### Methods
 
@@ -2407,7 +2425,7 @@ Defined in: [candidate-execution/types.ts:120](https://github.com/tangle-network
 
 > **resolve**(`input`): `Promise`\<`AgentCandidateResolvedModel`\>
 
-Defined in: [candidate-execution/types.ts:121](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L121)
+Defined in: [candidate-execution/types.ts:122](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L122)
 
 ###### Parameters
 
@@ -2433,7 +2451,7 @@ Defined in: [candidate-execution/types.ts:121](https://github.com/tangle-network
 
 > **reserveGrant**(`input`): `Promise`\<[`AgentCandidateProtectedModelReservation`](#agentcandidateprotectedmodelreservation)\>
 
-Defined in: [candidate-execution/types.ts:131](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L131)
+Defined in: [candidate-execution/types.ts:132](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L132)
 
 Reserve a stable access identity without creating a live credential.
 The reservation is scoped to `preparationId` and must automatically expire
@@ -2479,7 +2497,7 @@ at `expiresAtMs`, even if this call returns ambiguously to the runtime.
 
 > **activateGrant**(`input`): `Promise`\<[`AgentCandidateProtectedModelActivation`](#agentcandidateprotectedmodelactivation)\>
 
-Defined in: [candidate-execution/types.ts:141](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L141)
+Defined in: [candidate-execution/types.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L142)
 
 Create the live scoped credential only after the execution attempt is durably claimed.
 
@@ -2515,7 +2533,7 @@ Create the live scoped credential only after the execution attempt is durably cl
 
 > **settleGrant**(`input`): `Promise`\<[`AgentCandidateProtectedModelSettlement`](#agentcandidateprotectedmodelsettlement)\>
 
-Defined in: [candidate-execution/types.ts:154](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L154)
+Defined in: [candidate-execution/types.ts:155](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L155)
 
 Atomically revoke the grant, drain in-flight calls, and return its immutable final ledger.
 This operation must be idempotent for the exact preparation and must also
@@ -2552,35 +2570,9 @@ different preparation, even when both reservations report the same digest.
 
 ***
 
-### AgentCandidateBenchmarkGraderIdentity
-
-Defined in: [candidate-execution/types.ts:169](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L169)
-
-#### Properties
-
-##### name
-
-> **name**: `string`
-
-Defined in: [candidate-execution/types.ts:170](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L170)
-
-##### version
-
-> **version**: `string`
-
-Defined in: [candidate-execution/types.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L171)
-
-##### artifact
-
-> **artifact**: `AgentCandidateArtifactRef`
-
-Defined in: [candidate-execution/types.ts:172](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L172)
-
-***
-
 ### AgentCandidateProtectedModelReservation
 
-Defined in: [candidate-execution/types.ts:175](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L175)
+Defined in: [candidate-execution/types.ts:170](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L170)
 
 #### Properties
 
@@ -2588,19 +2580,19 @@ Defined in: [candidate-execution/types.ts:175](https://github.com/tangle-network
 
 > **preparationId**: `string`
 
-Defined in: [candidate-execution/types.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L176)
+Defined in: [candidate-execution/types.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L171)
 
 ##### digest
 
 > **digest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:177](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L177)
+Defined in: [candidate-execution/types.ts:172](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L172)
 
 ##### expiresAtMs
 
 > **expiresAtMs**: `number`
 
-Defined in: [candidate-execution/types.ts:179](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L179)
+Defined in: [candidate-execution/types.ts:174](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L174)
 
 Evaluator service must expire and revoke this reservation at this epoch millisecond.
 
@@ -2608,7 +2600,7 @@ Evaluator service must expire and revoke this reservation at this epoch millisec
 
 > **enforcedLimits**: [`AgentCandidateModelLimits`](#agentcandidatemodellimits)
 
-Defined in: [candidate-execution/types.ts:181](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L181)
+Defined in: [candidate-execution/types.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L176)
 
 The gateway must stop calls before any one of these limits is exceeded.
 
@@ -2616,7 +2608,7 @@ The gateway must stop calls before any one of these limits is exceeded.
 
 > **network**: `AgentCandidateModelAccessNetwork`
 
-Defined in: [candidate-execution/types.ts:183](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L183)
+Defined in: [candidate-execution/types.ts:178](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L178)
 
 Exact public endpoint exception; every other candidate destination stays blocked.
 
@@ -2624,7 +2616,7 @@ Exact public endpoint exception; every other candidate destination stays blocked
 
 ### AgentCandidateProtectedModelActivation
 
-Defined in: [candidate-execution/types.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L186)
+Defined in: [candidate-execution/types.ts:181](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L181)
 
 #### Properties
 
@@ -2632,7 +2624,7 @@ Defined in: [candidate-execution/types.ts:186](https://github.com/tangle-network
 
 > **env**: `Readonly`\<`Record`\<`string`, `string`\>\>
 
-Defined in: [candidate-execution/types.ts:188](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L188)
+Defined in: [candidate-execution/types.ts:183](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L183)
 
 Injected only into the trusted executor after all pre-launch checks pass.
 
@@ -2640,7 +2632,7 @@ Injected only into the trusted executor after all pre-launch checks pass.
 
 ### AgentCandidateProtectedModelSettlement
 
-Defined in: [candidate-execution/types.ts:191](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L191)
+Defined in: [candidate-execution/types.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L186)
 
 #### Properties
 
@@ -2648,31 +2640,31 @@ Defined in: [candidate-execution/types.ts:191](https://github.com/tangle-network
 
 > **preparationId**: `string`
 
-Defined in: [candidate-execution/types.ts:192](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L192)
+Defined in: [candidate-execution/types.ts:187](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L187)
 
 ##### grantDigest
 
 > **grantDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:193](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L193)
+Defined in: [candidate-execution/types.ts:188](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L188)
 
 ##### closed
 
 > **closed**: `true`
 
-Defined in: [candidate-execution/types.ts:194](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L194)
+Defined in: [candidate-execution/types.ts:189](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L189)
 
 ##### calls
 
 > **calls**: readonly `AgentCandidateModelSettlementCall`[]
 
-Defined in: [candidate-execution/types.ts:195](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L195)
+Defined in: [candidate-execution/types.ts:190](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L190)
 
 ***
 
 ### AgentCandidateMemoryResetResult
 
-Defined in: [candidate-execution/types.ts:198](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L198)
+Defined in: [candidate-execution/types.ts:193](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L193)
 
 #### Properties
 
@@ -2680,43 +2672,43 @@ Defined in: [candidate-execution/types.ts:198](https://github.com/tangle-network
 
 > **preparationId**: `string`
 
-Defined in: [candidate-execution/types.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L199)
+Defined in: [candidate-execution/types.ts:194](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L194)
 
 ##### accessDigest
 
 > **accessDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:200](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L200)
+Defined in: [candidate-execution/types.ts:195](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L195)
 
 ##### expiresAtMs
 
 > **expiresAtMs**: `number`
 
-Defined in: [candidate-execution/types.ts:201](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L201)
+Defined in: [candidate-execution/types.ts:196](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L196)
 
 ##### evidence
 
 > **evidence**: `AgentCandidateCapturedArtifact`
 
-Defined in: [candidate-execution/types.ts:202](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L202)
+Defined in: [candidate-execution/types.ts:197](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L197)
 
 ##### emptyStateDigest
 
 > **emptyStateDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:203](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L203)
+Defined in: [candidate-execution/types.ts:198](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L198)
 
 ##### beforeState
 
 > **beforeState**: `AgentCandidateWorkspaceSnapshotEvidence`
 
-Defined in: [candidate-execution/types.ts:204](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L204)
+Defined in: [candidate-execution/types.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L199)
 
 ***
 
 ### AgentCandidateMemoryPort
 
-Defined in: [candidate-execution/types.ts:207](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L207)
+Defined in: [candidate-execution/types.ts:202](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L202)
 
 #### Methods
 
@@ -2724,7 +2716,7 @@ Defined in: [candidate-execution/types.ts:207](https://github.com/tangle-network
 
 > **reset**(`input`): `Promise`\<[`AgentCandidateMemoryResetResult`](#agentcandidatememoryresetresult)\>
 
-Defined in: [candidate-execution/types.ts:213](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L213)
+Defined in: [candidate-execution/types.ts:208](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L208)
 
 Reset and reserve exact task memory without returning live access.
 The service must scope the reservation to `preparationId`, automatically
@@ -2766,7 +2758,7 @@ revoke it at `expiresAtMs`, and never reuse it for another preparation.
 
 > **activate**(`input`): `Promise`\<\{ `env`: `Readonly`\<`Record`\<`string`, `string`\>\>; \}\>
 
-Defined in: [candidate-execution/types.ts:225](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L225)
+Defined in: [candidate-execution/types.ts:220](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L220)
 
 Create live scoped access only after the execution attempt is durably claimed.
 Activation must match the exact preparation/access pair and may not extend expiry.
@@ -2803,7 +2795,7 @@ Activation must match the exact preparation/access pair and may not extend expir
 
 > **close**(`input`): `Promise`\<\{ `closed`: `true`; \}\>
 
-Defined in: [candidate-execution/types.ts:237](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L237)
+Defined in: [candidate-execution/types.ts:232](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L232)
 
 Revoke evaluator-owned access after process death or a failed preparation.
 Must be idempotent and concurrency-safe for the exact preparation/access
@@ -2841,7 +2833,7 @@ pair and must never close a different preparation.
 
 ### AgentCandidateExecutionPorts
 
-Defined in: [candidate-execution/types.ts:246](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L246)
+Defined in: [candidate-execution/types.ts:241](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L241)
 
 #### Extends
 
@@ -2853,7 +2845,7 @@ Defined in: [candidate-execution/types.ts:246](https://github.com/tangle-network
 
 > **artifacts**: [`AgentCandidateArtifactPort`](#agentcandidateartifactport)
 
-Defined in: [candidate-execution/types.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L85)
+Defined in: [candidate-execution/types.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L86)
 
 ###### Inherited from
 
@@ -2863,7 +2855,7 @@ Defined in: [candidate-execution/types.ts:85](https://github.com/tangle-network/
 
 > **repositories**: [`AgentCandidateRepositoryPort`](#agentcandidaterepositoryport)
 
-Defined in: [candidate-execution/types.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L86)
+Defined in: [candidate-execution/types.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L87)
 
 ###### Inherited from
 
@@ -2873,33 +2865,33 @@ Defined in: [candidate-execution/types.ts:86](https://github.com/tangle-network/
 
 > **workspaces**: [`AgentCandidateWorkspacePort`](#agentcandidateworkspaceport)
 
-Defined in: [candidate-execution/types.ts:247](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L247)
+Defined in: [candidate-execution/types.ts:242](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L242)
 
 ##### containers
 
 > **containers**: [`AgentCandidateContainerPort`](#agentcandidatecontainerport)
 
-Defined in: [candidate-execution/types.ts:248](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L248)
+Defined in: [candidate-execution/types.ts:243](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L243)
 
 ##### models
 
 > **models**: [`AgentCandidateModelPort`](#agentcandidatemodelport)
 
-Defined in: [candidate-execution/types.ts:249](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L249)
+Defined in: [candidate-execution/types.ts:244](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L244)
 
 ##### memory
 
 > **memory**: [`AgentCandidateMemoryPort`](#agentcandidatememoryport)
 
-Defined in: [candidate-execution/types.ts:250](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L250)
+Defined in: [candidate-execution/types.ts:245](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L245)
 
 ***
 
 ### AgentCandidateTaskExecution
 
-Defined in: [candidate-execution/types.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L254)
+Defined in: [candidate-execution/types.ts:249](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L249)
 
-One signed benchmark task and the exact result shape its executor must capture.
+Runtime placement for one exact cell from a signed candidate experiment.
 
 #### Properties
 
@@ -2907,85 +2899,31 @@ One signed benchmark task and the exact result shape its executor must capture.
 
 > **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:255](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L255)
-
-##### benchmark
-
-> **benchmark**: `string`
-
-Defined in: [candidate-execution/types.ts:256](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L256)
-
-##### benchmarkVersion
-
-> **benchmarkVersion**: `string`
-
-Defined in: [candidate-execution/types.ts:257](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L257)
-
-##### taskId
-
-> **taskId**: `string`
-
-Defined in: [candidate-execution/types.ts:258](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L258)
-
-##### splitDigest
-
-> **splitDigest**: `` `sha256:${string}` ``
-
-Defined in: [candidate-execution/types.ts:259](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L259)
-
-##### instruction
-
-> **instruction**: `string`
-
-Defined in: [candidate-execution/types.ts:261](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L261)
-
-Exact agent-visible task instruction. The runtime rejects malformed Unicode.
-
-##### repository?
-
-> `optional` **repository?**: `AgentCandidateTaskRepository`
-
-Defined in: [candidate-execution/types.ts:263](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L263)
-
-Optional source identity, required when the expected outcome is a workspace.
-
-##### outcome
-
-> **outcome**: `AgentCandidateTaskOutcomeSpec`
-
-Defined in: [candidate-execution/types.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L264)
-
-##### attempt
-
-> **attempt**: `AgentCandidateAttemptPolicy`
-
-Defined in: [candidate-execution/types.ts:265](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L265)
+Defined in: [candidate-execution/types.ts:250](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L250)
 
-##### model
+##### runCell
 
-> **model**: `object`
+> **runCell**: `AgentCandidateRunCell`
 
-Defined in: [candidate-execution/types.ts:266](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L266)
+Defined in: [candidate-execution/types.ts:251](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L251)
 
-###### requested
+##### benchmarkSuite
 
-> **requested**: `string`
-
-###### reasoningEffort
+> **benchmarkSuite**: `AgentCandidateBenchmarkSuite`
 
-> **reasoningEffort**: `ReasoningEffort`
+Defined in: [candidate-execution/types.ts:252](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L252)
 
-##### grader
+##### task
 
-> **grader**: [`AgentCandidateBenchmarkGraderIdentity`](#agentcandidatebenchmarkgraderidentity)
+> **task**: `AgentCandidateBenchmarkTask`
 
-Defined in: [candidate-execution/types.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L270)
+Defined in: [candidate-execution/types.ts:253](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L253)
 
 ##### executionRoots
 
 > **executionRoots**: `object`
 
-Defined in: [candidate-execution/types.ts:272](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L272)
+Defined in: [candidate-execution/types.ts:255](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L255)
 
 Absolute paths inside the evaluator-owned execution environment.
 
@@ -3001,7 +2939,7 @@ Absolute paths inside the evaluator-owned execution environment.
 
 > **stagingRoots**: `object`
 
-Defined in: [candidate-execution/types.ts:277](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L277)
+Defined in: [candidate-execution/types.ts:260](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L260)
 
 Host-side staging roots. These are verified but never signed as container paths.
 
@@ -3017,29 +2955,11 @@ Host-side staging roots. These are verified but never signed as container paths.
 
 > **profileRoot**: `string`
 
-##### workspace
-
-> **workspace**: `AgentCandidateWorkspaceSnapshotEvidence`
-
-Defined in: [candidate-execution/types.ts:282](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L282)
-
-##### evaluatorTaskContainer?
-
-> `optional` **evaluatorTaskContainer?**: [`ResolvedAgentCandidateContainer`](#resolvedagentcandidatecontainer)
-
-Defined in: [candidate-execution/types.ts:283](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L283)
-
-##### limits
-
-> **limits**: `AgentCandidateExecutionLimits`
-
-Defined in: [candidate-execution/types.ts:284](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L284)
-
 ***
 
 ### VerifiedAgentCandidate
 
-Defined in: [candidate-execution/types.ts:287](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L287)
+Defined in: [candidate-execution/types.ts:267](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L267)
 
 #### Properties
 
@@ -3047,25 +2967,25 @@ Defined in: [candidate-execution/types.ts:287](https://github.com/tangle-network
 
 > `readonly` **bundle**: `AgentCandidateBundle`
 
-Defined in: [candidate-execution/types.ts:288](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L288)
+Defined in: [candidate-execution/types.ts:268](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L268)
 
 ##### materializedTree?
 
 > `readonly` `optional` **materializedTree?**: `string`
 
-Defined in: [candidate-execution/types.ts:289](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L289)
+Defined in: [candidate-execution/types.ts:269](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L269)
 
 ##### \[verifiedCandidateBrand\]
 
 > `readonly` **\[verifiedCandidateBrand\]**: `true`
 
-Defined in: [candidate-execution/types.ts:290](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L290)
+Defined in: [candidate-execution/types.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L270)
 
 ***
 
 ### CanonicalCandidateDocument
 
-Defined in: [candidate-execution/types.ts:293](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L293)
+Defined in: [candidate-execution/types.ts:273](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L273)
 
 #### Type Parameters
 
@@ -3079,13 +2999,13 @@ Defined in: [candidate-execution/types.ts:293](https://github.com/tangle-network
 
 > `readonly` **value**: `T`
 
-Defined in: [candidate-execution/types.ts:294](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L294)
+Defined in: [candidate-execution/types.ts:274](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L274)
 
 ##### bytes
 
 > `readonly` **bytes**: `Uint8Array`
 
-Defined in: [candidate-execution/types.ts:296](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L296)
+Defined in: [candidate-execution/types.ts:276](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L276)
 
 Canonical UTF-8 bytes of `value` with its top-level digest omitted.
 
@@ -3093,13 +3013,13 @@ Canonical UTF-8 bytes of `value` with its top-level digest omitted.
 
 > `readonly` **digest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:297](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L297)
+Defined in: [candidate-execution/types.ts:277](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L277)
 
 ***
 
 ### PreparedAgentCandidateLaunch
 
-Defined in: [candidate-execution/types.ts:300](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L300)
+Defined in: [candidate-execution/types.ts:280](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L280)
 
 #### Properties
 
@@ -3107,13 +3027,13 @@ Defined in: [candidate-execution/types.ts:300](https://github.com/tangle-network
 
 > **executable**: `string`
 
-Defined in: [candidate-execution/types.ts:301](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L301)
+Defined in: [candidate-execution/types.ts:281](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L281)
 
 ##### args
 
 > **args**: readonly `string`[]
 
-Defined in: [candidate-execution/types.ts:303](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L303)
+Defined in: [candidate-execution/types.ts:283](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L283)
 
 Complete fixed argv, including profile materializer flags but excluding task delivery.
 
@@ -3121,13 +3041,13 @@ Complete fixed argv, including profile materializer flags but excluding task del
 
 > **env**: `Readonly`\<`Record`\<`string`, `string`\>\>
 
-Defined in: [candidate-execution/types.ts:304](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L304)
+Defined in: [candidate-execution/types.ts:284](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L284)
 
 ##### flags
 
 > **flags**: readonly `string`[]
 
-Defined in: [candidate-execution/types.ts:306](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L306)
+Defined in: [candidate-execution/types.ts:286](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L286)
 
 Informational subset already present at the tail of `args`; executors must not append twice.
 
@@ -3135,13 +3055,13 @@ Informational subset already present at the tail of `args`; executors must not a
 
 > **cwd**: `string`
 
-Defined in: [candidate-execution/types.ts:307](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L307)
+Defined in: [candidate-execution/types.ts:287](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L287)
 
 ***
 
 ### PreparedAgentCandidateInstruction
 
-Defined in: [candidate-execution/types.ts:310](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L310)
+Defined in: [candidate-execution/types.ts:290](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L290)
 
 #### Properties
 
@@ -3149,19 +3069,19 @@ Defined in: [candidate-execution/types.ts:310](https://github.com/tangle-network
 
 > **bytes**: `Uint8Array`
 
-Defined in: [candidate-execution/types.ts:311](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L311)
+Defined in: [candidate-execution/types.ts:291](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L291)
 
 ##### delivery
 
 > **delivery**: `AgentCandidateInstructionDelivery`
 
-Defined in: [candidate-execution/types.ts:312](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L312)
+Defined in: [candidate-execution/types.ts:292](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L292)
 
 ***
 
 ### PreparedAgentCandidateKnowledge
 
-Defined in: [candidate-execution/types.ts:316](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L316)
+Defined in: [candidate-execution/types.ts:296](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L296)
 
 Exact file-backed knowledge admitted by the candidate bundle.
 
@@ -3171,31 +3091,31 @@ Exact file-backed knowledge admitted by the candidate bundle.
 
 > `readonly` **candidate**: `AgentCandidateKnowledgeRef`
 
-Defined in: [candidate-execution/types.ts:317](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L317)
+Defined in: [candidate-execution/types.ts:297](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L297)
 
 ##### snapshot
 
 > `readonly` **snapshot**: `AgentCandidateWorkspaceSnapshotEvidence`
 
-Defined in: [candidate-execution/types.ts:318](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L318)
+Defined in: [candidate-execution/types.ts:298](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L298)
 
 ##### files
 
 > `readonly` **files**: readonly [`AgentCandidateExecutorWorkspaceFile`](#agentcandidateexecutorworkspacefile)[]
 
-Defined in: [candidate-execution/types.ts:319](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L319)
+Defined in: [candidate-execution/types.ts:299](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L299)
 
 ##### retrievalConfig?
 
 > `readonly` `optional` **retrievalConfig?**: `Uint8Array`\<`ArrayBufferLike`\>
 
-Defined in: [candidate-execution/types.ts:320](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L320)
+Defined in: [candidate-execution/types.ts:300](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L300)
 
 ***
 
 ### PreparedAgentCandidateTrace
 
-Defined in: [candidate-execution/types.ts:323](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L323)
+Defined in: [candidate-execution/types.ts:303](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L303)
 
 #### Properties
 
@@ -3203,25 +3123,25 @@ Defined in: [candidate-execution/types.ts:323](https://github.com/tangle-network
 
 > **runId**: `string`
 
-Defined in: [candidate-execution/types.ts:324](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L324)
+Defined in: [candidate-execution/types.ts:304](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L304)
 
 ##### tags
 
 > **tags**: `Readonly`\<`Record`\<`string`, `string`\>\>
 
-Defined in: [candidate-execution/types.ts:325](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L325)
+Defined in: [candidate-execution/types.ts:305](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L305)
 
 ##### env
 
 > **env**: `Readonly`\<`Record`\<`string`, `string`\>\>
 
-Defined in: [candidate-execution/types.ts:326](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L326)
+Defined in: [candidate-execution/types.ts:306](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L306)
 
 ***
 
 ### PreparedAgentCandidateExecution
 
-Defined in: [candidate-execution/types.ts:329](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L329)
+Defined in: [candidate-execution/types.ts:309](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L309)
 
 #### Properties
 
@@ -3229,19 +3149,33 @@ Defined in: [candidate-execution/types.ts:329](https://github.com/tangle-network
 
 > `readonly` **bundle**: `AgentCandidateBundle`
 
-Defined in: [candidate-execution/types.ts:330](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L330)
+Defined in: [candidate-execution/types.ts:310](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L310)
+
+##### benchmark
+
+> `readonly` **benchmark**: `object`
+
+Defined in: [candidate-execution/types.ts:311](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L311)
+
+###### suite
+
+> `readonly` **suite**: `AgentCandidateBenchmarkSuite`
+
+###### task
+
+> `readonly` **task**: `AgentCandidateBenchmarkTask`
 
 ##### executionId
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:331](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L331)
+Defined in: [candidate-execution/types.ts:315](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L315)
 
 ##### roots
 
 > `readonly` **roots**: `object`
 
-Defined in: [candidate-execution/types.ts:332](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L332)
+Defined in: [candidate-execution/types.ts:316](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L316)
 
 ###### execution
 
@@ -3275,7 +3209,7 @@ Defined in: [candidate-execution/types.ts:332](https://github.com/tangle-network
 
 > `readonly` **profilePlan**: `object`
 
-Defined in: [candidate-execution/types.ts:343](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L343)
+Defined in: [candidate-execution/types.ts:327](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L327)
 
 ###### value
 
@@ -3293,13 +3227,13 @@ Defined in: [candidate-execution/types.ts:343](https://github.com/tangle-network
 
 > `readonly` **profileActivation**: `AgentCandidateProfileActivation`
 
-Defined in: [candidate-execution/types.ts:348](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L348)
+Defined in: [candidate-execution/types.ts:332](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L332)
 
 ##### executionPlan
 
 > `readonly` **executionPlan**: `object`
 
-Defined in: [candidate-execution/types.ts:349](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L349)
+Defined in: [candidate-execution/types.ts:333](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L333)
 
 ###### value
 
@@ -3313,55 +3247,55 @@ Defined in: [candidate-execution/types.ts:349](https://github.com/tangle-network
 
 > `readonly` **materializationReceipt**: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateMaterializationReceipt`\>
 
-Defined in: [candidate-execution/types.ts:353](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L353)
+Defined in: [candidate-execution/types.ts:337](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L337)
 
 ##### launch
 
 > `readonly` **launch**: [`PreparedAgentCandidateLaunch`](#preparedagentcandidatelaunch)
 
-Defined in: [candidate-execution/types.ts:354](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L354)
+Defined in: [candidate-execution/types.ts:338](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L338)
 
 ##### instruction
 
 > `readonly` **instruction**: [`PreparedAgentCandidateInstruction`](#preparedagentcandidateinstruction)
 
-Defined in: [candidate-execution/types.ts:355](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L355)
+Defined in: [candidate-execution/types.ts:339](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L339)
 
 ##### resolvedModel
 
 > `readonly` **resolvedModel**: `AgentCandidateResolvedModel`
 
-Defined in: [candidate-execution/types.ts:356](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L356)
+Defined in: [candidate-execution/types.ts:340](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L340)
 
 ##### knowledge?
 
 > `readonly` `optional` **knowledge?**: [`PreparedAgentCandidateKnowledge`](#preparedagentcandidateknowledge)
 
-Defined in: [candidate-execution/types.ts:357](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L357)
+Defined in: [candidate-execution/types.ts:341](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L341)
 
 ##### trace
 
 > `readonly` **trace**: [`PreparedAgentCandidateTrace`](#preparedagentcandidatetrace)
 
-Defined in: [candidate-execution/types.ts:358](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L358)
+Defined in: [candidate-execution/types.ts:342](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L342)
 
 ##### memory
 
 > `readonly` **memory**: `AgentCandidateEffectiveMemory`
 
-Defined in: [candidate-execution/types.ts:359](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L359)
+Defined in: [candidate-execution/types.ts:343](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L343)
 
 ##### \[preparedCandidateBrand\]
 
 > `readonly` **\[preparedCandidateBrand\]**: `true`
 
-Defined in: [candidate-execution/types.ts:360](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L360)
+Defined in: [candidate-execution/types.ts:344](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L344)
 
 ***
 
 ### AgentCandidateProtectedRunCapture
 
-Defined in: [candidate-execution/types.ts:363](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L363)
+Defined in: [candidate-execution/types.ts:349](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L349)
 
 #### Properties
 
@@ -3369,19 +3303,19 @@ Defined in: [candidate-execution/types.ts:363](https://github.com/tangle-network
 
 > **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:364](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L364)
+Defined in: [candidate-execution/types.ts:350](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L350)
 
 ##### termination
 
 > **termination**: `AgentCandidateTermination`
 
-Defined in: [candidate-execution/types.ts:365](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L365)
+Defined in: [candidate-execution/types.ts:351](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L351)
 
 ***
 
 ### AgentCandidateExecutorMemoryCapture
 
-Defined in: [candidate-execution/types.ts:388](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L388)
+Defined in: [candidate-execution/types.ts:374](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L374)
 
 Raw isolated-memory capture made only after access has been revoked.
 
@@ -3391,19 +3325,19 @@ Raw isolated-memory capture made only after access has been revoked.
 
 > `readonly` **afterState**: `AgentCandidateWorkspaceManifestMaterial`
 
-Defined in: [candidate-execution/types.ts:389](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L389)
+Defined in: [candidate-execution/types.ts:375](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L375)
 
 ##### archive
 
 > `readonly` **archive**: `Uint8Array`
 
-Defined in: [candidate-execution/types.ts:390](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L390)
+Defined in: [candidate-execution/types.ts:376](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L376)
 
 ***
 
 ### AgentCandidateExecutorFinalCapture
 
-Defined in: [candidate-execution/types.ts:394](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L394)
+Defined in: [candidate-execution/types.ts:380](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L380)
 
 Replayable evaluator result captured only after process death and trace drain.
 
@@ -3413,13 +3347,13 @@ Replayable evaluator result captured only after process death and trace drain.
 
 > `readonly` `optional` **taskOutcome?**: [`AgentCandidateExecutorTaskOutcomeCapture`](#agentcandidateexecutortaskoutcomecapture)
 
-Defined in: [candidate-execution/types.ts:395](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L395)
+Defined in: [candidate-execution/types.ts:381](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L381)
 
 ##### memoryAfter?
 
 > `readonly` `optional` **memoryAfter?**: [`AgentCandidateExecutorMemoryCapture`](#agentcandidateexecutormemorycapture)
 
-Defined in: [candidate-execution/types.ts:397](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L397)
+Defined in: [candidate-execution/types.ts:383](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L383)
 
 Required only when the prepared candidate uses isolated task memory.
 
@@ -3427,7 +3361,7 @@ Required only when the prepared candidate uses isolated task memory.
 
 > `readonly` `optional` **evidence?**: `Uint8Array`\<`ArrayBufferLike`\>
 
-Defined in: [candidate-execution/types.ts:399](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L399)
+Defined in: [candidate-execution/types.ts:385](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L385)
 
 Executor-native bytes preserved when a fresh worker cannot reconstruct a verified outcome.
 
@@ -3435,7 +3369,7 @@ Executor-native bytes preserved when a fresh worker cannot reconstruct a verifie
 
 ### AgentCandidateBenchmarkGraderPort
 
-Defined in: [candidate-execution/types.ts:436](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L436)
+Defined in: [candidate-execution/types.ts:422](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L422)
 
 Evaluator-owned executable grader, pinned by immutable implementation bytes.
 
@@ -3451,19 +3385,19 @@ copying an expected digest from ambient configuration.
 
 > `readonly` **name**: `string`
 
-Defined in: [candidate-execution/types.ts:437](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L437)
+Defined in: [candidate-execution/types.ts:423](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L423)
 
 ##### version
 
 > `readonly` **version**: `string`
 
-Defined in: [candidate-execution/types.ts:438](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L438)
+Defined in: [candidate-execution/types.ts:424](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L424)
 
 ##### artifact
 
 > `readonly` **artifact**: `AgentCandidateArtifactRef`
 
-Defined in: [candidate-execution/types.ts:439](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L439)
+Defined in: [candidate-execution/types.ts:425](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L425)
 
 #### Methods
 
@@ -3471,7 +3405,7 @@ Defined in: [candidate-execution/types.ts:439](https://github.com/tangle-network
 
 > **run**(`input`): `Promise`\<\{ `evaluation`: `BenchmarkEvaluation`; `evidence`: `Uint8Array`; `binding`: \{ `implementationDigest`: `` `sha256:${string}` ``; `taskOutcomeDigest`: `` `sha256:${string}` ``; `outputDigest`: `` `sha256:${string}` ``; \}; \}\>
 
-Defined in: [candidate-execution/types.ts:440](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L440)
+Defined in: [candidate-execution/types.ts:426](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L426)
 
 ###### Parameters
 
@@ -3517,7 +3451,7 @@ Frozen result deadline; runners must stop work and side effects when aborted.
 
 ### AgentCandidateExecutorRequest
 
-Defined in: [candidate-execution/types.ts:468](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L468)
+Defined in: [candidate-execution/types.ts:454](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L454)
 
 One detached request passed to the trusted environment-specific executor.
 
@@ -3527,13 +3461,27 @@ One detached request passed to the trusted environment-specific executor.
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:469](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L469)
+Defined in: [candidate-execution/types.ts:455](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L455)
+
+##### benchmark
+
+> `readonly` **benchmark**: `object`
+
+Defined in: [candidate-execution/types.ts:456](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L456)
+
+###### suite
+
+> `readonly` **suite**: `AgentCandidateBenchmarkSuite`
+
+###### task
+
+> `readonly` **task**: `AgentCandidateBenchmarkTask`
 
 ##### inputs
 
 > `readonly` **inputs**: `object`
 
-Defined in: [candidate-execution/types.ts:471](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L471)
+Defined in: [candidate-execution/types.ts:458](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L458)
 
 Immutable bytes from which the executor creates fresh isolated workspaces.
 
@@ -3551,13 +3499,13 @@ Immutable bytes from which the executor creates fresh isolated workspaces.
 
 ###### profile.files
 
-> `readonly` **files**: readonly `AgentCandidateExecutorProfileFile`[]
+> `readonly` **files**: readonly [`AgentCandidateExecutorProfileFile`](#agentcandidateexecutorprofilefile)[]
 
 ##### roots
 
 > `readonly` **roots**: `object`
 
-Defined in: [candidate-execution/types.ts:478](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L478)
+Defined in: [candidate-execution/types.ts:465](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L465)
 
 ###### taskRoot
 
@@ -3571,7 +3519,7 @@ Defined in: [candidate-execution/types.ts:478](https://github.com/tangle-network
 
 > `readonly` **profilePlan**: `object`
 
-Defined in: [candidate-execution/types.ts:479](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L479)
+Defined in: [candidate-execution/types.ts:466](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L466)
 
 ###### value
 
@@ -3589,13 +3537,13 @@ Defined in: [candidate-execution/types.ts:479](https://github.com/tangle-network
 
 > `readonly` **profileActivation**: `AgentCandidateProfileActivation`
 
-Defined in: [candidate-execution/types.ts:480](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L480)
+Defined in: [candidate-execution/types.ts:467](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L467)
 
 ##### executionPlan
 
 > `readonly` **executionPlan**: `object`
 
-Defined in: [candidate-execution/types.ts:481](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L481)
+Defined in: [candidate-execution/types.ts:468](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L468)
 
 ###### value
 
@@ -3609,31 +3557,31 @@ Defined in: [candidate-execution/types.ts:481](https://github.com/tangle-network
 
 > `readonly` **materializationReceipt**: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateMaterializationReceipt`\>
 
-Defined in: [candidate-execution/types.ts:482](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L482)
+Defined in: [candidate-execution/types.ts:469](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L469)
 
 ##### launch
 
 > `readonly` **launch**: [`PreparedAgentCandidateLaunch`](#preparedagentcandidatelaunch)
 
-Defined in: [candidate-execution/types.ts:483](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L483)
+Defined in: [candidate-execution/types.ts:470](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L470)
 
 ##### instruction
 
 > `readonly` **instruction**: [`PreparedAgentCandidateInstruction`](#preparedagentcandidateinstruction)
 
-Defined in: [candidate-execution/types.ts:484](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L484)
+Defined in: [candidate-execution/types.ts:471](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L471)
 
 ##### resolvedModel
 
 > `readonly` **resolvedModel**: `AgentCandidateResolvedModel`
 
-Defined in: [candidate-execution/types.ts:485](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L485)
+Defined in: [candidate-execution/types.ts:472](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L472)
 
 ##### hardLimits
 
 > `readonly` **hardLimits**: `Pick`\<`AgentCandidateExecutionLimits`, `"timeoutMs"`\>
 
-Defined in: [candidate-execution/types.ts:487](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L487)
+Defined in: [candidate-execution/types.ts:474](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L474)
 
 Mechanically enforced by the runtime plus executor process-death acknowledgement.
 
@@ -3641,7 +3589,7 @@ Mechanically enforced by the runtime plus executor process-death acknowledgement
 
 > `readonly` **observedLimits**: `Pick`\<`AgentCandidateExecutionLimits`, `"maxSteps"`\>
 
-Defined in: [candidate-execution/types.ts:489](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L489)
+Defined in: [candidate-execution/types.ts:476](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L476)
 
 Validity bound checked against protected traces; generic black-box executors cannot preempt it.
 
@@ -3649,25 +3597,25 @@ Validity bound checked against protected traces; generic black-box executors can
 
 > `readonly` `optional` **knowledge?**: [`PreparedAgentCandidateKnowledge`](#preparedagentcandidateknowledge)
 
-Defined in: [candidate-execution/types.ts:490](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L490)
+Defined in: [candidate-execution/types.ts:477](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L477)
 
 ##### trace
 
 > `readonly` **trace**: [`PreparedAgentCandidateTrace`](#preparedagentcandidatetrace)
 
-Defined in: [candidate-execution/types.ts:491](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L491)
+Defined in: [candidate-execution/types.ts:478](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L478)
 
 ##### memory
 
 > `readonly` **memory**: `AgentCandidateEffectiveMemory`
 
-Defined in: [candidate-execution/types.ts:492](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L492)
+Defined in: [candidate-execution/types.ts:479](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L479)
 
 ***
 
 ### AgentCandidateExecutorPort
 
-Defined in: [candidate-execution/types.ts:503](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L503)
+Defined in: [candidate-execution/types.ts:490](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L490)
 
 Executes one prepared request inside an evaluator-owned isolation boundary.
 
@@ -3682,7 +3630,7 @@ no candidate-authored usage or score fields.
 
 > **execute**(`request`, `context`): `Promise`\<[`AgentCandidateProtectedRunCapture`](#agentcandidateprotectedruncapture)\>
 
-Defined in: [candidate-execution/types.ts:504](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L504)
+Defined in: [candidate-execution/types.ts:491](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L491)
 
 ###### Parameters
 
@@ -3716,7 +3664,7 @@ Absolute epoch-millisecond deadline owned by the runtime.
 
 > **stop**(`request`, `context`): `Promise`\<\{ `stopped`: `true`; \}\>
 
-Defined in: [candidate-execution/types.ts:515](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L515)
+Defined in: [candidate-execution/types.ts:502](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L502)
 
 Kill the exact process/container and drain trace writes. Must be idempotent.
 
@@ -3756,7 +3704,7 @@ Absolute execution deadline; a later stop acknowledgement cannot produce success
 
 > **capture**(`request`, `context`): `Promise`\<[`AgentCandidateExecutorFinalCapture`](#agentcandidateexecutorfinalcapture)\>
 
-Defined in: [candidate-execution/types.ts:527](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L527)
+Defined in: [candidate-execution/types.ts:514](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L514)
 
 Capture immutable final evidence after stop. Must be replayable by a fresh worker.
 
@@ -3782,11 +3730,35 @@ Aborted at the frozen execution deadline or evaluator cleanup deadline.
 
 `Promise`\<[`AgentCandidateExecutorFinalCapture`](#agentcandidateexecutorfinalcapture)\>
 
+##### dispose()?
+
+> `optional` **dispose**(`request`, `context`): `Promise`\<\{ `disposed`: `true`; \}\>
+
+Defined in: [candidate-execution/types.ts:523](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L523)
+
+Remove evaluator-owned execution resources after final capture. Must be idempotent.
+
+###### Parameters
+
+###### request
+
+[`AgentCandidateExecutorStopRequest`](#agentcandidateexecutorstoprequest)
+
+###### context
+
+###### signal
+
+`AbortSignal`
+
+###### Returns
+
+`Promise`\<\{ `disposed`: `true`; \}\>
+
 ***
 
 ### AgentCandidateExecutorStopRequest
 
-Defined in: [candidate-execution/types.ts:538](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L538)
+Defined in: [candidate-execution/types.ts:532](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L532)
 
 Opaque process identity used for termination without re-exposing launch credentials.
 
@@ -3796,19 +3768,19 @@ Opaque process identity used for termination without re-exposing launch credenti
 
 > `readonly` **executionId**: `string`
 
-Defined in: [candidate-execution/types.ts:539](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L539)
+Defined in: [candidate-execution/types.ts:533](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L533)
 
 ##### executionPlanDigest
 
 > `readonly` **executionPlanDigest**: `` `sha256:${string}` ``
 
-Defined in: [candidate-execution/types.ts:540](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L540)
+Defined in: [candidate-execution/types.ts:534](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L534)
 
 ***
 
 ### AgentCandidateExecutorWorkspaceInput
 
-Defined in: [candidate-execution/types.ts:543](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L543)
+Defined in: [candidate-execution/types.ts:537](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L537)
 
 #### Properties
 
@@ -3816,19 +3788,19 @@ Defined in: [candidate-execution/types.ts:543](https://github.com/tangle-network
 
 > `readonly` **snapshot**: `AgentCandidateWorkspaceSnapshotEvidence`
 
-Defined in: [candidate-execution/types.ts:544](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L544)
+Defined in: [candidate-execution/types.ts:538](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L538)
 
 ##### files
 
 > `readonly` **files**: readonly [`AgentCandidateExecutorWorkspaceFile`](#agentcandidateexecutorworkspacefile)[]
 
-Defined in: [candidate-execution/types.ts:545](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L545)
+Defined in: [candidate-execution/types.ts:539](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L539)
 
 ***
 
 ### AgentCandidateExecutorWorkspaceFile
 
-Defined in: [candidate-execution/types.ts:548](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L548)
+Defined in: [candidate-execution/types.ts:542](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L542)
 
 #### Properties
 
@@ -3836,20 +3808,48 @@ Defined in: [candidate-execution/types.ts:548](https://github.com/tangle-network
 
 > `readonly` **path**: `string`
 
-Defined in: [candidate-execution/types.ts:549](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L549)
+Defined in: [candidate-execution/types.ts:543](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L543)
 
 ##### mode
 
 > `readonly` **mode**: `number`
 
-Defined in: [candidate-execution/types.ts:550](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L550)
+Defined in: [candidate-execution/types.ts:544](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L544)
 
 ##### bytes
 
 > `readonly` **bytes**: `Uint8Array`
 
+Defined in: [candidate-execution/types.ts:545](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L545)
+
+***
+
+### AgentCandidateExecutorProfileFile
+
+Defined in: [candidate-execution/types.ts:549](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L549)
+
+One exact profile file supplied to an evaluator-owned executor.
+
+#### Properties
+
+##### path
+
+> `readonly` **path**: `string`
+
+Defined in: [candidate-execution/types.ts:550](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L550)
+
+##### mode
+
+> `readonly` **mode**: `number`
+
 Defined in: [candidate-execution/types.ts:551](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L551)
 
+##### bytes
+
+> `readonly` **bytes**: `Uint8Array`
+
+Defined in: [candidate-execution/types.ts:552](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L552)
+
 ***
 
 ### AgentCandidateWorkspaceArchiveLimits
@@ -6464,7 +6464,7 @@ Defined in: [improvement/reflective-generator.ts:22](https://github.com/tangle-n
 
 ### RunKnowledgeImprovementJobOptions
 
-Defined in: [knowledge/improvement-job.ts:42](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L42)
+Defined in: [knowledge/improvement-job.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L46)
 
 #### Extends
 
@@ -6476,61 +6476,61 @@ Defined in: [knowledge/improvement-job.ts:42](https://github.com/tangle-network/
 
 > **budget**: [`Budget`](runtime.md#budget-12)
 
-Defined in: [knowledge/improvement-job.ts:44](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L44)
+Defined in: [knowledge/improvement-job.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L48)
 
 ##### readinessCheck?
 
 > `optional` **readinessCheck?**: [`KnowledgeReadinessCheck`](#knowledgereadinesscheck)
 
-Defined in: [knowledge/improvement-job.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L45)
+Defined in: [knowledge/improvement-job.ts:49](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L49)
 
 ##### backend?
 
 > `optional` **backend?**: [`ExecutorConfig`](runtime.md#executorconfig)
 
-Defined in: [knowledge/improvement-job.ts:46](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L46)
+Defined in: [knowledge/improvement-job.ts:50](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L50)
 
 ##### makeWorkerAgent?
 
 > `optional` **makeWorkerAgent?**: [`MakeWorkerAgent`](runtime.md#makeworkeragent)
 
-Defined in: [knowledge/improvement-job.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L47)
+Defined in: [knowledge/improvement-job.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L51)
 
 ##### harness?
 
 > `optional` **harness?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L48)
+Defined in: [knowledge/improvement-job.ts:52](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L52)
 
 ##### supervisorModel?
 
 > `optional` **supervisorModel?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:49](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L49)
+Defined in: [knowledge/improvement-job.ts:53](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L53)
 
 ##### supervisorSystemPrompt?
 
 > `optional` **supervisorSystemPrompt?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:50](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L50)
+Defined in: [knowledge/improvement-job.ts:54](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L54)
 
 ##### superviseOptions?
 
 > `optional` **superviseOptions?**: `Partial`\<`Omit`\<[`SuperviseOptions`](runtime.md#superviseoptions), `"backend"` \| `"budget"` \| `"makeWorkerAgent"` \| `"deliverable"` \| `"allowedModels"`\>\>
 
-Defined in: [knowledge/improvement-job.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L51)
+Defined in: [knowledge/improvement-job.ts:55](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L55)
 
 ##### allowedModels?
 
 > `optional` **allowedModels?**: readonly `string`[]
 
-Defined in: [knowledge/improvement-job.ts:57](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L57)
+Defined in: [knowledge/improvement-job.ts:61](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L61)
 
 ##### runSupervised?
 
 > `optional` **runSupervised?**: (`profile`, `task`, `opts`) => `Promise`\<[`SupervisedResult`](runtime.md#supervisedresult)\<`unknown`\>\>
 
-Defined in: [knowledge/improvement-job.ts:58](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L58)
+Defined in: [knowledge/improvement-job.ts:62](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L62)
 
 ###### Parameters
 
@@ -6554,19 +6554,19 @@ Defined in: [knowledge/improvement-job.ts:58](https://github.com/tangle-network/
 
 > `optional` **candidateArtifacts?**: [`AgentCandidateOutputArtifactPort`](#agentcandidateoutputartifactport)
 
-Defined in: [knowledge/improvement-job.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L63)
+Defined in: [knowledge/improvement-job.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L67)
 
 ##### approval?
 
 > `optional` **approval?**: [`ApprovedKnowledgeImprovementCandidate`](#approvedknowledgeimprovementcandidate)
 
-Defined in: [knowledge/improvement-job.ts:64](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L64)
+Defined in: [knowledge/improvement-job.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L68)
 
 ##### onMeasurement?
 
 > `optional` **onMeasurement?**: (`measurement`) => `void` \| `Promise`\<`void`\>
 
-Defined in: [knowledge/improvement-job.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L65)
+Defined in: [knowledge/improvement-job.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L69)
 
 ###### Parameters
 
@@ -6582,7 +6582,7 @@ Defined in: [knowledge/improvement-job.ts:65](https://github.com/tangle-network/
 
 ### ApprovedKnowledgeImprovementCandidate
 
-Defined in: [knowledge/improvement-job.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L68)
+Defined in: [knowledge/improvement-job.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L72)
 
 #### Properties
 
@@ -6590,30 +6590,40 @@ Defined in: [knowledge/improvement-job.ts:68](https://github.com/tangle-network/
 
 > **proposal**: `AgentImprovementProposal`
 
-Defined in: [knowledge/improvement-job.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L69)
+Defined in: [knowledge/improvement-job.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L73)
 
 ##### review
 
 > **review**: `AgentImprovementReview`
 
-Defined in: [knowledge/improvement-job.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L70)
+Defined in: [knowledge/improvement-job.ts:74](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L74)
 
-##### authorizeReview
+##### activation
 
-> **authorizeReview**: (`review`, `proposal`) => `boolean` \| `Promise`\<`boolean`\>
+> **activation**: `AgentImprovementActivation`
 
-Defined in: [knowledge/improvement-job.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L71)
+Defined in: [knowledge/improvement-job.ts:75](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L75)
+
+##### authorizeActivation
+
+> **authorizeActivation**: (`activation`, `proposal`, `review`) => `boolean` \| `Promise`\<`boolean`\>
+
+Defined in: [knowledge/improvement-job.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L76)
 
 ###### Parameters
 
-###### review
+###### activation
 
-`AgentImprovementReview`
+`AgentImprovementActivation`
 
 ###### proposal
 
 `AgentImprovementProposal`
 
+###### review
+
+`AgentImprovementReview`
+
 ###### Returns
 
 `boolean` \| `Promise`\<`boolean`\>
@@ -6622,7 +6632,7 @@ Defined in: [knowledge/improvement-job.ts:71](https://github.com/tangle-network/
 
 ### KnowledgeImprovementJobMeasurement
 
-Defined in: [knowledge/improvement-job.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L77)
+Defined in: [knowledge/improvement-job.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L83)
 
 #### Properties
 
@@ -6630,37 +6640,37 @@ Defined in: [knowledge/improvement-job.ts:77](https://github.com/tangle-network/
 
 > **startedAt**: `string`
 
-Defined in: [knowledge/improvement-job.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L78)
+Defined in: [knowledge/improvement-job.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L84)
 
 ##### finishedAt
 
 > **finishedAt**: `string`
 
-Defined in: [knowledge/improvement-job.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L79)
+Defined in: [knowledge/improvement-job.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L85)
 
 ##### durationMs
 
 > **durationMs**: `number`
 
-Defined in: [knowledge/improvement-job.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L80)
+Defined in: [knowledge/improvement-job.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L86)
 
 ##### updateCalls
 
 > **updateCalls**: `number`
 
-Defined in: [knowledge/improvement-job.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L81)
+Defined in: [knowledge/improvement-job.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L87)
 
 ##### updateDurationMs
 
 > **updateDurationMs**: `number`
 
-Defined in: [knowledge/improvement-job.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L82)
+Defined in: [knowledge/improvement-job.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L88)
 
 ##### supervisedSpent
 
 > **supervisedSpent**: `object`
 
-Defined in: [knowledge/improvement-job.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L83)
+Defined in: [knowledge/improvement-job.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L89)
 
 ###### iterations
 
@@ -6690,7 +6700,7 @@ Defined in: [knowledge/improvement-job.ts:83](https://github.com/tangle-network/
 
 ### KnowledgeImprovementJobResult
 
-Defined in: [knowledge/improvement-job.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L93)
+Defined in: [knowledge/improvement-job.ts:99](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L99)
 
 #### Properties
 
@@ -6698,37 +6708,37 @@ Defined in: [knowledge/improvement-job.ts:93](https://github.com/tangle-network/
 
 > **improvement**: `KnowledgeImprovementResult`
 
-Defined in: [knowledge/improvement-job.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L94)
+Defined in: [knowledge/improvement-job.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L100)
 
 ##### candidateKnowledge?
 
 > `optional` **candidateKnowledge?**: `AgentCandidateKnowledge`
 
-Defined in: [knowledge/improvement-job.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L95)
+Defined in: [knowledge/improvement-job.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L101)
 
 ##### measurement
 
 > **measurement**: [`KnowledgeImprovementJobMeasurement`](#knowledgeimprovementjobmeasurement)
 
-Defined in: [knowledge/improvement-job.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L96)
+Defined in: [knowledge/improvement-job.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L102)
 
 ##### promoted
 
 > **promoted**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L97)
+Defined in: [knowledge/improvement-job.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L103)
 
 ##### blocked
 
 > **blocked**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L98)
+Defined in: [knowledge/improvement-job.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L104)
 
 ***
 
 ### AgentKnowledgeReadinessCheckOptions
 
-Defined in: [knowledge/improvement-job.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L101)
+Defined in: [knowledge/improvement-job.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L107)
 
 #### Properties
 
@@ -6736,37 +6746,37 @@ Defined in: [knowledge/improvement-job.ts:101](https://github.com/tangle-network
 
 > **goal**: `string`
 
-Defined in: [knowledge/improvement-job.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L102)
+Defined in: [knowledge/improvement-job.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L108)
 
 ##### readinessSpecs?
 
 > `optional` **readinessSpecs?**: readonly `KnowledgeReadinessSpec`[]
 
-Defined in: [knowledge/improvement-job.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L103)
+Defined in: [knowledge/improvement-job.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L109)
 
 ##### readinessTaskId?
 
 > `optional` **readinessTaskId?**: `string`
 
-Defined in: [knowledge/improvement-job.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L104)
+Defined in: [knowledge/improvement-job.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L110)
 
 ##### readiness?
 
 > `optional` **readiness?**: `Omit`\<`BuildEvalKnowledgeBundleOptions`, `"taskId"` \| `"index"` \| `"specs"`\>
 
-Defined in: [knowledge/improvement-job.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L105)
+Defined in: [knowledge/improvement-job.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L111)
 
 ##### strict?
 
 > `optional` **strict?**: `boolean`
 
-Defined in: [knowledge/improvement-job.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L106)
+Defined in: [knowledge/improvement-job.ts:112](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L112)
 
 ##### kbQuality?
 
 > `optional` **kbQuality?**: `KnowledgeBaseQualityOptions`
 
-Defined in: [knowledge/improvement-job.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L107)
+Defined in: [knowledge/improvement-job.ts:113](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L113)
 
 ***
 
@@ -10577,7 +10587,7 @@ A complete profile that can be frozen without losing behavior.
 
 > **diffs**: readonly `AgentProfileDiff`[]
 
-Applied in order. Each exact diff is content-addressed into lineage.
+Applied in order before the resulting profile is frozen into the bundle.
 
 ***
 
@@ -10833,7 +10843,7 @@ Secret-free response from the service's reservation endpoint.
 
 ### AgentCandidateOutputPurpose
 
-> **AgentCandidateOutputPurpose** = `"execution-plan"` \| `"materialization-receipt"` \| `"candidate-workspace-manifest"` \| `"candidate-workspace-archive"` \| `"task-manifest"` \| `"task-archive"` \| `"task-patch"` \| `"task-output"` \| `"task-outcome"` \| `"memory-after-manifest"` \| `"memory-after-archive"` \| `"grader-evidence"` \| `"benchmark-result"` \| `"model-settlement"` \| `"trace"` \| `"executor-capture"` \| `"run-receipt"` \| `"knowledge-retrieval-config"` \| `"knowledge-evaluation"` \| `"failure-evidence"`
+> **AgentCandidateOutputPurpose** = `"execution-plan"` \| `"materialization-receipt"` \| `"candidate-workspace-manifest"` \| `"candidate-workspace-archive"` \| `"task-manifest"` \| `"task-archive"` \| `"task-patch"` \| `"task-output"` \| `"task-outcome"` \| `"memory-after-manifest"` \| `"memory-after-archive"` \| `"grader-evidence"` \| `"benchmark-result"` \| `"model-settlement"` \| `"trace"` \| `"executor-native-evidence"` \| `"executor-capture"` \| `"run-receipt"` \| `"knowledge-retrieval-config"` \| `"knowledge-evaluation"` \| `"failure-evidence"`
 
 Defined in: [candidate-execution/types.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L45)
 
@@ -10843,7 +10853,7 @@ Defined in: [candidate-execution/types.ts:45](https://github.com/tangle-network/
 
 > **AgentCandidateModelLimits** = `Pick`\<`AgentCandidateExecutionLimits`, `"maxModelCalls"` \| `"maxInputTokens"` \| `"maxOutputTokens"` \| `"maxCostUsd"`\>
 
-Defined in: [candidate-execution/types.ts:164](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L164)
+Defined in: [candidate-execution/types.ts:165](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L165)
 
 Limits mechanically enforced by the evaluator-owned model gateway.
 
@@ -10853,7 +10863,7 @@ Limits mechanically enforced by the evaluator-owned model gateway.
 
 > **AgentCandidateExecutorTaskOutcomeCapture** = \{ `kind`: `"workspace"`; `resultTree`: `string`; `afterState`: `AgentCandidateWorkspaceManifestMaterial`; `archive`: `Uint8Array`; `gitDiff`: `Uint8Array`; \} \| \{ `kind`: `"output"`; `bytes`: `Uint8Array`; \}
 
-Defined in: [candidate-execution/types.ts:369](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L369)
+Defined in: [candidate-execution/types.ts:355](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L355)
 
 Raw evaluator capture made only after the candidate process is dead.
 
@@ -10913,7 +10923,7 @@ Exact evaluator-captured final output bytes.
 
 > **VerifiedAgentCandidateTaskOutcome** = \{ `kind`: `"workspace"`; `evidence`: `PersistedTaskOutcomeEvidence`\<`"workspace"`\>; `patch`: `Uint8Array`; `[verifiedTaskOutcomeBrand]`: `true`; \} \| \{ `kind`: `"output"`; `evidence`: `PersistedTaskOutcomeEvidence`\<`"output"`\>; `spec`: `AgentCandidateTaskOutputSpec`; `bytes`: `Uint8Array`; `[verifiedTaskOutcomeBrand]`: `true`; \}
 
-Defined in: [candidate-execution/types.ts:412](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L412)
+Defined in: [candidate-execution/types.ts:398](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L398)
 
 Branded task outcome that has survived independent evaluator verification.
 
@@ -10923,7 +10933,7 @@ Branded task outcome that has survived independent evaluator verification.
 
 > **AgentCandidateRunFinalization** = \{ `succeeded`: `true`; `receipt`: [`CanonicalCandidateDocument`](#canonicalcandidatedocument)\<`AgentCandidateRunReceipt`\>; `artifacts`: \{ `executorCapture`: `AgentCandidateArtifactRef`; `modelSettlement`: `AgentCandidateArtifactRef`; `taskOutcome`: `AgentCandidateArtifactRef`; `benchmarkResult`: `AgentCandidateArtifactRef`; `runReceipt`: `AgentCandidateArtifactRef`; \}; \} \| \{ `succeeded`: `false`; `reason`: `string`; `partial`: \{ `executionId`: `string`; `bundleDigest`: `Sha256Digest`; `executionPlanDigest`: `Sha256Digest`; `materializationReceiptDigest`: `Sha256Digest`; `termination?`: `AgentCandidateTermination`; \}; `usage`: `AgentCandidateFixedSpend` \| `null`; \}
 
-Defined in: [candidate-execution/types.ts:560](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L560)
+Defined in: [candidate-execution/types.ts:555](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L555)
 
 #### Union Members
 
@@ -11876,7 +11886,7 @@ Environment variable containing the materialized retrieval configuration path.
 
 > `const` **CANDIDATE\_TRACE\_TAGS**: `object`
 
-Defined in: [candidate-execution/types.ts:587](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L587)
+Defined in: [candidate-execution/types.ts:582](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L582)
 
 Protected trace tags that bind a run to one prepared candidate execution.
 
@@ -11904,7 +11914,7 @@ Protected trace tags that bind a run to one prepared candidate execution.
 
 > `const` **CANDIDATE\_TRACE\_ENV**: `object`
 
-Defined in: [candidate-execution/types.ts:595](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L595)
+Defined in: [candidate-execution/types.ts:590](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/types.ts#L590)
 
 Environment keys used to propagate immutable candidate trace identity.
 
@@ -12276,7 +12286,7 @@ Maximum completion tokens, sent as OpenAI-compatible `max_tokens`. Omit for prov
 
 > **buildAgentCandidateBundle**(`input`): `AgentCandidateBundle`
 
-Defined in: [candidate-execution/builder.ts:74](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/builder.ts#L74)
+Defined in: [candidate-execution/builder.ts:73](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/builder.ts#L73)
 
 Compile one measured profile/code candidate into the immutable execution
 contract. Code bytes are re-read and verified by agent-eval before they are
@@ -12374,7 +12384,7 @@ Revoke reservations held by a prepared candidate that will not be executed.
 
 > **executePreparedAgentCandidate**(`prepared`, `options`): `Promise`\<[`AgentCandidateRunFinalization`](#agentcandidaterunfinalization)\>
 
-Defined in: [candidate-execution/execute.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L77)
+Defined in: [candidate-execution/execute.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/execute.ts#L80)
 
 Executes and finalizes one durably claimed candidate without exposing an unproven result.
 
@@ -12468,7 +12478,7 @@ Persist evaluator evidence, read it back, and bind the returned locator to the e
 
 > **prepareAgentCandidateExecution**(`candidate`, `task`, `ports`, `options?`): `Promise`\<[`PreparedAgentCandidateExecution`](#preparedagentcandidateexecution)\>
 
-Defined in: [candidate-execution/prepare.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L88)
+Defined in: [candidate-execution/prepare.ts:92](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/prepare.ts#L92)
 
 Materializes a verified candidate into one immutable evaluator-owned execution plan.
 
@@ -12500,7 +12510,7 @@ Materializes a verified candidate into one immutable evaluator-owned execution p
 
 > **parseExactAgentProfile**(`input`, `label`): `AgentProfile`
 
-Defined in: [candidate-execution/profile.ts:187](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L187)
+Defined in: [candidate-execution/profile.ts:190](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L190)
 
 Parse a complete profile without silently discarding unsupported fields.
 
@@ -12524,7 +12534,7 @@ Parse a complete profile without silently discarding unsupported fields.
 
 > **parseExactAgentProfileDiff**(`input`, `label`): `AgentProfileDiff`
 
-Defined in: [candidate-execution/profile.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L199)
+Defined in: [candidate-execution/profile.ts:202](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L202)
 
 Parse a profile diff without silently discarding unsupported fields.
 
@@ -12548,7 +12558,7 @@ Parse a profile diff without silently discarding unsupported fields.
 
 > **applyExactAgentProfileDiff**(`baseInput`, `diffInput`, `label`): `AgentProfile`
 
-Defined in: [candidate-execution/profile.ts:211](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L211)
+Defined in: [candidate-execution/profile.ts:214](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/profile.ts#L214)
 
 Apply one exact diff and reject any value that cannot be preserved canonically.
 
@@ -12600,7 +12610,7 @@ it to cross into candidate execution or durable receipt finalization.
 
 > **recoverExpiredAgentCandidateExecution**(`options`): `Promise`\<[`AgentCandidateExecutionFinishResult`](#agentcandidateexecutionfinishresult)\>
 
-Defined in: [candidate-execution/recover.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L51)
+Defined in: [candidate-execution/recover.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/candidate-execution/recover.ts#L48)
 
 Close an expired crashed attempt from persisted non-secret handles, then record failure.
 
@@ -13290,7 +13300,7 @@ Build the starting instruction for a coder agent tasked with implementing a new
 
 > **applyImprovementWinnerToProfile**(`profile`, `surface`, `winner`): `AgentProfile`
 
-Defined in: [improvement/improve.ts:444](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L444)
+Defined in: [improvement/improve.ts:456](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L456)
 
 Apply a promoted winner surface back into the profile field for `surface`.
  Returns a shallow copy; never mutates the input profile.
@@ -13319,7 +13329,7 @@ Apply a promoted winner surface back into the profile field for `surface`.
 
 > **improve**\<`TScenario`, `TArtifact`\>(`profile`, `findings`, `opts`): `Promise`\<[`ImproveResult`](#improveresult)\<`TScenario`, `TArtifact`\>\>
 
-Defined in: [improvement/improve.ts:507](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L507)
+Defined in: [improvement/improve.ts:519](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L519)
 
 Run the held-out-gated self-improvement loop on ONE profile surface.
 
@@ -13501,7 +13511,7 @@ Cheap no-sandbox `CandidateGenerator` (the `shots=1` setting): draft surface edi
 
 > **createAgentKnowledgeReadinessCheck**(`options`): [`KnowledgeReadinessCheck`](#knowledgereadinesscheck)
 
-Defined in: [knowledge/improvement-job.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L111)
+Defined in: [knowledge/improvement-job.ts:117](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L117)
 
 Build the default readiness check backed by `@tangle-network/agent-knowledge` validation and scoring.
 
@@ -13521,7 +13531,7 @@ Build the default readiness check backed by `@tangle-network/agent-knowledge` va
 
 > **runKnowledgeImprovementJob**(`options`): `Promise`\<[`KnowledgeImprovementJobResult`](#knowledgeimprovementjobresult)\>
 
-Defined in: [knowledge/improvement-job.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L142)
+Defined in: [knowledge/improvement-job.ts:148](https://github.com/tangle-network/agent-runtime/blob/main/src/knowledge/improvement-job.ts#L148)
 
 Produce a frozen KB candidate, and promote it only when an exact signed review is supplied.
 
diff --git a/docs/api/intelligence.md b/docs/api/intelligence.md
index 85b1535a..79e73852 100644
--- a/docs/api/intelligence.md
+++ b/docs/api/intelligence.md
@@ -65,6 +65,122 @@ Defined in: [intelligence/capability.ts:247](https://github.com/tangle-network/a
 
 Defined in: [intelligence/capability.ts:248](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/capability.ts#L248)
 
+***
+
+### AgentCandidateExperimentCellExecutionError
+
+Defined in: [intelligence/improvement-cycle.ts:123](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L123)
+
+A failed baseline or candidate cell with its complete Runtime failure result.
+
+#### Extends
+
+- `Error`
+
+#### Constructors
+
+##### Constructor
+
+> **new AgentCandidateExperimentCellExecutionError**(`finalization`): [`AgentCandidateExperimentCellExecutionError`](#agentcandidateexperimentcellexecutionerror)
+
+Defined in: [intelligence/improvement-cycle.ts:126](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L126)
+
+###### Parameters
+
+###### finalization
+
+###### succeeded
+
+`false`
+
+###### reason
+
+`string`
+
+###### partial
+
+\{ `executionId`: `string`; `bundleDigest`: `` `sha256:${string}` ``; `executionPlanDigest`: `` `sha256:${string}` ``; `materializationReceiptDigest`: `` `sha256:${string}` ``; `termination?`: `AgentCandidateTermination`; \}
+
+###### partial.executionId
+
+`string`
+
+###### partial.bundleDigest
+
+`` `sha256:${string}` ``
+
+###### partial.executionPlanDigest
+
+`` `sha256:${string}` ``
+
+###### partial.materializationReceiptDigest
+
+`` `sha256:${string}` ``
+
+###### partial.termination?
+
+`AgentCandidateTermination`
+
+###### usage
+
+`AgentCandidateFixedSpend` \| `null`
+
+Independent evaluator-gateway usage, even when execution or trace capture failed.
+
+###### Returns
+
+[`AgentCandidateExperimentCellExecutionError`](#agentcandidateexperimentcellexecutionerror)
+
+###### Overrides
+
+`Error.constructor`
+
+#### Properties
+
+##### finalization
+
+> `readonly` **finalization**: `object`
+
+Defined in: [intelligence/improvement-cycle.ts:124](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L124)
+
+###### succeeded
+
+> **succeeded**: `false`
+
+###### reason
+
+> **reason**: `string`
+
+###### partial
+
+> **partial**: `object`
+
+###### partial.executionId
+
+> **executionId**: `string`
+
+###### partial.bundleDigest
+
+> **bundleDigest**: `` `sha256:${string}` ``
+
+###### partial.executionPlanDigest
+
+> **executionPlanDigest**: `` `sha256:${string}` ``
+
+###### partial.materializationReceiptDigest
+
+> **materializationReceiptDigest**: `` `sha256:${string}` ``
+
+###### partial.termination?
+
+> `optional` **termination?**: `AgentCandidateTermination`
+
+###### usage
+
+> **usage**: `AgentCandidateFixedSpend` \| `null`
+
+Independent evaluator-gateway usage, even when execution or trace capture failed.
+
 ## Interfaces
 
 ### CredentialRef
@@ -1076,121 +1192,301 @@ Intelligence-class spend ceiling. `0` refuses every intelligence spawn; `null` u
 
 ***
 
-### ProposeAgentImprovementOptions
+### AgentCandidateExperimentCellPlacement
 
-Defined in: [intelligence/improvement-cycle.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L85)
+Defined in: [intelligence/improvement-cycle.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L83)
 
-#### Type Parameters
+#### Extended by
 
-##### TScenario
+- [`ExecuteAgentCandidateExperimentCellOptions`](#executeagentcandidateexperimentcelloptions)
 
-`TScenario` *extends* `Scenario`
+#### Properties
 
-##### TArtifact
+##### executionId
 
-`TArtifact`
+> **executionId**: `string`
 
-#### Properties
+Defined in: [intelligence/improvement-cycle.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L84)
 
-##### runId
+##### attempt?
 
-> **runId**: `string`
+> `optional` **attempt?**: `number`
+
+Defined in: [intelligence/improvement-cycle.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L85)
+
+##### executionRoots
+
+> **executionRoots**: `object`
 
 Defined in: [intelligence/improvement-cycle.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L86)
 
-##### profile
+###### taskRoot
 
-> **profile**: `AgentProfile`
+> **taskRoot**: `string`
+
+###### candidateRoot?
+
+> `optional` **candidateRoot?**: `string`
+
+##### stagingRoots
+
+> **stagingRoots**: `object`
 
 Defined in: [intelligence/improvement-cycle.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L87)
 
-##### analysis
+###### taskRoot
 
-> **analysis**: `Omit`\<[`RunAnalystLoopOpts`](analyst-loop.md#runanalystloopopts), `"runId"` \| `"improvementAdapter"` \| `"autoApply"`\>
+> **taskRoot**: `string`
+
+###### candidateRoot?
+
+> `optional` **candidateRoot?**: `string`
+
+###### profileRoot
+
+> **profileRoot**: `string`
+
+##### ports
+
+> **ports**: [`AgentCandidateExecutionPorts`](index.md#agentcandidateexecutionports)
 
 Defined in: [intelligence/improvement-cycle.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L88)
 
-##### improvement
+##### preparation?
 
-> **improvement**: [`ImproveOptions`](index.md#improveoptions)\<`TScenario`, `TArtifact`\>
+> `optional` **preparation?**: [`PrepareAgentCandidateExecutionOptions`](index.md#prepareagentcandidateexecutionoptions)
 
 Defined in: [intelligence/improvement-cycle.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L89)
 
-##### buildCandidate
+##### execution
+
+> **execution**: [`ExecutePreparedAgentCandidateOptions`](index.md#executepreparedagentcandidateoptions)
+
+Defined in: [intelligence/improvement-cycle.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L90)
+
+***
+
+### RunAgentCandidateExperimentOptions
 
-> **buildCandidate**: (`input`) => `AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput) \| `Promise`\<`AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)\>
+Defined in: [intelligence/improvement-cycle.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L93)
 
-Defined in: [intelligence/improvement-cycle.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L91)
+#### Extends
 
-Freezes the measured winner into the exact bundle reviewed for execution.
+- `Omit`\<`CompareCandidateExperimentOptions`, `"experiment"` \| `"measurements"`\>
+
+#### Properties
+
+##### experiment
+
+> **experiment**: `AgentCandidateExperiment`
+
+Defined in: [intelligence/improvement-cycle.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L95)
+
+##### placeCell
+
+> **placeCell**: (`input`) => [`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement) \| `Promise`\<[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement)\>
+
+Defined in: [intelligence/improvement-cycle.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L96)
 
 ###### Parameters
 
 ###### input
 
-###### analysis
+`CandidateExperimentExecutionInput`
 
-[`RunAnalystLoopResult`](analyst-loop.md#runanalystloopresult)
+###### Returns
 
-###### improvement
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement) \| `Promise`\<[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement)\>
 
-[`ImproveResult`](index.md#improveresult)\<`TScenario`, `TArtifact`\>
+##### maxConcurrency?
 
-###### Returns
+> `optional` **maxConcurrency?**: `number`
 
-`AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput) \| `Promise`\<`AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)\>
+Defined in: [intelligence/improvement-cycle.ts:99](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L99)
 
-##### now?
+##### signal?
 
-> `optional` **now?**: () => `Date`
+> `optional` **signal?**: `AbortSignal`
 
-Defined in: [intelligence/improvement-cycle.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L98)
+Defined in: [intelligence/improvement-cycle.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L100)
 
-###### Returns
+***
 
-`Date`
+### RunAgentCandidateExperimentResult
+
+Defined in: [intelligence/improvement-cycle.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L103)
+
+#### Properties
+
+##### experiment
+
+> **experiment**: `AgentCandidateExperiment`
+
+Defined in: [intelligence/improvement-cycle.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L104)
+
+##### measurements
+
+> **measurements**: `AgentCandidateExperimentMeasurement`[]
+
+Defined in: [intelligence/improvement-cycle.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L105)
+
+##### evaluation
+
+> **evaluation**: `AgentImprovementMeasuredComparison`
+
+Defined in: [intelligence/improvement-cycle.ts:106](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L106)
 
 ***
 
-### ProposeAgentImprovementResult
+### ExecuteAgentCandidateExperimentCellOptions
 
-Defined in: [intelligence/improvement-cycle.ts:101](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L101)
+Defined in: [intelligence/improvement-cycle.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L109)
 
-#### Type Parameters
+#### Extends
 
-##### TScenario
+- `CandidateExperimentExecutionInput`.[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement)
 
-`TScenario` *extends* `Scenario`
+#### Properties
 
-##### TArtifact
+##### executionId
 
-`TArtifact`
+> **executionId**: `string`
+
+Defined in: [intelligence/improvement-cycle.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L84)
+
+###### Inherited from
+
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement).[`executionId`](#executionid)
+
+##### attempt?
+
+> `optional` **attempt?**: `number`
+
+Defined in: [intelligence/improvement-cycle.ts:85](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L85)
+
+###### Inherited from
+
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement).[`attempt`](#attempt)
+
+##### executionRoots
+
+> **executionRoots**: `object`
+
+Defined in: [intelligence/improvement-cycle.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L86)
+
+###### taskRoot
+
+> **taskRoot**: `string`
+
+###### candidateRoot?
+
+> `optional` **candidateRoot?**: `string`
+
+###### Inherited from
+
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement).[`executionRoots`](#executionroots)
+
+##### stagingRoots
+
+> **stagingRoots**: `object`
+
+Defined in: [intelligence/improvement-cycle.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L87)
+
+###### taskRoot
+
+> **taskRoot**: `string`
+
+###### candidateRoot?
+
+> `optional` **candidateRoot?**: `string`
+
+###### profileRoot
+
+> **profileRoot**: `string`
+
+###### Inherited from
+
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement).[`stagingRoots`](#stagingroots)
+
+##### ports
+
+> **ports**: [`AgentCandidateExecutionPorts`](index.md#agentcandidateexecutionports)
+
+Defined in: [intelligence/improvement-cycle.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L88)
+
+###### Inherited from
+
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement).[`ports`](#ports-1)
+
+##### preparation?
+
+> `optional` **preparation?**: [`PrepareAgentCandidateExecutionOptions`](index.md#prepareagentcandidateexecutionoptions)
+
+Defined in: [intelligence/improvement-cycle.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L89)
+
+###### Inherited from
+
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement).[`preparation`](#preparation)
+
+##### execution
+
+> **execution**: [`ExecutePreparedAgentCandidateOptions`](index.md#executepreparedagentcandidateoptions)
+
+Defined in: [intelligence/improvement-cycle.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L90)
+
+###### Inherited from
+
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement).[`execution`](#execution)
+
+***
+
+### VerifyCandidateExecutionEvidenceOptions
+
+Defined in: [intelligence/improvement-cycle.ts:113](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L113)
 
 #### Properties
 
-##### analysis
+##### experiment
 
-> **analysis**: [`RunAnalystLoopResult`](analyst-loop.md#runanalystloopresult)
+> **experiment**: `AgentCandidateExperiment`
 
-Defined in: [intelligence/improvement-cycle.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L102)
+Defined in: [intelligence/improvement-cycle.ts:114](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L114)
 
-##### improvement
+##### arm
 
-> **improvement**: [`ImproveResult`](index.md#improveresult)\<`TScenario`, `TArtifact`\>
+> **arm**: `"candidate"` \| `"baseline"`
 
-Defined in: [intelligence/improvement-cycle.ts:103](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L103)
+Defined in: [intelligence/improvement-cycle.ts:115](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L115)
 
-##### proposal
+##### benchmarkCell
 
-> **proposal**: `AgentImprovementProposal`
+> **benchmarkCell**: `AgentCandidateBenchmarkCellRef`
 
-Defined in: [intelligence/improvement-cycle.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L104)
+Defined in: [intelligence/improvement-cycle.ts:116](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L116)
+
+##### seed
+
+> **seed**: `number`
+
+Defined in: [intelligence/improvement-cycle.ts:117](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L117)
+
+##### attempt?
+
+> `optional` **attempt?**: `number`
+
+Defined in: [intelligence/improvement-cycle.ts:118](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L118)
+
+##### resolvedResources?
+
+> `optional` **resolvedResources?**: `ReadonlyMap`\<`` `sha256:${string}` ``, `string`\>
+
+Defined in: [intelligence/improvement-cycle.ts:119](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L119)
 
 ***
 
 ### CreateAgentImprovementProposalOptions
 
-Defined in: [intelligence/improvement-cycle.ts:107](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L107)
+Defined in: [intelligence/improvement-cycle.ts:133](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L133)
 
 #### Properties
 
@@ -1198,37 +1494,25 @@ Defined in: [intelligence/improvement-cycle.ts:107](https://github.com/tangle-ne
 
 > **runId**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:108](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L108)
-
-##### baselineProfile
-
-> **baselineProfile**: `AgentProfile`
-
-Defined in: [intelligence/improvement-cycle.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L109)
+Defined in: [intelligence/improvement-cycle.ts:134](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L134)
 
 ##### findings
 
 > **findings**: readonly `AnalystFinding`[]
 
-Defined in: [intelligence/improvement-cycle.ts:110](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L110)
+Defined in: [intelligence/improvement-cycle.ts:135](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L135)
 
 ##### evaluation
 
 > **evaluation**: `AgentImprovementMeasuredComparison`
 
-Defined in: [intelligence/improvement-cycle.ts:111](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L111)
-
-##### candidateBundle
-
-> **candidateBundle**: `AgentCandidateBundle` \| [`AgentCandidateBundleInput`](index.md#agentcandidatebundleinput)
-
-Defined in: [intelligence/improvement-cycle.ts:112](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L112)
+Defined in: [intelligence/improvement-cycle.ts:136](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L136)
 
 ##### now?
 
 > `optional` **now?**: () => `Date`
 
-Defined in: [intelligence/improvement-cycle.ts:113](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L113)
+Defined in: [intelligence/improvement-cycle.ts:137](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L137)
 
 ###### Returns
 
@@ -1238,7 +1522,7 @@ Defined in: [intelligence/improvement-cycle.ts:113](https://github.com/tangle-ne
 
 ### ReviewAgentImprovementInput
 
-Defined in: [intelligence/improvement-cycle.ts:126](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L126)
+Defined in: [intelligence/improvement-cycle.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L142)
 
 #### Properties
 
@@ -1246,31 +1530,31 @@ Defined in: [intelligence/improvement-cycle.ts:126](https://github.com/tangle-ne
 
 > **decision**: `AgentImprovementReviewDecision`
 
-Defined in: [intelligence/improvement-cycle.ts:127](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L127)
+Defined in: [intelligence/improvement-cycle.ts:143](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L143)
 
 ##### reviewedBy
 
 > **reviewedBy**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:128](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L128)
+Defined in: [intelligence/improvement-cycle.ts:144](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L144)
 
 ##### reason
 
 > **reason**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:129](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L129)
+Defined in: [intelligence/improvement-cycle.ts:145](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L145)
 
 ##### feedback?
 
 > `optional` **feedback?**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:130](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L130)
+Defined in: [intelligence/improvement-cycle.ts:146](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L146)
 
 ##### now?
 
 > `optional` **now?**: () => `Date`
 
-Defined in: [intelligence/improvement-cycle.ts:131](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L131)
+Defined in: [intelligence/improvement-cycle.ts:147](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L147)
 
 ###### Returns
 
@@ -1278,107 +1562,211 @@ Defined in: [intelligence/improvement-cycle.ts:131](https://github.com/tangle-ne
 
 ***
 
-### ExecuteApprovedAgentCandidateOptions
+### CreateAgentImprovementActivationOptions
 
-Defined in: [intelligence/improvement-cycle.ts:134](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L134)
+Defined in: [intelligence/improvement-cycle.ts:150](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L150)
 
 #### Properties
 
-##### proposal
+##### targets
 
-> **proposal**: `AgentImprovementProposal`
+> **targets**: \[`AgentImprovementActivationTarget`, `...AgentImprovementActivationTarget[]`\]
 
-Defined in: [intelligence/improvement-cycle.ts:135](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L135)
+Defined in: [intelligence/improvement-cycle.ts:151](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L151)
 
-##### review
+##### fundingOwner
 
-> **review**: `AgentImprovementReview`
+> **fundingOwner**: `string`
 
-Defined in: [intelligence/improvement-cycle.ts:136](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L136)
+Defined in: [intelligence/improvement-cycle.ts:152](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L152)
+
+##### authorizedBy
+
+> **authorizedBy**: `string`
+
+Defined in: [intelligence/improvement-cycle.ts:153](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L153)
+
+##### now?
+
+> `optional` **now?**: () => `Date`
+
+Defined in: [intelligence/improvement-cycle.ts:154](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L154)
+
+###### Returns
+
+`Date`
+
+***
+
+### ProposeAgentImprovementOptions
+
+Defined in: [intelligence/improvement-cycle.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L157)
+
+#### Type Parameters
+
+##### TScenario
 
-##### authorizeReview
+`TScenario` *extends* `Scenario`
 
-> **authorizeReview**: (`review`, `proposal`) => `boolean` \| `Promise`\<`boolean`\>
+##### TArtifact
 
-Defined in: [intelligence/improvement-cycle.ts:138](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L138)
+`TArtifact`
 
-Product-owned authentication check for the persisted approval record.
+#### Properties
+
+##### runId
+
+> **runId**: `string`
+
+Defined in: [intelligence/improvement-cycle.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L158)
+
+##### profile
+
+> **profile**: `AgentProfile`
+
+Defined in: [intelligence/improvement-cycle.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L159)
+
+##### analysis
+
+> **analysis**: `Omit`\<[`RunAnalystLoopOpts`](analyst-loop.md#runanalystloopopts), `"runId"` \| `"improvementAdapter"` \| `"autoApply"`\>
+
+Defined in: [intelligence/improvement-cycle.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L160)
+
+##### improvement
+
+> **improvement**: [`ImproveOptions`](index.md#improveoptions)\<`TScenario`, `TArtifact`\>
+
+Defined in: [intelligence/improvement-cycle.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L161)
+
+##### buildExperiment
+
+> **buildExperiment**: (`input`) => `AgentCandidateExperiment` \| `Promise`\<`AgentCandidateExperiment`\>
+
+Defined in: [intelligence/improvement-cycle.ts:162](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L162)
 
 ###### Parameters
 
-###### review
+###### input
 
-`AgentImprovementReview`
+###### analysis
 
-###### proposal
+[`RunAnalystLoopResult`](analyst-loop.md#runanalystloopresult)
 
-`AgentImprovementProposal`
+###### improvement
+
+[`ImproveResult`](index.md#improveresult)\<`TScenario`, `TArtifact`\>
 
 ###### Returns
 
-`boolean` \| `Promise`\<`boolean`\>
+`AgentCandidateExperiment` \| `Promise`\<`AgentCandidateExperiment`\>
 
-##### task
+##### placeCell
 
-> **task**: [`AgentCandidateTaskExecution`](index.md#agentcandidatetaskexecution)
+> **placeCell**: (`input`) => [`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement) \| `Promise`\<[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement)\>
 
-Defined in: [intelligence/improvement-cycle.ts:142](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L142)
+Defined in: [intelligence/improvement-cycle.ts:166](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L166)
 
-##### ports
+###### Parameters
 
-> **ports**: [`AgentCandidateExecutionPorts`](index.md#agentcandidateexecutionports)
+###### input
 
-Defined in: [intelligence/improvement-cycle.ts:143](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L143)
+`CandidateExperimentExecutionInput`
 
-##### preparation?
+###### Returns
 
-> `optional` **preparation?**: [`PrepareAgentCandidateExecutionOptions`](index.md#prepareagentcandidateexecutionoptions)
+[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement) \| `Promise`\<[`AgentCandidateExperimentCellPlacement`](#agentcandidateexperimentcellplacement)\>
 
-Defined in: [intelligence/improvement-cycle.ts:144](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L144)
+##### maxConcurrency?
 
-##### execution
+> `optional` **maxConcurrency?**: `number`
 
-> **execution**: [`ExecutePreparedAgentCandidateOptions`](index.md#executepreparedagentcandidateoptions)
+Defined in: [intelligence/improvement-cycle.ts:167](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L167)
 
-Defined in: [intelligence/improvement-cycle.ts:145](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L145)
+##### signal?
+
+> `optional` **signal?**: `AbortSignal`
+
+Defined in: [intelligence/improvement-cycle.ts:168](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L168)
+
+##### candidate?
+
+> `optional` **candidate?**: `object`
+
+Defined in: [intelligence/improvement-cycle.ts:169](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L169)
+
+##### metadata?
+
+> `optional` **metadata?**: `object`
+
+Defined in: [intelligence/improvement-cycle.ts:170](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L170)
+
+###### Index Signature
+
+\[`key`: `string`\]: `AgentCandidateJsonValue`
+
+##### now?
+
+> `optional` **now?**: () => `Date`
+
+Defined in: [intelligence/improvement-cycle.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L171)
+
+###### Returns
+
+`Date`
 
 ***
 
-### VerifyCandidateExecutionEvidenceOptions
+### ProposeAgentImprovementResult
 
-Defined in: [intelligence/improvement-cycle.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L158)
+Defined in: [intelligence/improvement-cycle.ts:174](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L174)
+
+#### Type Parameters
+
+##### TScenario
+
+`TScenario` *extends* `Scenario`
+
+##### TArtifact
+
+`TArtifact`
 
 #### Properties
 
-##### proposal
+##### analysis
 
-> **proposal**: `AgentImprovementProposal`
+> **analysis**: [`RunAnalystLoopResult`](analyst-loop.md#runanalystloopresult)
+
+Defined in: [intelligence/improvement-cycle.ts:175](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L175)
+
+##### improvement
+
+> **improvement**: [`ImproveResult`](index.md#improveresult)\<`TScenario`, `TArtifact`\>
 
-Defined in: [intelligence/improvement-cycle.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L159)
+Defined in: [intelligence/improvement-cycle.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L176)
 
-##### review
+##### experiment
 
-> **review**: `AgentImprovementReview`
+> **experiment**: `AgentCandidateExperiment`
 
-Defined in: [intelligence/improvement-cycle.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L160)
+Defined in: [intelligence/improvement-cycle.ts:177](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L177)
 
-##### expectedCount
+##### measurements
 
-> **expectedCount**: `number`
+> **measurements**: `AgentCandidateExperimentMeasurement`[]
 
-Defined in: [intelligence/improvement-cycle.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L161)
+Defined in: [intelligence/improvement-cycle.ts:178](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L178)
 
-##### resolvedResources?
+##### proposal
 
-> `optional` **resolvedResources?**: `ReadonlyMap`\<`` `sha256:${string}` ``, `string`\>
+> **proposal**: `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:162](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L162)
+Defined in: [intelligence/improvement-cycle.ts:179](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L179)
 
 ***
 
 ### UsageSplit
 
-Defined in: [intelligence/index.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L157)
+Defined in: [intelligence/index.ts:165](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L165)
 
 The per-class cost split carried by every trace and outcome. `off` ⇒
 `intelligenceUsd: 0` by construction — there is no intelligence spawn to
@@ -1390,7 +1778,7 @@ bill. This is a classification on the trace, NOT a budget-pool split.
 
 > **inferenceUsd**: `number`
 
-Defined in: [intelligence/index.ts:159](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L159)
+Defined in: [intelligence/index.ts:167](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L167)
 
 Base-stream (model) spend in USD.
 
@@ -1398,7 +1786,7 @@ Base-stream (model) spend in USD.
 
 > **intelligenceUsd**: `number`
 
-Defined in: [intelligence/index.ts:161](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L161)
+Defined in: [intelligence/index.ts:169](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L169)
 
 Intelligence-spawn spend in USD. Provably `0` at the OFF tier.
 
@@ -1406,7 +1794,7 @@ Intelligence-spawn spend in USD. Provably `0` at the OFF tier.
 
 ### RunRecord
 
-Defined in: [intelligence/index.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L171)
+Defined in: [intelligence/index.ts:179](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L179)
 
 The typed record `withIntelligence` sends per call — serialized through the
 shipped OTLP builders to the plane's `/v1/otlp` ingest. `input`/`output` are
@@ -1420,43 +1808,43 @@ tree under the same `traceId`.
 
 > **runId**: `string`
 
-Defined in: [intelligence/index.ts:172](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L172)
+Defined in: [intelligence/index.ts:180](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L180)
 
 ##### traceId
 
 > **traceId**: `string`
 
-Defined in: [intelligence/index.ts:173](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L173)
+Defined in: [intelligence/index.ts:181](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L181)
 
 ##### project
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:174](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L174)
+Defined in: [intelligence/index.ts:182](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L182)
 
 ##### target
 
 > **target**: `string`
 
-Defined in: [intelligence/index.ts:175](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L175)
+Defined in: [intelligence/index.ts:183](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L183)
 
 ##### input
 
 > **input**: `unknown`
 
-Defined in: [intelligence/index.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L176)
+Defined in: [intelligence/index.ts:184](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L184)
 
 ##### output
 
 > **output**: `unknown`
 
-Defined in: [intelligence/index.ts:177](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L177)
+Defined in: [intelligence/index.ts:185](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L185)
 
 ##### outcome
 
 > **outcome**: `object`
 
-Defined in: [intelligence/index.ts:178](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L178)
+Defined in: [intelligence/index.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L186)
 
 ###### success?
 
@@ -1474,61 +1862,61 @@ Defined in: [intelligence/index.ts:178](https://github.com/tangle-network/agent-
 
 > `optional` **model?**: `string`
 
-Defined in: [intelligence/index.ts:183](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L183)
+Defined in: [intelligence/index.ts:191](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L191)
 
 ##### provider?
 
 > `optional` **provider?**: `string`
 
-Defined in: [intelligence/index.ts:184](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L184)
+Defined in: [intelligence/index.ts:192](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L192)
 
 ##### loopEvents?
 
 > `optional` **loopEvents?**: [`LoopTraceEvent`](runtime.md#looptraceevent)[]
 
-Defined in: [intelligence/index.ts:185](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L185)
+Defined in: [intelligence/index.ts:193](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L193)
 
 ##### runtimeEvents?
 
 > `optional` **runtimeEvents?**: [`RuntimeStreamEvent`](index.md#runtimestreamevent)[]
 
-Defined in: [intelligence/index.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L186)
+Defined in: [intelligence/index.ts:194](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L194)
 
 ##### profile?
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [intelligence/index.ts:187](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L187)
+Defined in: [intelligence/index.ts:195](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L195)
 
 ##### sessionId?
 
 > `optional` **sessionId?**: `string`
 
-Defined in: [intelligence/index.ts:188](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L188)
+Defined in: [intelligence/index.ts:196](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L196)
 
 ##### harness?
 
 > `optional` **harness?**: `string`
 
-Defined in: [intelligence/index.ts:189](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L189)
+Defined in: [intelligence/index.ts:197](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L197)
 
 ##### repository?
 
 > `optional` **repository?**: `string`
 
-Defined in: [intelligence/index.ts:190](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L190)
+Defined in: [intelligence/index.ts:198](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L198)
 
 ##### commitSha?
 
 > `optional` **commitSha?**: `string`
 
-Defined in: [intelligence/index.ts:191](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L191)
+Defined in: [intelligence/index.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L199)
 
 ##### timing?
 
 > `optional` **timing?**: `object`
 
-Defined in: [intelligence/index.ts:192](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L192)
+Defined in: [intelligence/index.ts:200](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L200)
 
 ###### startedAt
 
@@ -1546,7 +1934,7 @@ Defined in: [intelligence/index.ts:192](https://github.com/tangle-network/agent-
 
 > `optional` **tokens?**: `object`
 
-Defined in: [intelligence/index.ts:193](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L193)
+Defined in: [intelligence/index.ts:201](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L201)
 
 ###### input
 
@@ -1568,7 +1956,7 @@ Defined in: [intelligence/index.ts:193](https://github.com/tangle-network/agent-
 
 > `optional` **error?**: `object`
 
-Defined in: [intelligence/index.ts:199](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L199)
+Defined in: [intelligence/index.ts:207](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L207)
 
 ###### name
 
@@ -1586,7 +1974,7 @@ Defined in: [intelligence/index.ts:199](https://github.com/tangle-network/agent-
 
 > `optional` **candidateExecution?**: `CandidateExecutionEvidence`
 
-Defined in: [intelligence/index.ts:201](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L201)
+Defined in: [intelligence/index.ts:209](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L209)
 
 Exact proposal → review → execution → receipt linkage for candidate runs.
 
@@ -1594,7 +1982,7 @@ Exact proposal → review → execution → receipt linkage for candidate runs.
 
 ### RunReport
 
-Defined in: [intelligence/index.ts:210](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L210)
+Defined in: [intelligence/index.ts:218](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L218)
 
 What an agent reports (via `applied.record`) to enrich the [RunRecord](#runrecord)
 sent for its call. All optional — an un-recorded run still sends input/output
@@ -1607,79 +1995,79 @@ as pure inference (the base stream).
 
 > `optional` **success?**: `boolean`
 
-Defined in: [intelligence/index.ts:211](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L211)
+Defined in: [intelligence/index.ts:219](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L219)
 
 ##### score?
 
 > `optional` **score?**: `number`
 
-Defined in: [intelligence/index.ts:212](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L212)
+Defined in: [intelligence/index.ts:220](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L220)
 
 ##### usage?
 
 > `optional` **usage?**: `Partial`\<[`UsageSplit`](#usagesplit)\>
 
-Defined in: [intelligence/index.ts:213](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L213)
+Defined in: [intelligence/index.ts:221](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L221)
 
 ##### costUsd?
 
 > `optional` **costUsd?**: `number`
 
-Defined in: [intelligence/index.ts:214](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L214)
+Defined in: [intelligence/index.ts:222](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L222)
 
 ##### model?
 
 > `optional` **model?**: `string`
 
-Defined in: [intelligence/index.ts:215](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L215)
+Defined in: [intelligence/index.ts:223](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L223)
 
 ##### provider?
 
 > `optional` **provider?**: `string`
 
-Defined in: [intelligence/index.ts:216](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L216)
+Defined in: [intelligence/index.ts:224](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L224)
 
 ##### loopEvents?
 
 > `optional` **loopEvents?**: [`LoopTraceEvent`](runtime.md#looptraceevent)[]
 
-Defined in: [intelligence/index.ts:217](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L217)
+Defined in: [intelligence/index.ts:225](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L225)
 
 ##### runtimeEvents?
 
 > `optional` **runtimeEvents?**: [`RuntimeStreamEvent`](index.md#runtimestreamevent)[]
 
-Defined in: [intelligence/index.ts:218](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L218)
+Defined in: [intelligence/index.ts:226](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L226)
 
 ##### profile?
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [intelligence/index.ts:219](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L219)
+Defined in: [intelligence/index.ts:227](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L227)
 
 ##### sessionId?
 
 > `optional` **sessionId?**: `string`
 
-Defined in: [intelligence/index.ts:220](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L220)
+Defined in: [intelligence/index.ts:228](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L228)
 
 ##### harness?
 
 > `optional` **harness?**: `string`
 
-Defined in: [intelligence/index.ts:221](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L221)
+Defined in: [intelligence/index.ts:229](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L229)
 
 ##### commitSha?
 
 > `optional` **commitSha?**: `string`
 
-Defined in: [intelligence/index.ts:222](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L222)
+Defined in: [intelligence/index.ts:230](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L230)
 
 ##### tokens?
 
 > `optional` **tokens?**: `object`
 
-Defined in: [intelligence/index.ts:223](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L223)
+Defined in: [intelligence/index.ts:231](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L231)
 
 ###### input
 
@@ -1701,7 +2089,7 @@ Defined in: [intelligence/index.ts:223](https://github.com/tangle-network/agent-
 
 > `optional` **error?**: `object`
 
-Defined in: [intelligence/index.ts:224](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L224)
+Defined in: [intelligence/index.ts:232](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L232)
 
 ###### name
 
@@ -1719,13 +2107,13 @@ Defined in: [intelligence/index.ts:224](https://github.com/tangle-network/agent-
 
 > `optional` **candidateExecution?**: `CandidateExecutionEvidence`
 
-Defined in: [intelligence/index.ts:225](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L225)
+Defined in: [intelligence/index.ts:233](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L233)
 
 ***
 
 ### RepoConfig
 
-Defined in: [intelligence/index.ts:231](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L231)
+Defined in: [intelligence/index.ts:239](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L239)
 
 Repo coordinates a product may declare for the (later) Gated-PR mode. The
  Observe slice only records their PRESENCE for `doctor()`; it never touches
@@ -1737,25 +2125,25 @@ Repo coordinates a product may declare for the (later) Gated-PR mode. The
 
 > **owner**: `string`
 
-Defined in: [intelligence/index.ts:232](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L232)
+Defined in: [intelligence/index.ts:240](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L240)
 
 ##### name
 
 > **name**: `string`
 
-Defined in: [intelligence/index.ts:233](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L233)
+Defined in: [intelligence/index.ts:241](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L241)
 
 ##### baseBranch
 
 > **baseBranch**: `string`
 
-Defined in: [intelligence/index.ts:234](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L234)
+Defined in: [intelligence/index.ts:242](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L242)
 
 ***
 
 ### IntelligenceConfig
 
-Defined in: [intelligence/index.ts:240](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L240)
+Defined in: [intelligence/index.ts:248](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L248)
 
 Client configuration. `project` + `apiKey` are the Observe minimum; the
  rest tune effort, endpoint, redaction, and (for `doctor()` readiness)
@@ -1771,7 +2159,7 @@ Client configuration. `project` + `apiKey` are the Observe minimum; the
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:242](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L242)
+Defined in: [intelligence/index.ts:250](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L250)
 
 Stable project id — the tenant dimension every trace is tagged with.
 
@@ -1779,7 +2167,7 @@ Stable project id — the tenant dimension every trace is tagged with.
 
 > `optional` **apiKey?**: `string`
 
-Defined in: [intelligence/index.ts:244](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L244)
+Defined in: [intelligence/index.ts:252](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L252)
 
 Bearer key for the Intelligence ingest. Reads `TANGLE_API_KEY` when omitted.
 
@@ -1787,7 +2175,7 @@ Bearer key for the Intelligence ingest. Reads `TANGLE_API_KEY` when omitted.
 
 > `optional` **effort?**: [`EffortTier`](#efforttier) \| \{ `tier`: [`EffortTier`](#efforttier); `overrides?`: `Partial`\<[`EffortSettings`](#effortsettings)\>; \}
 
-Defined in: [intelligence/index.ts:246](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L246)
+Defined in: [intelligence/index.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L254)
 
 Effort tier (default `'standard'`) plus optional per-field overrides.
 
@@ -1795,7 +2183,7 @@ Effort tier (default `'standard'`) plus optional per-field overrides.
 
 > `optional` **baseUrl?**: `string`
 
-Defined in: [intelligence/index.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L254)
+Defined in: [intelligence/index.ts:262](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L262)
 
 The ONE Tangle Intelligence base URL — both the send (OTLP `/v1/otlp`) and
 receive (`/v1/profiles/:target/composed`) paths derive from it. Reads
@@ -1807,7 +2195,7 @@ key the ingest requires); absent a key, export is a no-op.
 
 > `optional` **redact?**: `false` \| [`Redactor`](#redactor)
 
-Defined in: [intelligence/index.ts:260](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L260)
+Defined in: [intelligence/index.ts:268](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L268)
 
 Redaction hook run over every exported input/output. A function replaces
 the default scrubber; `false` opts out entirely (raw fidelity, caller has
@@ -1817,7 +2205,7 @@ sanitized upstream); omitted ⇒ the built-in `defaultRedactor`.
 
 > `optional` **surfaces?**: `string`[]
 
-Defined in: [intelligence/index.ts:262](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L262)
+Defined in: [intelligence/index.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L270)
 
 Mutable surfaces a later PR mode would edit. Recorded for `doctor()` only.
 
@@ -1825,7 +2213,7 @@ Mutable surfaces a later PR mode would edit. Recorded for `doctor()` only.
 
 > `optional` **checks?**: `string`[]
 
-Defined in: [intelligence/index.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L264)
+Defined in: [intelligence/index.ts:272](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L272)
 
 Verification checks a later PR mode would gate on. Recorded for `doctor()` only.
 
@@ -1833,7 +2221,7 @@ Verification checks a later PR mode would gate on. Recorded for `doctor()` only.
 
 > `optional` **repo?**: [`RepoConfig`](#repoconfig)
 
-Defined in: [intelligence/index.ts:266](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L266)
+Defined in: [intelligence/index.ts:274](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L274)
 
 Repo access a later PR mode would need. Recorded for `doctor()` only.
 
@@ -1841,7 +2229,7 @@ Repo access a later PR mode would need. Recorded for `doctor()` only.
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [intelligence/index.ts:268](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L268)
+Defined in: [intelligence/index.ts:276](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L276)
 
 Full canonical profile used for this agent. Exported redacted with a stable hash.
 
@@ -1849,7 +2237,7 @@ Full canonical profile used for this agent. Exported redacted with a stable hash
 
 > `optional` **commitSha?**: `string`
 
-Defined in: [intelligence/index.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L270)
+Defined in: [intelligence/index.ts:278](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L278)
 
 Commit that produced the running agent, when known.
 
@@ -1857,7 +2245,7 @@ Commit that produced the running agent, when known.
 
 > `optional` **runtimeTelemetry?**: [`RuntimeTelemetryOptions`](index.md#runtimetelemetryoptions)
 
-Defined in: [intelligence/index.ts:272](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L272)
+Defined in: [intelligence/index.ts:280](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L280)
 
 Runtime-event payload policy. Tool inputs/results remain off unless explicitly enabled.
 
@@ -1865,7 +2253,7 @@ Runtime-event payload policy. Tool inputs/results remain off unless explicitly e
 
 > `optional` **payloadAttributes?**: `"metadata"` \| `"full"`
 
-Defined in: [intelligence/index.ts:279](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L279)
+Defined in: [intelligence/index.ts:287](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L287)
 
 Payloads are metadata-only by default: the run span carries a stable hash
 and UTF-8 byte count, but not the redacted content. Set `full` only when
@@ -1876,7 +2264,7 @@ inputs, outputs, and profiles.
 
 ### TraceMeta
 
-Defined in: [intelligence/index.ts:283](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L283)
+Defined in: [intelligence/index.ts:291](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L291)
 
 Metadata describing one traced run. `runId`/`traceId` default to fresh ids.
 
@@ -1886,7 +2274,7 @@ Metadata describing one traced run. `runId`/`traceId` default to fresh ids.
 
 > `optional` **input?**: `unknown`
 
-Defined in: [intelligence/index.ts:285](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L285)
+Defined in: [intelligence/index.ts:293](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L293)
 
 The run's input — exported through the redactor.
 
@@ -1894,7 +2282,7 @@ The run's input — exported through the redactor.
 
 > `optional` **runId?**: `string`
 
-Defined in: [intelligence/index.ts:287](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L287)
+Defined in: [intelligence/index.ts:295](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L295)
 
 Stable run id. Defaults to a fresh id.
 
@@ -1902,7 +2290,7 @@ Stable run id. Defaults to a fresh id.
 
 > `optional` **traceId?**: `string`
 
-Defined in: [intelligence/index.ts:289](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L289)
+Defined in: [intelligence/index.ts:297](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L297)
 
 32-hex trace id. Defaults to a fresh id.
 
@@ -1910,7 +2298,7 @@ Defined in: [intelligence/index.ts:289](https://github.com/tangle-network/agent-
 
 > `optional` **model?**: `string`
 
-Defined in: [intelligence/index.ts:291](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L291)
+Defined in: [intelligence/index.ts:299](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L299)
 
 Model id, when known — stamped on the span.
 
@@ -1918,7 +2306,7 @@ Model id, when known — stamped on the span.
 
 > `optional` **provider?**: `string`
 
-Defined in: [intelligence/index.ts:293](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L293)
+Defined in: [intelligence/index.ts:301](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L301)
 
 Provider name, when known — stamped on the span.
 
@@ -1926,7 +2314,7 @@ Provider name, when known — stamped on the span.
 
 > `optional` **labels?**: `Record`\<`string`, `string` \| `number` \| `boolean`\>
 
-Defined in: [intelligence/index.ts:295](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L295)
+Defined in: [intelligence/index.ts:303](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L303)
 
 Arbitrary extra labels (string/number/boolean) stamped on the span.
 
@@ -1934,7 +2322,7 @@ Arbitrary extra labels (string/number/boolean) stamped on the span.
 
 ### TraceHandle
 
-Defined in: [intelligence/index.ts:304](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L304)
+Defined in: [intelligence/index.ts:312](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L312)
 
 The trace handle a `traceRun` body records into. `recordOutput` captures the
 agent's result (redacted on export); `recordOutcome` captures the scored
@@ -1947,7 +2335,7 @@ an un-recorded run still exports a span with whatever was set.
 
 > **recordOutput**(`output`): `void`
 
-Defined in: [intelligence/index.ts:306](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L306)
+Defined in: [intelligence/index.ts:314](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L314)
 
 Capture the run's output. Exported through the redactor.
 
@@ -1965,7 +2353,7 @@ Capture the run's output. Exported through the redactor.
 
 > **recordOutcome**(`outcome`): `void`
 
-Defined in: [intelligence/index.ts:313](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L313)
+Defined in: [intelligence/index.ts:321](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L321)
 
 Capture the run's outcome. `usage` defaults to inference-only
 (`intelligenceUsd: 0`) — the OFF baseline; an intelligence-enabled run
@@ -2000,7 +2388,7 @@ treated as pure inference.
 
 ### RecordTraceMeta
 
-Defined in: [intelligence/index.ts:322](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L322)
+Defined in: [intelligence/index.ts:330](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L330)
 
 Metadata for [IntelligenceClient.recordTrace](#recordtrace).
 
@@ -2010,7 +2398,7 @@ Metadata for [IntelligenceClient.recordTrace](#recordtrace).
 
 > `optional` **traceId?**: `string`
 
-Defined in: [intelligence/index.ts:324](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L324)
+Defined in: [intelligence/index.ts:332](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L332)
 
 32-hex trace id to anchor every span to. Defaults to a fresh id.
 
@@ -2018,7 +2406,7 @@ Defined in: [intelligence/index.ts:324](https://github.com/tangle-network/agent-
 
 > `optional` **rootParentSpanId?**: `string`
 
-Defined in: [intelligence/index.ts:327](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L327)
+Defined in: [intelligence/index.ts:335](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L335)
 
 Span id of an enclosing span the loop root should parent under (e.g. a
  `traceRun` span). Omitted ⇒ the loop root is the trace root.
@@ -2027,7 +2415,7 @@ Span id of an enclosing span the loop root should parent under (e.g. a
 
 ### TraceOutcome
 
-Defined in: [intelligence/index.ts:332](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L332)
+Defined in: [intelligence/index.ts:340](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L340)
 
 The resolved outcome of one traced run, surfaced on the export span and
  available to the caller for downstream billing assertions.
@@ -2038,25 +2426,25 @@ The resolved outcome of one traced run, surfaced on the export span and
 
 > **runId**: `string`
 
-Defined in: [intelligence/index.ts:333](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L333)
+Defined in: [intelligence/index.ts:341](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L341)
 
 ##### traceId
 
 > **traceId**: `string`
 
-Defined in: [intelligence/index.ts:334](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L334)
+Defined in: [intelligence/index.ts:342](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L342)
 
 ##### project
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:335](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L335)
+Defined in: [intelligence/index.ts:343](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L343)
 
 ##### effort
 
 > **effort**: [`EffortSettings`](#effortsettings)
 
-Defined in: [intelligence/index.ts:337](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L337)
+Defined in: [intelligence/index.ts:345](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L345)
 
 The resolved effort settings this run executed under.
 
@@ -2064,7 +2452,7 @@ The resolved effort settings this run executed under.
 
 > **intelligenceOff**: `boolean`
 
-Defined in: [intelligence/index.ts:339](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L339)
+Defined in: [intelligence/index.ts:347](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L347)
 
 True when this run ran as pure passthrough (the OFF floor).
 
@@ -2072,19 +2460,19 @@ True when this run ran as pure passthrough (the OFF floor).
 
 > `optional` **success?**: `boolean`
 
-Defined in: [intelligence/index.ts:340](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L340)
+Defined in: [intelligence/index.ts:348](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L348)
 
 ##### score?
 
 > `optional` **score?**: `number`
 
-Defined in: [intelligence/index.ts:341](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L341)
+Defined in: [intelligence/index.ts:349](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L349)
 
 ##### usage
 
 > **usage**: [`UsageSplit`](#usagesplit)
 
-Defined in: [intelligence/index.ts:343](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L343)
+Defined in: [intelligence/index.ts:351](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L351)
 
 Per-class billing split. `intelligenceUsd` is `0` at the OFF tier.
 
@@ -2092,7 +2480,7 @@ Per-class billing split. `intelligenceUsd` is `0` at the OFF tier.
 
 ### IntelligenceClient
 
-Defined in: [intelligence/index.ts:347](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L347)
+Defined in: [intelligence/index.ts:355](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L355)
 
 The Observe-mode Intelligence client.
 
@@ -2102,7 +2490,7 @@ The Observe-mode Intelligence client.
 
 > `readonly` **project**: `string`
 
-Defined in: [intelligence/index.ts:349](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L349)
+Defined in: [intelligence/index.ts:357](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L357)
 
 The resolved project id.
 
@@ -2110,7 +2498,7 @@ The resolved project id.
 
 > `readonly` **effort**: [`EffortSettings`](#effortsettings)
 
-Defined in: [intelligence/index.ts:351](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L351)
+Defined in: [intelligence/index.ts:359](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L359)
 
 The resolved effort settings.
 
@@ -2120,7 +2508,7 @@ The resolved effort settings.
 
 > **traceRun**\<`T`\>(`meta`, `fn`): `Promise`\<`T`\>
 
-Defined in: [intelligence/index.ts:357](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L357)
+Defined in: [intelligence/index.ts:365](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L365)
 
 Run `fn` under a trace, export one span best-effort, and return whatever
 `fn` returns. Telemetry-export failures are swallowed; an error THROWN by
@@ -2150,7 +2538,7 @@ Run `fn` under a trace, export one span best-effort, and return whatever
 
 > **recordTrace**(`events`, `meta?`): `string`
 
-Defined in: [intelligence/index.ts:367](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L367)
+Defined in: [intelligence/index.ts:375](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L375)
 
 Export a run's full loop topology — the ordered `LoopTraceEvent` stream a
 `runLoop`/`Supervisor` run emits — as a nested OTLP span tree (loop → round →
@@ -2178,7 +2566,7 @@ readonly [`LoopTraceEvent`](runtime.md#looptraceevent)[]
 
 > **exportRunRecord**(`record`): `string`
 
-Defined in: [intelligence/index.ts:375](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L375)
+Defined in: [intelligence/index.ts:383](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L383)
 
 Send one typed [RunRecord](#runrecord) — the run's flat span (input/output/outcome/
 usage/model/provider, redacted) plus, when `loopEvents` are present, the
@@ -2200,7 +2588,7 @@ Best-effort: export failures are swallowed. Returns the record's `traceId`.
 
 > **freshRunId**(): `string`
 
-Defined in: [intelligence/index.ts:377](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L377)
+Defined in: [intelligence/index.ts:385](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L385)
 
 Mint a fresh run id (`run-<hex>`).
 
@@ -2212,7 +2600,7 @@ Mint a fresh run id (`run-<hex>`).
 
 > **freshTraceId**(): `string`
 
-Defined in: [intelligence/index.ts:379](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L379)
+Defined in: [intelligence/index.ts:387](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L387)
 
 Mint a fresh 32-hex trace id.
 
@@ -2224,7 +2612,7 @@ Mint a fresh 32-hex trace id.
 
 > **doctor**(): [`DoctorReport`](#doctorreport)
 
-Defined in: [intelligence/index.ts:385](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L385)
+Defined in: [intelligence/index.ts:393](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L393)
 
 Network-free readiness report: which adoption modes are reachable given
 this config. Observe is always reachable; Recommend needs outcomes; PR
@@ -2238,7 +2626,7 @@ needs checks + surfaces + repo.
 
 > **flush**(): `Promise`\<`void`\>
 
-Defined in: [intelligence/index.ts:387](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L387)
+Defined in: [intelligence/index.ts:395](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L395)
 
 Flush any pending export spans. Best-effort; resolves even if export fails.
 
@@ -2250,7 +2638,7 @@ Flush any pending export spans. Best-effort; resolves even if export fails.
 
 ### ModeReadiness
 
-Defined in: [intelligence/index.ts:391](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L391)
+Defined in: [intelligence/index.ts:399](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L399)
 
 One mode's readiness verdict.
 
@@ -2260,13 +2648,13 @@ One mode's readiness verdict.
 
 > **ready**: `boolean`
 
-Defined in: [intelligence/index.ts:392](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L392)
+Defined in: [intelligence/index.ts:400](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L400)
 
 ##### missing
 
 > **missing**: `string`[]
 
-Defined in: [intelligence/index.ts:394](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L394)
+Defined in: [intelligence/index.ts:402](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L402)
 
 Inputs this mode still needs, when not ready. Empty when ready.
 
@@ -2274,7 +2662,7 @@ Inputs this mode still needs, when not ready. Empty when ready.
 
 ### DoctorReport
 
-Defined in: [intelligence/index.ts:398](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L398)
+Defined in: [intelligence/index.ts:406](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L406)
 
 The `doctor()` readiness report — Mode-readiness without any network call.
 
@@ -2284,19 +2672,19 @@ The `doctor()` readiness report — Mode-readiness without any network call.
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:399](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L399)
+Defined in: [intelligence/index.ts:407](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L407)
 
 ##### effort
 
 > **effort**: [`EffortSettings`](#effortsettings)
 
-Defined in: [intelligence/index.ts:400](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L400)
+Defined in: [intelligence/index.ts:408](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L408)
 
 ##### exportConfigured
 
 > **exportConfigured**: `boolean`
 
-Defined in: [intelligence/index.ts:402](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L402)
+Defined in: [intelligence/index.ts:410](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L410)
 
 True when an OTLP endpoint is configured (export will actually ship).
 
@@ -2304,7 +2692,7 @@ True when an OTLP endpoint is configured (export will actually ship).
 
 > **modes**: `object`
 
-Defined in: [intelligence/index.ts:403](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L403)
+Defined in: [intelligence/index.ts:411](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L411)
 
 ###### observe
 
@@ -2528,9 +2916,9 @@ carrying an un-admitted binding kind is a hard error, not a soft drop).
 
 ***
 
-### CreateSandboxApprovedCandidateExecutorOptions
+### CreateSandboxCandidateExperimentExecutorOptions
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:65](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L65)
+Defined in: [intelligence/sandbox-approved-candidate.ts:77](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L77)
 
 #### Properties
 
@@ -2538,63 +2926,43 @@ Defined in: [intelligence/sandbox-approved-candidate.ts:65](https://github.com/t
 
 > **client**: `SandboxClientPort`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:66](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L66)
+Defined in: [intelligence/sandbox-approved-candidate.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L78)
 
 ##### ports
 
 > **ports**: [`AgentCandidateExecutionPorts`](index.md#agentcandidateexecutionports)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:67](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L67)
+Defined in: [intelligence/sandbox-approved-candidate.ts:79](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L79)
 
 ##### grader
 
 > **grader**: [`AgentCandidateBenchmarkGraderPort`](index.md#agentcandidatebenchmarkgraderport)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:68](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L68)
+Defined in: [intelligence/sandbox-approved-candidate.ts:80](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L80)
 
 ##### outputArtifacts
 
 > **outputArtifacts**: [`AgentCandidateOutputArtifactPort`](index.md#agentcandidateoutputartifactport)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:69](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L69)
+Defined in: [intelligence/sandbox-approved-candidate.ts:81](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L81)
 
 ##### traceStore
 
 > **traceStore**: `TraceStore`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:70](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L70)
+Defined in: [intelligence/sandbox-approved-candidate.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L82)
 
 ##### claimStore
 
 > **claimStore**: [`AgentCandidateExecutionClaimStore`](index.md#agentcandidateexecutionclaimstore)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:71](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L71)
-
-##### authorizeReview
-
-> **authorizeReview**: (`review`, `proposal`) => `boolean` \| `Promise`\<`boolean`\>
-
-Defined in: [intelligence/sandbox-approved-candidate.ts:72](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L72)
-
-###### Parameters
-
-###### review
-
-`AgentImprovementReview`
-
-###### proposal
-
-`AgentImprovementProposal`
-
-###### Returns
-
-`boolean` \| `Promise`\<`boolean`\>
+Defined in: [intelligence/sandbox-approved-candidate.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L83)
 
 ##### sandbox?
 
 > `optional` **sandbox?**: `object`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:76](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L76)
+Defined in: [intelligence/sandbox-approved-candidate.ts:84](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L84)
 
 ###### teamId?
 
@@ -2616,51 +2984,81 @@ Defined in: [intelligence/sandbox-approved-candidate.ts:76](https://github.com/t
 
 > `optional` **cleanupTimeoutMs?**: `number`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:82](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L82)
+Defined in: [intelligence/sandbox-approved-candidate.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L90)
 
 ##### resultTimeoutMs?
 
 > `optional` **resultTimeoutMs?**: `number`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:83](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L83)
+Defined in: [intelligence/sandbox-approved-candidate.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L91)
 
 ***
 
-### SandboxApprovedCandidateExecution
+### SandboxCandidateExperimentExecution
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:94](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L94)
+
+#### Extends
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:86](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L86)
+- `CandidateExperimentExecutionInput`
 
 #### Properties
 
-##### proposal
+##### executionId
 
-> **proposal**: `AgentImprovementProposal`
+> **executionId**: `string`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L95)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L87)
+##### attempt?
 
-##### review
+> `optional` **attempt?**: `number`
 
-> **review**: `AgentImprovementReview`
+Defined in: [intelligence/sandbox-approved-candidate.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L96)
+
+##### executionRoots
+
+> **executionRoots**: `object`
+
+Defined in: [intelligence/sandbox-approved-candidate.ts:97](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L97)
+
+###### taskRoot
+
+> **taskRoot**: `string`
+
+###### candidateRoot?
+
+> `optional` **candidateRoot?**: `string`
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:88](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L88)
+##### stagingRoots
 
-##### task
+> **stagingRoots**: `object`
 
-> **task**: [`AgentCandidateTaskExecution`](index.md#agentcandidatetaskexecution)
+Defined in: [intelligence/sandbox-approved-candidate.ts:98](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L98)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:89](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L89)
+###### taskRoot
+
+> **taskRoot**: `string`
+
+###### candidateRoot?
+
+> `optional` **candidateRoot?**: `string`
+
+###### profileRoot
+
+> **profileRoot**: `string`
 
 ##### preparation?
 
 > `optional` **preparation?**: [`PrepareAgentCandidateExecutionOptions`](index.md#prepareagentcandidateexecutionoptions)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:90](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L90)
+Defined in: [intelligence/sandbox-approved-candidate.ts:99](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L99)
 
 ***
 
-### SandboxApprovedCandidateExecutor
+### SandboxCandidateExperimentExecutor
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:93](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L93)
+Defined in: [intelligence/sandbox-approved-candidate.ts:102](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L102)
 
 #### Properties
 
@@ -2668,7 +3066,7 @@ Defined in: [intelligence/sandbox-approved-candidate.ts:93](https://github.com/t
 
 > `readonly` **executor**: [`AgentCandidateExecutorPort`](index.md#agentcandidateexecutorport)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:95](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L95)
+Defined in: [intelligence/sandbox-approved-candidate.ts:104](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L104)
 
 The same port is usable by Runtime's expired-claim recovery path.
 
@@ -2678,13 +3076,13 @@ The same port is usable by Runtime's expired-claim recovery path.
 
 > **execute**(`input`): `Promise`\<`CandidateExecutionEvidence`\>
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:96](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L96)
+Defined in: [intelligence/sandbox-approved-candidate.ts:105](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L105)
 
 ###### Parameters
 
 ###### input
 
-[`SandboxApprovedCandidateExecution`](#sandboxapprovedcandidateexecution)
+[`SandboxCandidateExperimentExecution`](#sandboxcandidateexperimentexecution)
 
 ###### Returns
 
@@ -2816,7 +3214,7 @@ Defined in: [intelligence/with-intelligence.ts:83](https://github.com/tangle-net
 
 > **project**: `string`
 
-Defined in: [intelligence/index.ts:242](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L242)
+Defined in: [intelligence/index.ts:250](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L250)
 
 Stable project id — the tenant dimension every trace is tagged with.
 
@@ -2828,7 +3226,7 @@ Stable project id — the tenant dimension every trace is tagged with.
 
 > `optional` **apiKey?**: `string`
 
-Defined in: [intelligence/index.ts:244](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L244)
+Defined in: [intelligence/index.ts:252](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L252)
 
 Bearer key for the Intelligence ingest. Reads `TANGLE_API_KEY` when omitted.
 
@@ -2840,7 +3238,7 @@ Bearer key for the Intelligence ingest. Reads `TANGLE_API_KEY` when omitted.
 
 > `optional` **effort?**: [`EffortTier`](#efforttier) \| \{ `tier`: [`EffortTier`](#efforttier); `overrides?`: `Partial`\<[`EffortSettings`](#effortsettings)\>; \}
 
-Defined in: [intelligence/index.ts:246](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L246)
+Defined in: [intelligence/index.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L254)
 
 Effort tier (default `'standard'`) plus optional per-field overrides.
 
@@ -2852,7 +3250,7 @@ Effort tier (default `'standard'`) plus optional per-field overrides.
 
 > `optional` **baseUrl?**: `string`
 
-Defined in: [intelligence/index.ts:254](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L254)
+Defined in: [intelligence/index.ts:262](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L262)
 
 The ONE Tangle Intelligence base URL — both the send (OTLP `/v1/otlp`) and
 receive (`/v1/profiles/:target/composed`) paths derive from it. Reads
@@ -2868,7 +3266,7 @@ key the ingest requires); absent a key, export is a no-op.
 
 > `optional` **redact?**: `false` \| [`Redactor`](#redactor)
 
-Defined in: [intelligence/index.ts:260](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L260)
+Defined in: [intelligence/index.ts:268](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L268)
 
 Redaction hook run over every exported input/output. A function replaces
 the default scrubber; `false` opts out entirely (raw fidelity, caller has
@@ -2882,7 +3280,7 @@ sanitized upstream); omitted ⇒ the built-in `defaultRedactor`.
 
 > `optional` **surfaces?**: `string`[]
 
-Defined in: [intelligence/index.ts:262](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L262)
+Defined in: [intelligence/index.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L270)
 
 Mutable surfaces a later PR mode would edit. Recorded for `doctor()` only.
 
@@ -2894,7 +3292,7 @@ Mutable surfaces a later PR mode would edit. Recorded for `doctor()` only.
 
 > `optional` **checks?**: `string`[]
 
-Defined in: [intelligence/index.ts:264](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L264)
+Defined in: [intelligence/index.ts:272](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L272)
 
 Verification checks a later PR mode would gate on. Recorded for `doctor()` only.
 
@@ -2906,7 +3304,7 @@ Verification checks a later PR mode would gate on. Recorded for `doctor()` only.
 
 > `optional` **repo?**: [`RepoConfig`](#repoconfig)
 
-Defined in: [intelligence/index.ts:266](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L266)
+Defined in: [intelligence/index.ts:274](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L274)
 
 Repo access a later PR mode would need. Recorded for `doctor()` only.
 
@@ -2918,7 +3316,7 @@ Repo access a later PR mode would need. Recorded for `doctor()` only.
 
 > `optional` **profile?**: `AgentProfile`
 
-Defined in: [intelligence/index.ts:268](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L268)
+Defined in: [intelligence/index.ts:276](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L276)
 
 Full canonical profile used for this agent. Exported redacted with a stable hash.
 
@@ -2930,7 +3328,7 @@ Full canonical profile used for this agent. Exported redacted with a stable hash
 
 > `optional` **commitSha?**: `string`
 
-Defined in: [intelligence/index.ts:270](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L270)
+Defined in: [intelligence/index.ts:278](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L278)
 
 Commit that produced the running agent, when known.
 
@@ -2942,7 +3340,7 @@ Commit that produced the running agent, when known.
 
 > `optional` **runtimeTelemetry?**: [`RuntimeTelemetryOptions`](index.md#runtimetelemetryoptions)
 
-Defined in: [intelligence/index.ts:272](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L272)
+Defined in: [intelligence/index.ts:280](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L280)
 
 Runtime-event payload policy. Tool inputs/results remain off unless explicitly enabled.
 
@@ -2954,7 +3352,7 @@ Runtime-event payload policy. Tool inputs/results remain off unless explicitly e
 
 > `optional` **payloadAttributes?**: `"metadata"` \| `"full"`
 
-Defined in: [intelligence/index.ts:279](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L279)
+Defined in: [intelligence/index.ts:287](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L287)
 
 Payloads are metadata-only by default: the run span carries a stable hash
 and UTF-8 byte count, but not the redacted content. Set `full` only when
@@ -3154,19 +3552,11 @@ Per-field overrides applied on top of a tier preset. Any subset of the
 
 ***
 
-### ExecuteApprovedAgentCandidateResult
-
-> **ExecuteApprovedAgentCandidateResult** = \{ `finalization`: `Extract`\<[`AgentCandidateRunFinalization`](index.md#agentcandidaterunfinalization), \{ `succeeded`: `false`; \}\>; `evidence?`: `never`; \} \| \{ `finalization`: `Extract`\<[`AgentCandidateRunFinalization`](index.md#agentcandidaterunfinalization), \{ `succeeded`: `true`; \}\>; `evidence`: `CandidateExecutionEvidence`; \}
-
-Defined in: [intelligence/improvement-cycle.ts:148](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L148)
-
-***
-
 ### UsageClass
 
 > **UsageClass** = `"inference"` \| `"intelligence"`
 
-Defined in: [intelligence/index.ts:150](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L150)
+Defined in: [intelligence/index.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L158)
 
 Usage class for billing. Base-stream tokens bill `'inference'`; every
  intelligence spawn (analyst, corpus, loop) bills `'intelligence'`. The
@@ -3291,11 +3681,11 @@ The default tier when a client declares no effort. `'standard'` turns
 
 ***
 
-### sandboxApprovedCandidateExecutionSupport
+### sandboxCandidateExperimentExecutionSupport
 
-> `const` **sandboxApprovedCandidateExecutionSupport**: `Readonly`\<\{ `outcomes`: readonly \[`"output"`\]; `outputMediaTypes`: readonly \[`"text/*"`, `"application/json"`, `"*+json"`\]; `code`: readonly \[`"disabled"`\]; `memory`: readonly \[`"disabled"`\]; `knowledge`: `true`; `profile`: `Readonly`\<\{ `mcpTransports`: readonly \[`"stdio"`\]; `remoteMcp`: `false`; `tools`: `false`; `permissions`: `false`; `modes`: `false`; `confidential`: `false`; \}\>; `isolation`: `Readonly`\<\{ `freshSandbox`: `true`; `exactProcess`: `true`; `egress`: readonly \[`"blocked"`, `"strict"`\]; \}\>; \}\>
+> `const` **sandboxCandidateExperimentExecutionSupport**: `Readonly`\<\{ `outcomes`: readonly \[`"output"`\]; `outputMediaTypes`: readonly \[`"text/*"`, `"application/json"`, `"*+json"`\]; `code`: readonly \[`"disabled"`\]; `memory`: readonly \[`"disabled"`\]; `knowledge`: `true`; `profile`: `Readonly`\<\{ `mcpTransports`: readonly \[`"stdio"`\]; `remoteMcp`: `false`; `tools`: `false`; `permissions`: `false`; `modes`: `false`; `confidential`: `false`; \}\>; `isolation`: `Readonly`\<\{ `freshSandbox`: `true`; `exactProcess`: `true`; `egress`: readonly \[`"blocked"`, `"strict"`\]; \}\>; \}\>
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:51](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L51)
+Defined in: [intelligence/sandbox-approved-candidate.ts:63](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L63)
 
 Declares the exact candidate surfaces the sandbox executor can run.
 
@@ -3554,13 +3944,73 @@ compile to `withAnalyst: true`, the tier's `fanout`, and `withLoops: true`.
 
 ***
 
+### runAgentCandidateExperiment()
+
+> **runAgentCandidateExperiment**(`options`): `Promise`\<[`RunAgentCandidateExperimentResult`](#runagentcandidateexperimentresult)\>
+
+Defined in: [intelligence/improvement-cycle.ts:187](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L187)
+
+Execute both arms of one immutable experiment and derive its paired result.
+
+#### Parameters
+
+##### options
+
+[`RunAgentCandidateExperimentOptions`](#runagentcandidateexperimentoptions)
+
+#### Returns
+
+`Promise`\<[`RunAgentCandidateExperimentResult`](#runagentcandidateexperimentresult)\>
+
+***
+
+### executeAgentCandidateExperimentCell()
+
+> **executeAgentCandidateExperimentCell**(`options`): `Promise`\<`CandidateExecutionEvidence`\>
+
+Defined in: [intelligence/improvement-cycle.ts:218](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L218)
+
+Execute one exact arm, task, repetition, seed, and attempt through Runtime.
+
+#### Parameters
+
+##### options
+
+[`ExecuteAgentCandidateExperimentCellOptions`](#executeagentcandidateexperimentcelloptions)
+
+#### Returns
+
+`Promise`\<`CandidateExecutionEvidence`\>
+
+***
+
+### createAgentImprovementMeasuredComparison()
+
+> **createAgentImprovementMeasuredComparison**(`options`): `AgentImprovementMeasuredComparison`
+
+Defined in: [intelligence/improvement-cycle.ts:274](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L274)
+
+Delegate all statistics and promotion checks to agent-eval's receipt-based comparison.
+
+#### Parameters
+
+##### options
+
+`CompareCandidateExperimentOptions`
+
+#### Returns
+
+`AgentImprovementMeasuredComparison`
+
+***
+
 ### proposeAgentImprovement()
 
 > **proposeAgentImprovement**\<`TScenario`, `TArtifact`\>(`options`): `Promise`\<[`ProposeAgentImprovementResult`](#proposeagentimprovementresult)\<`TScenario`, `TArtifact`\>\>
 
-Defined in: [intelligence/improvement-cycle.ts:166](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L166)
+Defined in: [intelligence/improvement-cycle.ts:281](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L281)
 
-Analyze one run and freeze its measured winner into one exact proposal.
+Analyze, search, then remeasure the resulting exact candidate before proposing it.
 
 #### Type Parameters
 
@@ -3588,11 +4038,9 @@ Analyze one run and freeze its measured winner into one exact proposal.
 
 > **createAgentImprovementProposal**(`options`): `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:218](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L218)
+Defined in: [intelligence/improvement-cycle.ts:342](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L342)
 
-Freeze an already-measured improvement into the one reviewable proposal
-contract. Products that run analysis or evaluation in separate workers use
-this constructor instead of rerunning either phase or rebuilding digests.
+Create the reviewable record only from a complete, recomputable experiment result.
 
 #### Parameters
 
@@ -3610,9 +4058,9 @@ this constructor instead of rerunning either phase or rebuilding digests.
 
 > **reviewAgentImprovementProposal**(`inputProposal`, `input`): `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:256](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L256)
+Defined in: [intelligence/improvement-cycle.ts:373](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L373)
 
-Persist an approve/reject/change-request decision bound to one exact proposal.
+Persist a human or tenant-policy decision bound to one exact proposal.
 
 #### Parameters
 
@@ -3630,33 +4078,41 @@ Persist an approve/reject/change-request decision bound to one exact proposal.
 
 ***
 
-### executeApprovedAgentCandidate()
+### createAgentImprovementActivation()
 
-> **executeApprovedAgentCandidate**(`options`): `Promise`\<[`ExecuteApprovedAgentCandidateResult`](#executeapprovedagentcandidateresult)\>
+> **createAgentImprovementActivation**(`inputProposal`, `inputReview`, `options`): `AgentImprovementActivation`
 
-Defined in: [intelligence/improvement-cycle.ts:284](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L284)
+Defined in: [intelligence/improvement-cycle.ts:397](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L397)
 
-Verify, materialize, run, grade, and receipt only the exact approved bundle.
+Authorize product-owned writes only after the exact candidate was measured and approved.
 
 #### Parameters
 
+##### inputProposal
+
+`AgentImprovementProposal`
+
+##### inputReview
+
+`AgentImprovementReview`
+
 ##### options
 
-[`ExecuteApprovedAgentCandidateOptions`](#executeapprovedagentcandidateoptions)
+[`CreateAgentImprovementActivationOptions`](#createagentimprovementactivationoptions)
 
 #### Returns
 
-`Promise`\<[`ExecuteApprovedAgentCandidateResult`](#executeapprovedagentcandidateresult)\>
+`AgentImprovementActivation`
 
 ***
 
-### verifyCandidateExecutionEvidence()
+### verifyAgentImprovementProposal()
 
-> **verifyCandidateExecutionEvidence**(`input`, `options`): `CandidateExecutionEvidence`[]
+> **verifyAgentImprovementProposal**(`input`): `AgentImprovementProposal`
 
-Defined in: [intelligence/improvement-cycle.ts:333](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L333)
+Defined in: [intelligence/improvement-cycle.ts:428](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L428)
 
-Verify approval, receipts, uniqueness, and the exact native profile files executed.
+Validate a proposal and recompute every binding to its measured experiment.
 
 #### Parameters
 
@@ -3664,23 +4120,19 @@ Verify approval, receipts, uniqueness, and the exact native profile files execut
 
 `unknown`
 
-##### options
-
-[`VerifyCandidateExecutionEvidenceOptions`](#verifycandidateexecutionevidenceoptions)
-
 #### Returns
 
-`CandidateExecutionEvidence`[]
+`AgentImprovementProposal`
 
 ***
 
-### verifyAgentImprovementProposal()
+### verifyAgentImprovementReview()
 
-> **verifyAgentImprovementProposal**(`input`): `AgentImprovementProposal`
+> **verifyAgentImprovementReview**(`input`): `AgentImprovementReview`
 
-Defined in: [intelligence/improvement-cycle.ts:436](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L436)
+Defined in: [intelligence/improvement-cycle.ts:452](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L452)
 
-Validate a proposal's schema, profile, sealed bundle, and canonical digest.
+Validate the canonical identity and wire shape of an improvement review.
 
 #### Parameters
 
@@ -3690,57 +4142,61 @@ Validate a proposal's schema, profile, sealed bundle, and canonical digest.
 
 #### Returns
 
-`AgentImprovementProposal`
+`AgentImprovementReview`
 
 ***
 
-### verifyAgentImprovementReview()
+### verifyAgentImprovementActivation()
 
-> **verifyAgentImprovementReview**(`input`): `AgentImprovementReview`
+> **verifyAgentImprovementActivation**(`input`): `AgentImprovementActivation`
 
-Defined in: [intelligence/improvement-cycle.ts:670](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L670)
+Defined in: [intelligence/improvement-cycle.ts:460](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L460)
 
-Validate a review's decision fields and canonical digest.
+Validate activation authority against the exact proposal, review, experiment, and base state.
 
 #### Parameters
 
 ##### input
 
-`unknown`
+###### proposal
 
-#### Returns
+`unknown`
 
-`AgentImprovementReview`
+###### review
 
-***
+`unknown`
 
-### createAgentImprovementMeasuredComparison()
+###### activation
 
-> **createAgentImprovementMeasuredComparison**\<`TScenario`, `TArtifact`\>(`options`): `AgentImprovementMeasuredComparison`
+`unknown`
 
-Defined in: [intelligence/improvement-cycle.ts:676](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L676)
+#### Returns
 
-Convert agent-eval's paired result into the portable Interface comparison.
+`AgentImprovementActivation`
 
-#### Type Parameters
+***
 
-##### TScenario
+### verifyCandidateExecutionEvidence()
 
-`TScenario` *extends* `Scenario`
+> **verifyCandidateExecutionEvidence**(`input`, `options`): `CandidateExecutionEvidence`
 
-##### TArtifact
+Defined in: [intelligence/improvement-cycle.ts:487](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L487)
 
-`TArtifact`
+Recheck one Runtime receipt against its exact signed experiment cell.
 
 #### Parameters
 
+##### input
+
+`unknown`
+
 ##### options
 
-`CreateAgentImprovementMeasuredComparisonOptions`\<`TScenario`, `TArtifact`\>
+[`VerifyCandidateExecutionEvidenceOptions`](#verifycandidateexecutionevidenceoptions)
 
 #### Returns
 
-`AgentImprovementMeasuredComparison`
+`CandidateExecutionEvidence`
 
 ***
 
@@ -3748,7 +4204,7 @@ Convert agent-eval's paired result into the portable Interface comparison.
 
 > **createIntelligenceClient**(`config`): [`IntelligenceClient`](#intelligenceclient)
 
-Defined in: [intelligence/index.ts:469](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L469)
+Defined in: [intelligence/index.ts:477](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/index.ts#L477)
 
 Create an Observe-mode Intelligence client. Resolves effort, the base URL, and
 the redactor up front; the exporter is built lazily and is `undefined` when no
@@ -3882,23 +4338,23 @@ Lower a plane `CertifiedProfile` straight into a `ResolvedSurface` via
 
 ***
 
-### createSandboxApprovedCandidateExecutor()
+### createSandboxCandidateExperimentExecutor()
 
-> **createSandboxApprovedCandidateExecutor**(`options`): [`SandboxApprovedCandidateExecutor`](#sandboxapprovedcandidateexecutor)
+> **createSandboxCandidateExperimentExecutor**(`options`): [`SandboxCandidateExperimentExecutor`](#sandboxcandidateexperimentexecutor)
 
-Defined in: [intelligence/sandbox-approved-candidate.ts:100](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L100)
+Defined in: [intelligence/sandbox-approved-candidate.ts:109](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/sandbox-approved-candidate.ts#L109)
 
-Compose approved-candidate execution directly onto fresh Tangle sandboxes.
+Execute one signed experiment cell inside a fresh Tangle sandbox.
 
 #### Parameters
 
 ##### options
 
-[`CreateSandboxApprovedCandidateExecutorOptions`](#createsandboxapprovedcandidateexecutoroptions)
+[`CreateSandboxCandidateExperimentExecutorOptions`](#createsandboxcandidateexperimentexecutoroptions)
 
 #### Returns
 
-[`SandboxApprovedCandidateExecutor`](#sandboxapprovedcandidateexecutor)
+[`SandboxCandidateExperimentExecutor`](#sandboxcandidateexperimentexecutor)
 
 ***
 
diff --git a/docs/api/primitive-catalog.md b/docs/api/primitive-catalog.md
index dc38bb52..ab3795a1 100644
--- a/docs/api/primitive-catalog.md
+++ b/docs/api/primitive-catalog.md
@@ -7,7 +7,7 @@
 
 # Primitive catalog — the never-stale anti-reinvention inventory
 
-> **GENERATED** from `@tangle-network/agent-runtime@0.95.0` and `@tangle-network/agent-eval@0.121.0` by `scripts/gen-primitive-catalog.mjs`. Do NOT hand-edit — run `pnpm run docs:api`. This is the mechanical companion to the JUDGMENT in `canonical-api.md` (§2 decision table + §1.5 AgentProfile law): that doc says WHICH primitive to reach for and what NOT to build; this catalog proves WHAT exists. Per-symbol signatures + `file:line` live in the per-module pages under `docs/api/`.
+> **GENERATED** from `@tangle-network/agent-runtime@0.95.0` and `@tangle-network/agent-eval@0.122.1` by `scripts/gen-primitive-catalog.mjs`. Do NOT hand-edit — run `pnpm run docs:api`. This is the mechanical companion to the JUDGMENT in `canonical-api.md` (§2 decision table + §1.5 AgentProfile law): that doc says WHICH primitive to reach for and what NOT to build; this catalog proves WHAT exists. Per-symbol signatures + `file:line` live in the per-module pages under `docs/api/`.
 
 ## 1. agent-runtime — own public surface
 
@@ -15,7 +15,7 @@ Every subpath this package declares in `package.json` `exports`. Reach for these
 
 ### Root — task lifecycle, conversation, RSI verbs, observability
 
-Import from `@tangle-network/agent-runtime` — 343 exports.
+Import from `@tangle-network/agent-runtime` — 345 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -144,6 +144,7 @@ Import from `@tangle-network/agent-runtime` — 343 exports.
 | `SqlConversationJournal` | class | SQL-backed ConversationJournal. Two tables — runs (one row per runId, holds |
 | `ValidationError` | class | Caller passed invalid arguments (out of range, mutually-exclusive options, bad shape). |
 | `AgentCandidateArtifactPort` | interface | Reads one content-addressed object from the closed S3/IPFS locator set. |
+| `AgentCandidateBenchmarkGraderIdentity` | interface | Immutable grader identity admitted for one benchmark task. |
 | `AgentCandidateBenchmarkGraderPort` | interface | Evaluator-owned executable grader, pinned by immutable implementation bytes. |
 | `AgentCandidateCodeSurfaceSource` | interface | The only accepted path from an agent-eval code candidate to executable bytes. |
 | `AgentCandidateExecutionAttemptRecord` | interface | Persisted state available to a fresh trusted recovery worker after a crash. |
@@ -155,12 +156,13 @@ Import from `@tangle-network/agent-runtime` — 343 exports.
 | `AgentCandidateExecutorFinalCapture` | interface | Replayable evaluator result captured only after process death and trace drain. |
 | `AgentCandidateExecutorMemoryCapture` | interface | Raw isolated-memory capture made only after access has been revoked. |
 | `AgentCandidateExecutorPort` | interface | Executes one prepared request inside an evaluator-owned isolation boundary. |
+| `AgentCandidateExecutorProfileFile` | interface | One exact profile file supplied to an evaluator-owned executor. |
 | `AgentCandidateExecutorRequest` | interface | One detached request passed to the trusted environment-specific executor. |
 | `AgentCandidateExecutorStopRequest` | interface | Opaque process identity used for termination without re-exposing launch credentials. |
 | `AgentCandidateModelGrantClient` | interface | Narrow transport contract for a service that owns scoped model credentials |
 | `AgentCandidateOutputArtifactPort` | interface | Durable content-addressed evidence store controlled only by the evaluator. |
 | `AgentCandidateRepositoryPort` | interface | Resolves a declared GitHub repository to an already-present local Git object store. |
-| `AgentCandidateTaskExecution` | interface | One signed benchmark task and the exact result shape its executor must capture. |
+| `AgentCandidateTaskExecution` | interface | Runtime placement for one exact cell from a signed candidate experiment. |
 | `AgentCandidateWorkspacePort` | interface | Materializes an already-verified workspace archive. |
 | `BackendErrorDetail` | interface | Typed transport / backend failure detail. Carried on `backend_error` and |
 | `BuildAgentCandidateBundleInput` | interface | Complete measured surfaces and execution policy compiled into one candidate bundle. |
@@ -218,7 +220,7 @@ Import from `@tangle-network/agent-runtime` — 343 exports.
 | `VerifiedAgentCandidateTaskOutcome` | type | Branded task outcome that has survived independent evaluator verification. |
 | `Verifier` | type | Verifies the edited worktree. Sync or async; throws only on a setup fault |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentAdapter`, `AgentBackendContext`, `AgentBackendInput`, `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `AgentExecutionBackend`, `AgenticGeneratorOptions`, `AgenticGeneratorShotReceipt`, `AgentKnowledgeProvider`, `AgentKnowledgeReadinessCheckOptions`, `AgentTaskContext`, `AgentTaskRunResult`, `AgentTaskSpec`, `ApprovedKnowledgeImprovementCandidate`, `BackendCallPolicy`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `ChatTurnHooks`, `ChatTurnResult`, `ControlBudget`, `ControlEvalResult`, `ControlRunResult`, `ControlStep`, `Conversation`, `ConversationDriveState`, `ConversationJournal`, `ConversationParticipant`, `ConversationPolicy`, `ConversationResult`, `ConversationTurn`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `D1StmtLike`, `DataAcquisitionPlan`, `DelegatedLoopResult`, `DisposePreparedAgentCandidateOptions`, `EvalRunEvent`, `EvalRunGeneration`, `EvalRunsExportConfig`, `EvalRunsExportResult`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `HaltContext`, `HaltSignal`, `ImproveCodeOptions`, `ImproveMemoryOptions`, `ImprovementDriverOptions`, `ImproveResult`, `ImproveSkillsOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `KnowledgeReadinessReport`, `KnowledgeRequirement`, `LoopRunnerCliArgs`, `LoopRunnerCliResult`, `ManagedImprovementDriver`, `OtelAttribute`, `OtelExporter`, `OtelSpan`, `PersonaConversationResult`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `ProfileDiffProposerOptions`, `RecoverExpiredAgentCandidateOptions`, `ResearchLoopResult`, `ResearchLoopRunnerOptions`, `ResolveAgentBackendOptions`, `ResolvedAgentCandidateContainer`, `ResolvedChatModel`, `RunChatTurnInput`, `RunConversationOptions`, `RunDelegatedLoopOptions`, `RunKnowledgeImprovementJobOptions`, `RunPersonaConfig`, `RunPersonaConversationOptions`, `RuntimeDecisionEvidenceRef`, `RuntimeDecisionPoint`, `RuntimeEventCollector`, `RuntimeEventOtelOptions`, `RuntimeHookContext`, `RuntimeHookErrorContext`, `RuntimeHookEvent`, `RuntimeRunHandle`, `RuntimeRunPersistenceAdapter`, `RuntimeRunRow`, `RuntimeSessionStore`, `RuntimeStreamEventCollector`, `RuntimeTelemetryOptions`, `RunToolLoopOptions`, `SanitizedKnowledgeReadinessReport`, `StreamToolLoopOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `ToolLoopResult`, `VerifiedAgentCandidate`, `VetoedFact`, `WorktreeLoopRunnerOptions`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`, `AgentRuntimeEvent`, `AgentRuntimeEventSink`, `AgentTaskStatus`, `AuthSource`, `ControlDecision`, `ConversationStreamEvent`, `DelegatedLoopMode`, `DelegatedLoopRegistry`, `DelegatedLoopRunner`, `ForwardHeaderName`, `HaltPredicate`, `HaltReason`, `ImproveOptions`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `ProfileDiffProposerContext`, `RuntimeDecisionKind`, `RuntimeHookTarget`, `RuntimeStreamEvent`, `StreamToolLoopYield`, `SupervisedKnowledgeUpdater`, `ToolLoopEvent`, `TurnOrder`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentAdapter`, `AgentBackendContext`, `AgentBackendInput`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidatePreparationEvidence`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `AgentExecutionBackend`, `AgenticGeneratorOptions`, `AgenticGeneratorShotReceipt`, `AgentKnowledgeProvider`, `AgentKnowledgeReadinessCheckOptions`, `AgentTaskContext`, `AgentTaskRunResult`, `AgentTaskSpec`, `ApprovedKnowledgeImprovementCandidate`, `BackendCallPolicy`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `ChatTurnHooks`, `ChatTurnResult`, `ControlBudget`, `ControlEvalResult`, `ControlRunResult`, `ControlStep`, `Conversation`, `ConversationDriveState`, `ConversationJournal`, `ConversationParticipant`, `ConversationPolicy`, `ConversationResult`, `ConversationTurn`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `D1StmtLike`, `DataAcquisitionPlan`, `DelegatedLoopResult`, `DisposePreparedAgentCandidateOptions`, `EvalRunEvent`, `EvalRunGeneration`, `EvalRunsExportConfig`, `EvalRunsExportResult`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `HaltContext`, `HaltSignal`, `ImproveCodeOptions`, `ImproveMemoryOptions`, `ImprovementDriverOptions`, `ImproveResult`, `ImproveSkillsOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `KnowledgeReadinessReport`, `KnowledgeRequirement`, `LoopRunnerCliArgs`, `LoopRunnerCliResult`, `ManagedImprovementDriver`, `OtelAttribute`, `OtelExporter`, `OtelSpan`, `PersonaConversationResult`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `ProfileDiffProposerOptions`, `RecoverExpiredAgentCandidateOptions`, `ResearchLoopResult`, `ResearchLoopRunnerOptions`, `ResolveAgentBackendOptions`, `ResolvedAgentCandidateContainer`, `ResolvedChatModel`, `RunChatTurnInput`, `RunConversationOptions`, `RunDelegatedLoopOptions`, `RunKnowledgeImprovementJobOptions`, `RunPersonaConfig`, `RunPersonaConversationOptions`, `RuntimeDecisionEvidenceRef`, `RuntimeDecisionPoint`, `RuntimeEventCollector`, `RuntimeEventOtelOptions`, `RuntimeHookContext`, `RuntimeHookErrorContext`, `RuntimeHookEvent`, `RuntimeRunHandle`, `RuntimeRunPersistenceAdapter`, `RuntimeRunRow`, `RuntimeSessionStore`, `RuntimeStreamEventCollector`, `RuntimeTelemetryOptions`, `RunToolLoopOptions`, `SanitizedKnowledgeReadinessReport`, `StreamToolLoopOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `ToolLoopResult`, `VerifiedAgentCandidate`, `VetoedFact`, `WorktreeLoopRunnerOptions`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`, `AgentRuntimeEvent`, `AgentRuntimeEventSink`, `AgentTaskStatus`, `AuthSource`, `ControlDecision`, `ConversationStreamEvent`, `DelegatedLoopMode`, `DelegatedLoopRegistry`, `DelegatedLoopRunner`, `ForwardHeaderName`, `HaltPredicate`, `HaltReason`, `ImproveOptions`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `ProfileDiffProposerContext`, `RuntimeDecisionKind`, `RuntimeHookTarget`, `RuntimeStreamEvent`, `StreamToolLoopYield`, `SupervisedKnowledgeUpdater`, `ToolLoopEvent`, `TurnOrder`.
 
 ### Vertical agent — manifest + improvement adapter
 
@@ -305,7 +307,7 @@ Import from `@tangle-network/agent-runtime/conversation` — 53 exports.
 
 ### Intelligence SDK — Observe + provable-OFF billing
 
-Import from `@tangle-network/agent-runtime/intelligence` — 95 exports.
+Import from `@tangle-network/agent-runtime/intelligence` — 102 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -313,35 +315,39 @@ Import from `@tangle-network/agent-runtime/intelligence` — 95 exports.
 | `composeCertifiedProfile` | function | Compose a certified profile into a uniform `ResolvedSurface`. Additive over |
 | `composeCertifiedProfileFromWire` | function | Lower a plane `CertifiedProfile` straight into a `ResolvedSurface` via |
 | `composeCertifiedPrompt` | function | Fold the certified prompt surface (and any certified prompt-folding artifacts: |
-| `createAgentImprovementMeasuredComparison` | function | Convert agent-eval's paired result into the portable Interface comparison. |
-| `createAgentImprovementProposal` | function | Freeze an already-measured improvement into the one reviewable proposal |
+| `createAgentImprovementActivation` | function | Authorize product-owned writes only after the exact candidate was measured and approved. |
+| `createAgentImprovementMeasuredComparison` | function | Delegate all statistics and promotion checks to agent-eval's receipt-based comparison. |
+| `createAgentImprovementProposal` | function | Create the reviewable record only from a complete, recomputable experiment result. |
 | `createCertifiedPromptSource` | function | Create the cached certified-prompt source — the ONE module-scope-cache + |
 | `createIntelligenceClient` | function | Create an Observe-mode Intelligence client. Resolves effort, the base URL, and |
-| `createSandboxApprovedCandidateExecutor` | function | Compose approved-candidate execution directly onto fresh Tangle sandboxes. |
+| `createSandboxCandidateExperimentExecutor` | function | Execute one signed experiment cell inside a fresh Tangle sandbox. |
 | `defaultRedactor` | function | The built-in redactor. Walks objects and arrays; replaces values under |
-| `executeApprovedAgentCandidate` | function | Verify, materialize, run, grade, and receipt only the exact approved bundle. |
+| `executeAgentCandidateExperimentCell` | function | Execute one exact arm, task, repetition, seed, and attempt through Runtime. |
 | `isIntelligenceOff` | function | True when these settings admit NO intelligence spawn — the passthrough |
 | `manifestFromProfile` | function | Lower the EXISTING plane wire (`CertifiedProfile`) into a `CapabilityManifest`. |
 | `normalizeCertifiedProfile` | function | Deserialize the composed-endpoint response into a `CertifiedProfile`. The |
 | `parseAgentCandidateProfileActivation` | function | Parse and check every native file hash plus both canonical document digests. |
-| `proposeAgentImprovement` | function | Analyze one run and freeze its measured winner into one exact proposal. |
+| `proposeAgentImprovement` | function | Analyze, search, then remeasure the resulting exact candidate before proposing it. |
 | `pullCertified` | function | Pull the certified composed profile for a target. Fail-closed: a network |
 | `resolveEffort` | function | Compile a named tier (plus optional per-field overrides) into the flat |
 | `resolveIntelligenceBaseUrl` | function | Resolve the ONE Intelligence base URL — the single knob both the send and |
 | `resolveRedactor` | function | Resolve the redactor a client uses. A caller-supplied hook replaces the |
-| `reviewAgentImprovementProposal` | function | Persist an approve/reject/change-request decision bound to one exact proposal. |
-| `verifyAgentImprovementProposal` | function | Validate a proposal's schema, profile, sealed bundle, and canonical digest. |
-| `verifyAgentImprovementReview` | function | Validate a review's decision fields and canonical digest. |
-| `verifyCandidateExecutionEvidence` | function | Verify approval, receipts, uniqueness, and the exact native profile files executed. |
+| `reviewAgentImprovementProposal` | function | Persist a human or tenant-policy decision bound to one exact proposal. |
+| `runAgentCandidateExperiment` | function | Execute both arms of one immutable experiment and derive its paired result. |
+| `verifyAgentImprovementActivation` | function | Validate activation authority against the exact proposal, review, experiment, and base state. |
+| `verifyAgentImprovementProposal` | function | Validate a proposal and recompute every binding to its measured experiment. |
+| `verifyAgentImprovementReview` | function | Validate the canonical identity and wire shape of an improvement review. |
+| `verifyCandidateExecutionEvidence` | function | Recheck one Runtime receipt against its exact signed experiment cell. |
 | `withIntelligence` | function | Wrap an agent so it (a) RECEIVES the tenant's certified profile — the prompt |
 | `defaultEffortTier` | const | The default tier when a client declares no effort. `'standard'` turns |
-| `sandboxApprovedCandidateExecutionSupport` | const | Declares the exact candidate surfaces the sandbox executor can run. |
+| `sandboxCandidateExperimentExecutionSupport` | const | Declares the exact candidate surfaces the sandbox executor can run. |
+| `AgentCandidateExperimentCellExecutionError` | class | A failed baseline or candidate cell with its complete Runtime failure result. |
 | `CapabilityNotAdmittedError` | class | A binding kind whose resolver case is typed but not yet admitted (rag-index, |
 | `AgentCandidateProfileActivation` | interface | Exact native profile files and the canonical plan that activated them. |
 | `AgentImprovementMeasuredComparison` | interface | Portable paired held-out comparison produced by an evaluation package. |
 | `AgentImprovementReview` | interface | Human or tenant-policy decision bound to one exact proposal. |
 | `AppliedIntelligence` | interface | What the hook hands the agent each run. Additive over the prompt-only |
-| `CandidateExecutionEvidence` | interface | Successful post-approval execution, carrying the exact Runtime receipt. |
+| `CandidateExecutionEvidence` | interface | Complete execution of one exact experiment attempt. |
 | `CapabilityManifest` | interface | The strict generalization of `CertifiedProfile`. `promptSurface` is kept |
 | `CertifiedArtifact` | interface | A promoted, certified artifact (one entry in the composed profile). |
 | `CertifiedCapability` | interface | One certified unit of agent power. |
@@ -392,7 +398,7 @@ Import from `@tangle-network/agent-runtime/intelligence` — 95 exports.
 | `Redactor` | type | A redactor maps an arbitrary trace value to a safe-to-export value. Pure; |
 | `UsageClass` | type | Usage class for billing. Base-stream tokens bill `'inference'`; every |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentImprovementProposal`, `CreateAgentImprovementProposalOptions`, `CreateSandboxApprovedCandidateExecutorOptions`, `ExecuteApprovedAgentCandidateOptions`, `ProposeAgentImprovementOptions`, `ProposeAgentImprovementResult`, `PullCertifiedOptions`, `ReviewAgentImprovementInput`, `SandboxApprovedCandidateExecution`, `SandboxApprovedCandidateExecutor`, `VerifyCandidateExecutionEvidenceOptions`, `AgentImprovementReviewDecision`, `ExecuteApprovedAgentCandidateResult`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentCandidateExperimentCellPlacement`, `AgentImprovementProposal`, `CreateAgentImprovementActivationOptions`, `CreateAgentImprovementProposalOptions`, `CreateSandboxCandidateExperimentExecutorOptions`, `ExecuteAgentCandidateExperimentCellOptions`, `ProposeAgentImprovementOptions`, `ProposeAgentImprovementResult`, `PullCertifiedOptions`, `ReviewAgentImprovementInput`, `RunAgentCandidateExperimentOptions`, `RunAgentCandidateExperimentResult`, `SandboxCandidateExperimentExecution`, `SandboxCandidateExperimentExecutor`, `VerifyCandidateExecutionEvidenceOptions`, `AgentImprovementReviewDecision`.
 
 ### Recursive atom + loop kernel (alias of ./runtime)
 
@@ -948,7 +954,7 @@ Import from `@tangle-network/agent-runtime/primeintellect` — 27 exports.
 
 ### Candidate execution — immutable prepare, run, grade, and receipt
 
-Import from `@tangle-network/agent-runtime/candidate-execution` — 97 exports.
+Import from `@tangle-network/agent-runtime/candidate-execution` — 99 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -977,6 +983,7 @@ Import from `@tangle-network/agent-runtime/candidate-execution` — 97 exports.
 | `FileAgentCandidateExecutionClaimStore` | class | Cross-process lifecycle implemented as fsynced, create-if-absent records. |
 | `InMemoryAgentCandidateExecutionClaimStore` | class | Single-process lifecycle implementation. |
 | `AgentCandidateArtifactPort` | interface | Reads one content-addressed object from the closed S3/IPFS locator set. |
+| `AgentCandidateBenchmarkGraderIdentity` | interface | Immutable grader identity admitted for one benchmark task. |
 | `AgentCandidateBenchmarkGraderPort` | interface | Evaluator-owned executable grader, pinned by immutable implementation bytes. |
 | `AgentCandidateCodeSurfaceSource` | interface | The only accepted path from an agent-eval code candidate to executable bytes. |
 | `AgentCandidateExecutionAttemptRecord` | interface | Persisted state available to a fresh trusted recovery worker after a crash. |
@@ -988,12 +995,13 @@ Import from `@tangle-network/agent-runtime/candidate-execution` — 97 exports.
 | `AgentCandidateExecutorFinalCapture` | interface | Replayable evaluator result captured only after process death and trace drain. |
 | `AgentCandidateExecutorMemoryCapture` | interface | Raw isolated-memory capture made only after access has been revoked. |
 | `AgentCandidateExecutorPort` | interface | Executes one prepared request inside an evaluator-owned isolation boundary. |
+| `AgentCandidateExecutorProfileFile` | interface | One exact profile file supplied to an evaluator-owned executor. |
 | `AgentCandidateExecutorRequest` | interface | One detached request passed to the trusted environment-specific executor. |
 | `AgentCandidateExecutorStopRequest` | interface | Opaque process identity used for termination without re-exposing launch credentials. |
 | `AgentCandidateModelGrantClient` | interface | Narrow transport contract for a service that owns scoped model credentials |
 | `AgentCandidateOutputArtifactPort` | interface | Durable content-addressed evidence store controlled only by the evaluator. |
 | `AgentCandidateRepositoryPort` | interface | Resolves a declared GitHub repository to an already-present local Git object store. |
-| `AgentCandidateTaskExecution` | interface | One signed benchmark task and the exact result shape its executor must capture. |
+| `AgentCandidateTaskExecution` | interface | Runtime placement for one exact cell from a signed candidate experiment. |
 | `AgentCandidateWorkspacePort` | interface | Materializes an already-verified workspace archive. |
 | `BuildAgentCandidateBundleInput` | interface | Complete measured surfaces and execution policy compiled into one candidate bundle. |
 | `PreparedAgentCandidateKnowledge` | interface | Exact file-backed knowledge admitted by the candidate bundle. |
@@ -1013,7 +1021,7 @@ Import from `@tangle-network/agent-runtime/candidate-execution` — 97 exports.
 | `AgentCandidateProfileSource` | type | A complete profile that can be frozen without losing behavior. |
 | `VerifiedAgentCandidateTaskOutcome` | type | Branded task outcome that has survived independent evaluator verification. |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `DisposePreparedAgentCandidateOptions`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `RecoverExpiredAgentCandidateOptions`, `ResolvedAgentCandidateContainer`, `VerifiedAgentCandidate`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidatePreparationEvidence`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateVerificationPorts`, `AgentCandidateWorkspaceArchiveLimits`, `CanonicalCandidateDocument`, `CaptureAgentCandidateWorkspaceOptions`, `CapturedAgentCandidateWorkspace`, `CreateAgentCandidateWorkspacePortOptions`, `CreateProtectedAgentCandidateModelPortOptions`, `DisposePreparedAgentCandidateOptions`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `RecoverExpiredAgentCandidateOptions`, `ResolvedAgentCandidateContainer`, `VerifiedAgentCandidate`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`.
 
 ### MCP servers — delegate / coordination / detached-session
 
@@ -1245,7 +1253,7 @@ Import from `@tangle-network/agent-eval` — 51 exports.
 
 ### CAMPAIGN — profile matrix, gates, improvement loop
 
-Import from `@tangle-network/agent-eval/campaign` — 391 exports.
+Import from `@tangle-network/agent-eval/campaign` — 392 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -1267,6 +1275,7 @@ Import from `@tangle-network/agent-eval/campaign` — 391 exports.
 | `campaignScenarioIdentity` | function | Redacted but independently verifiable identity of one complete scenario. |
 | `campaignSplitDigest` | function | Canonical identity of the exact scenario payloads and replicate count. |
 | `campaignSplitDigestFromIdentities` | function | Canonical split identity reconstructed from redacted scenario identities. |
+| `canonicalDigest` | function | _(no summary — add a TSDoc line at the declaration)_ |
 | `classifyUngroundedLiterals` | function | Scan revised artifact text for single-quoted single-word literals (the |
 | `codeSurfaceIdentityMaterial` | function | Canonical, location-independent identity of a finalized code candidate. |
 | `compareProposers` | function | Run a head-to-head lift benchmark across surface proposers on a shared holdout, returning per-proposer lift CIs and pairwise "who wins" verdicts. |
diff --git a/docs/canonical-api.md b/docs/canonical-api.md
index d66a8965..f17b9f53 100644
--- a/docs/canonical-api.md
+++ b/docs/canonical-api.md
@@ -2,7 +2,7 @@
 
 <!-- This doc is the JUDGMENT layer: the mental model (§1), the AgentProfile law (§1.5), and the anti-reinvention decision table (§2) — WHICH primitive to reach for and what NOT to build. The export INVENTORY (WHAT exists) and per-symbol signatures + `file:line` are GENERATED into `docs/api/` (TypeDoc + `scripts/gen-primitive-catalog.mjs`, do NOT hand-edit) — that is the mechanical reference: `docs/api/primitive-catalog.md` is the never-stale list of every primitive to reuse. The freshness gate (`pnpm docs:freshness`) FAILS CI if a version pin, a cited `file:line`, a decision-table symbol, or the generated catalog drifts from source — see `docs/MAINTAINING.md`. Keep this file the small, hand-curated spine; never re-list the inventory here — point at the catalog. -->
 
-> **Version 0.95.0.** The export inventory + per-symbol signatures live in the generated `docs/api/` reference: **`docs/api/primitive-catalog.md`** is the never-stale, grouped list of every primitive to reuse (own surface + the agent-eval judge / authenticity / verification / statistics / campaign / token-usage surfaces), with each one's import path and one-line summary read live from source; the per-module pages hold the full signatures. The pinned substrate is agent-eval `>=0.121.0 <0.122.0`; the sandbox substrate that materializes profiles into harness shapes is `@tangle-network/sandbox` (peer `>=0.10.5 <1.0.0`). The neutral contract types (`AgentProfile`, `AgentProfileMcpServer`, `HarnessType`, `ReasoningEffort`, `Part`/`ToolPart`/`ToolState`, plus environment-provider types) are owned by **`@tangle-network/agent-interface`** (peer `>=0.28.0 <0.29.0`) — the single source of truth. Substrate primitives are re-exported through `@tangle-network/agent-eval/contract` (or `/campaign`), not local to this package — the catalog's §2 shows exactly which subpath each lives under.
+> **Version 0.95.0.** The export inventory + per-symbol signatures live in the generated `docs/api/` reference: **`docs/api/primitive-catalog.md`** is the never-stale, grouped list of every primitive to reuse (own surface + the agent-eval judge / authenticity / verification / statistics / campaign / token-usage surfaces), with each one's import path and one-line summary read live from source; the per-module pages hold the full signatures. The pinned substrate is agent-eval `>=0.122.1 <0.123.0`; the sandbox substrate that materializes profiles into harness shapes is `@tangle-network/sandbox` (peer `>=0.11.1 <1.0.0`). The neutral contract types (`AgentProfile`, `AgentProfileMcpServer`, `HarnessType`, `ReasoningEffort`, `Part`/`ToolPart`/`ToolState`, plus environment-provider types) are owned by **`@tangle-network/agent-interface`** (peer `>=0.30.0 <0.31.0`) — the single source of truth. Substrate primitives are re-exported through `@tangle-network/agent-eval/contract` (or `/campaign`), not local to this package — the catalog's §2 shows exactly which subpath each lives under.
 >
 > **`./loops` is the runtime barrel** — `package.json` maps it to `src/runtime/index.ts`. Everything below labelled `/loops` is the recursive-atom + loop-kernel surface.
 >
@@ -110,10 +110,11 @@ A general "loop" primitive is the single most common modelling error in this rep
 | Have a **supervisor spawn + live-drive workers in a backend you choose** and observe or steer them while the coordinator is alive | the **coordination MCP** via `createCoordinationTools` / `serveCoordinationMcp` over a live `Scope`; each worker's leaf is `createExecutor({ backend })` | `detachedSessionDelegate`, which is own-sandbox-session only and one-shot. Supervised-tree restart recovery is not implemented. |
 | Stand up a vertical agent in the eval loop | `defineAgent(manifest)` + `createSurfaceImprovementAdapter` — `/agent` | a per-vertical manifest parser, surface-validator, or bespoke `ImprovementAdapter` |
 | Observe + deliver Intelligence on a live agent (send RunRecords + receive certified profile/diffs) | `withIntelligence(agent, { project, target })` — `/intelligence` (proposals surfaced, never auto-applied; `effort: 'off'` proves inference-only billing) | a custom trace-wrapper, a second receive path, or hand-rolled effort/tier config |
-| Turn trace evidence into one measured, review-only agent proposal | `proposeAgentImprovement({ analysis, profile, improvement })` — `/intelligence` | manually joining analysts, optimizers, evidence, uncertainty, and candidate identity |
+| Turn trace evidence into one measured, review-only agent proposal | `proposeAgentImprovement({ analysis, profile, improvement, buildExperiment, placeCell })` in `/intelligence` | manually joining analysis, optimization, exact candidate execution, uncertainty, and candidate identity |
 | Freeze a measured profile/diff plus a content-addressed code surface into one executable candidate | `buildAgentCandidateBundle(...)`, then `verifyAgentCandidateBundle(...)` at execution — `/candidate-execution` | a product callback that converts profile fields, reproduces Git diff flags, hashes bytes, or assembles `AgentCandidateBundle` by hand |
 | Record approve/reject/change-request feedback against one exact proposal | `reviewAgentImprovementProposal(proposal, review)` — `/intelligence` | a mutable status row that is not bound to candidate bytes |
-| Execute and grade only an authenticated approved candidate | `executeApprovedAgentCandidate(options)` — `/intelligence`; low-level ports live at `/candidate-execution` | product-local claim/retry/isolation/receipt orchestration |
+| Run and grade the exact signed baseline-versus-candidate matrix before review | `runAgentCandidateExperiment({ experiment, placeCell })` in `/intelligence`; use `createSandboxCandidateExperimentExecutor(...)` for fresh Tangle sandboxes | product-local pairing, retry, isolation, receipt, and comparison code |
+| Authorize product writes only for the exact measured and approved candidate | `createAgentImprovementActivation(proposal, review, options)` and `verifyAgentImprovementActivation(...)` in `/intelligence` | a mutable approval flag that does not bind the experiment, candidate bytes, targets, or current base state |
 | Capture and restore exact task, candidate, or memory workspace bytes | `captureAgentCandidateWorkspace(...)` + `createAgentCandidateWorkspacePort(...)` — `/candidate-execution` | a product-specific archive format, ambient `git checkout`, or a materializer that skips byte/path/mode verification |
 | Fold **certified prompt additions into a system prompt you assemble yourself** (product chat routes) | `createCertifiedPromptSource({ target })` → `source.compose(base)` — `/intelligence` (cached, coalesced, fail-closed; `withIntelligence` rides the same source) | a module-scope cache + refresh-window + keep-last-known loop around `pullCertified` in product wiring |
 | Improve a KB with runtime agents, candidate workspaces, readiness checks, and measured supervised spend | `runKnowledgeImprovementJob(options)` from `/knowledge` | hand-wiring `improveKnowledgeBase` + a supervised updater + a readiness callback in every product |
diff --git a/package.json b/package.json
index fe05c16b..28a1cec8 100644
--- a/package.json
+++ b/package.json
@@ -118,9 +118,9 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.15",
-    "@tangle-network/agent-eval": "0.121.0",
-    "@tangle-network/agent-interface": "0.28.0",
-    "@tangle-network/sandbox": "^0.10.5",
+    "@tangle-network/agent-eval": "0.122.1",
+    "@tangle-network/agent-interface": "0.30.0",
+    "@tangle-network/sandbox": "^0.11.1",
     "@types/node": "^25.9.3",
     "@types/tar-stream": "3.1.4",
     "playwright": "^1.61.0",
@@ -150,9 +150,9 @@
   "license": "MIT",
   "packageManager": "pnpm@10.28.0",
   "peerDependencies": {
-    "@tangle-network/agent-eval": ">=0.121.0 <0.122.0",
-    "@tangle-network/agent-interface": ">=0.28.0 <0.29.0",
-    "@tangle-network/sandbox": ">=0.10.5 <1.0.0",
+    "@tangle-network/agent-eval": ">=0.122.1 <0.123.0",
+    "@tangle-network/agent-interface": ">=0.30.0 <0.31.0",
+    "@tangle-network/sandbox": ">=0.11.1 <1.0.0",
     "playwright": "^1.40.0"
   },
   "peerDependenciesMeta": {
@@ -164,8 +164,8 @@
     }
   },
   "dependencies": {
-    "@tangle-network/agent-knowledge": "^3.0.0",
-    "@tangle-network/agent-profile-materialize": "0.4.0",
+    "@tangle-network/agent-knowledge": "^3.0.1",
+    "@tangle-network/agent-profile-materialize": "0.5.1",
     "tar-stream": "3.2.0"
   }
 }
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index c960dccb..b9e4027f 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -9,11 +9,11 @@ importers:
   .:
     dependencies:
       '@tangle-network/agent-knowledge':
-        specifier: ^3.0.0
-        version: 3.0.0(typescript@5.9.3)
+        specifier: ^3.0.1
+        version: 3.0.1(typescript@5.9.3)
       '@tangle-network/agent-profile-materialize':
-        specifier: 0.4.0
-        version: 0.4.0
+        specifier: 0.5.1
+        version: 0.5.1
       tar-stream:
         specifier: 3.2.0
         version: 3.2.0
@@ -22,14 +22,14 @@ importers:
         specifier: ^2.4.15
         version: 2.4.15
       '@tangle-network/agent-eval':
-        specifier: 0.121.0
-        version: 0.121.0(typescript@5.9.3)
+        specifier: 0.122.1
+        version: 0.122.1(typescript@5.9.3)
       '@tangle-network/agent-interface':
-        specifier: 0.28.0
-        version: 0.28.0
+        specifier: 0.30.0
+        version: 0.30.0
       '@tangle-network/sandbox':
-        specifier: ^0.10.5
-        version: 0.10.5(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))
+        specifier: ^0.11.1
+        version: 0.11.1(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))
       '@types/node':
         specifier: ^25.9.3
         version: 25.9.3
@@ -61,17 +61,17 @@ importers:
   bench:
     dependencies:
       '@tangle-network/agent-eval':
-        specifier: 0.121.0
-        version: 0.121.0(typescript@5.9.3)
+        specifier: 0.122.1
+        version: 0.122.1(typescript@5.9.3)
       '@tangle-network/agent-interface':
-        specifier: 0.28.0
-        version: 0.28.0
+        specifier: 0.30.0
+        version: 0.30.0
       '@tangle-network/agent-runtime':
         specifier: workspace:*
         version: link:..
       '@tangle-network/sandbox':
-        specifier: ^0.10.5
-        version: 0.10.5(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))
+        specifier: ^0.11.1
+        version: 0.11.1(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))
     devDependencies:
       '@types/node':
         specifier: ^25.9.3
@@ -677,11 +677,8 @@ packages:
   '@tangle-network/agent-core@0.4.11':
     resolution: {integrity: sha512-5B1IjrJ8xDR7w8Hv/MSk2ixul6NEJQ5Ftzo+z6l7ipeFpc0yaf1+Kkml4HDUEmGW/rqdrNpBf9/MpT7i/SY0PA==}
 
-  '@tangle-network/agent-core@0.4.9':
-    resolution: {integrity: sha512-BFU4WV12Y6D28PX+o8LIVkIMD+cXwrL1uyV182l12Bn9zEu7VHt2oVMEEzbsN8L+X0xM9baEfMCHTFKFNJv7Aw==}
-
-  '@tangle-network/agent-eval@0.121.0':
-    resolution: {integrity: sha512-rBfwbK984Ro3Yeml2rVfJYojw1R9R1awRI0IIAf15V5jOrT7X4yHJYW8l73Qk3jHwdyISRLjB8jFmzdj3G14yQ==}
+  '@tangle-network/agent-eval@0.122.1':
+    resolution: {integrity: sha512-nBjoslpjtxDxBTeBKgxgSHb3CJH+nKmXX8u5X8fepB9OllmaqhnF4B/DoNp95b/KaezVnd4i2hT2Zip/WS2Kxg==}
     engines: {node: '>=20'}
     hasBin: true
 
@@ -691,28 +688,22 @@ packages:
   '@tangle-network/agent-interface@0.17.1':
     resolution: {integrity: sha512-B7dRJTo0HSUtgBCB1VMwkTFYkLUaRr/4BcRglrQuGhGUwOzKv1RYyMejOVh5M3a5AagY9N79f7GYbjcA3UmnIA==}
 
-  '@tangle-network/agent-interface@0.21.0':
-    resolution: {integrity: sha512-jDxhVJgxymrvU1RLWxWKueuaWQIpBAfrW8BuVTB5m2Y4eFMLo1SawDBEMDLdZN4/Gf34xrFrsRk+PAj9brGKMQ==}
-
-  '@tangle-network/agent-interface@0.24.0':
-    resolution: {integrity: sha512-iaHWNTYne49cBYXwb72NjGDw5bjT2KVJlfawXygbvkqnfUsQG57BS++Yjf/XYvwJJdDeEaGs+Xvy8SWYkvu0qA==}
-
   '@tangle-network/agent-interface@0.26.0':
     resolution: {integrity: sha512-/z4HavFr/9AbaHxi/13bFP9iSUt8oitZbw4wS9g9KAt2TeN2ec5/A2C106udWkKIETZLnu465TfU+K4KDVNkKw==}
 
-  '@tangle-network/agent-interface@0.28.0':
-    resolution: {integrity: sha512-OFI0HPVviHOV1SWG7s6sQMMAkofQkWDrhtUHZUlNAO48WOhT7RllskgS/qaWegOZWZlROoKY3NHg1nHopNgg+A==}
+  '@tangle-network/agent-interface@0.30.0':
+    resolution: {integrity: sha512-GmwPwamrzOFPn+Qx7W0nt9QRUF9VUiZrMNLmF6QpL23R+7TD21HllPblVOYB5MVqA728FpDLXit3HfqY8mStwg==}
 
-  '@tangle-network/agent-knowledge@3.0.0':
-    resolution: {integrity: sha512-gi7ZCaXxK6n/Btuv8i6B8J9nos8Yw9u8zVLDM0QgA3Nbn3JRqSD6Z0fUG7F1UEtokw6apmZxVIO1kasmqjy5og==}
+  '@tangle-network/agent-knowledge@3.0.1':
+    resolution: {integrity: sha512-bzlllxFx+ZNwdx+Zr1v/ED2qElttrTTqFbs/3FapaHp+p1t5Jhr2seoyLMhiSUJqPusm8d8VvSctlL4EbpOaoA==}
     engines: {node: '>=20.19.0'}
     hasBin: true
 
-  '@tangle-network/agent-profile-materialize@0.4.0':
-    resolution: {integrity: sha512-0chH3bIA1G9XngrLdrxZ81mKmqy4GKWymNtmSn5w2xPcI3GolkxJ4RbwVWdxtjGKlZaKS6eWxnkl7sDXofcBNg==}
+  '@tangle-network/agent-profile-materialize@0.5.1':
+    resolution: {integrity: sha512-gU4J6mglvy9bpBAnErOruO1goF1cSzGGd06Oo5CFSj7RjR1SXWDFMYQtVLSd/O68ldCVAMVIMq5b4rwboMBbcw==}
 
-  '@tangle-network/sandbox@0.10.5':
-    resolution: {integrity: sha512-u76Ht9f7HahLdIcZUz98FwomHbfWOsBusydbpd532nVF4UfcQ9tlntSXUJ4huiIvjplAp8HW/wCOeSt3aow7Qg==}
+  '@tangle-network/sandbox@0.11.1':
+    resolution: {integrity: sha512-9I5G1UHNhwTNFToe+Y0iQO1Joemd+7ba3LXIjfYUj+apRww6Eqq+1UGV2a11XaWUTlEqyhAc37F5OHBUj/xxCg==}
     peerDependencies:
       '@mastra/core': ^1.36.0
       '@modelcontextprotocol/sdk': ^1.29.0
@@ -1765,18 +1756,13 @@ snapshots:
       '@tangle-network/agent-interface': 0.26.0
       zod: 4.4.3
 
-  '@tangle-network/agent-core@0.4.9':
-    dependencies:
-      '@tangle-network/agent-interface': 0.24.0
-      zod: 4.4.3
-
-  '@tangle-network/agent-eval@0.121.0(typescript@5.9.3)':
+  '@tangle-network/agent-eval@0.122.1(typescript@5.9.3)':
     dependencies:
       '@asteasolutions/zod-to-openapi': 8.5.0(zod@4.4.3)
       '@ax-llm/ax': 23.0.1(zod@4.4.3)
       '@hono/node-server': 2.0.8(hono@4.12.30)
       '@tangle-network/agent-core': 0.4.11
-      '@tangle-network/agent-interface': 0.28.0
+      '@tangle-network/agent-interface': 0.30.0
       '@tangle-network/tcloud': 0.4.14(typescript@5.9.3)(zod@4.4.3)
       hono: 4.12.30
       zod: 4.4.3
@@ -1797,28 +1783,20 @@ snapshots:
     dependencies:
       zod: 4.4.3
 
-  '@tangle-network/agent-interface@0.21.0':
-    dependencies:
-      zod: 4.4.3
-
-  '@tangle-network/agent-interface@0.24.0':
-    dependencies:
-      zod: 4.4.3
-
   '@tangle-network/agent-interface@0.26.0':
     dependencies:
       '@noble/hashes': 1.8.0
       zod: 4.4.3
 
-  '@tangle-network/agent-interface@0.28.0':
+  '@tangle-network/agent-interface@0.30.0':
     dependencies:
       '@noble/hashes': 1.8.0
       zod: 4.4.3
 
-  '@tangle-network/agent-knowledge@3.0.0(typescript@5.9.3)':
+  '@tangle-network/agent-knowledge@3.0.1(typescript@5.9.3)':
     dependencies:
-      '@tangle-network/agent-eval': 0.121.0(typescript@5.9.3)
-      '@tangle-network/agent-interface': 0.28.0
+      '@tangle-network/agent-eval': 0.122.1(typescript@5.9.3)
+      '@tangle-network/agent-interface': 0.30.0
       proper-lockfile: 4.1.2
       zod: 4.4.3
     transitivePeerDependencies:
@@ -1830,14 +1808,14 @@ snapshots:
       - typescript
       - utf-8-validate
 
-  '@tangle-network/agent-profile-materialize@0.4.0':
+  '@tangle-network/agent-profile-materialize@0.5.1':
     dependencies:
-      '@tangle-network/agent-interface': 0.28.0
+      '@tangle-network/agent-interface': 0.30.0
 
-  '@tangle-network/sandbox@0.10.5(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))':
+  '@tangle-network/sandbox@0.11.1(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))':
     dependencies:
-      '@tangle-network/agent-core': 0.4.9
-      '@tangle-network/agent-interface': 0.21.0
+      '@tangle-network/agent-core': 0.4.11
+      '@tangle-network/agent-interface': 0.30.0
     optionalDependencies:
       viem: 2.54.6(typescript@5.9.3)(zod@4.4.3)
 
diff --git a/scripts/verify-package-exports.mjs b/scripts/verify-package-exports.mjs
index 37083527..91efd964 100644
--- a/scripts/verify-package-exports.mjs
+++ b/scripts/verify-package-exports.mjs
@@ -107,7 +107,10 @@ try {
           '@tangle-network/agent-runtime': `file:${tarballs[0]}`,
           ...peerDependencies,
         },
-        devDependencies: { typescript: repoPackageJson.devDependencies.typescript },
+        devDependencies: {
+          '@types/node': repoPackageJson.devDependencies['@types/node'],
+          typescript: repoPackageJson.devDependencies.typescript,
+        },
       },
       null,
       2,
@@ -140,18 +143,17 @@ try {
     `
       import type {
         AgentCandidateProfileActivation,
-        AgentImprovementProposal,
-        AgentImprovementReview,
         CandidateExecutionEvidence,
       } from '@tangle-network/agent-interface'
       import { SandboxClient } from '@tangle-network/sandbox'
-      import type { AgentCandidateTaskExecution } from '@tangle-network/agent-runtime/candidate-execution'
       import {
-        createSandboxApprovedCandidateExecutor,
+        createSandboxCandidateExperimentExecutor,
         parseAgentCandidateProfileActivation,
-        sandboxApprovedCandidateExecutionSupport,
+        sandboxCandidateExperimentExecutionSupport,
         verifyCandidateExecutionEvidence,
-        type CreateSandboxApprovedCandidateExecutorOptions,
+        type CreateSandboxCandidateExperimentExecutorOptions,
+        type SandboxCandidateExperimentExecution,
+        type VerifyCandidateExecutionEvidenceOptions,
       } from '@tangle-network/agent-runtime/intelligence'
 
       const client = new SandboxClient({
@@ -159,38 +161,31 @@ try {
         baseUrl: 'https://sandbox.example.com',
         trustLocalCliAuth: false,
       })
-      declare const ports: CreateSandboxApprovedCandidateExecutorOptions['ports']
-      declare const grader: CreateSandboxApprovedCandidateExecutorOptions['grader']
-      declare const outputArtifacts: CreateSandboxApprovedCandidateExecutorOptions['outputArtifacts']
-      declare const traceStore: CreateSandboxApprovedCandidateExecutorOptions['traceStore']
-      declare const claimStore: CreateSandboxApprovedCandidateExecutorOptions['claimStore']
-      declare const proposal: AgentImprovementProposal
-      declare const review: AgentImprovementReview
-      declare const task: AgentCandidateTaskExecution
+      declare const ports: CreateSandboxCandidateExperimentExecutorOptions['ports']
+      declare const grader: CreateSandboxCandidateExperimentExecutorOptions['grader']
+      declare const outputArtifacts: CreateSandboxCandidateExperimentExecutorOptions['outputArtifacts']
+      declare const traceStore: CreateSandboxCandidateExperimentExecutorOptions['traceStore']
+      declare const claimStore: CreateSandboxCandidateExperimentExecutorOptions['claimStore']
+      declare const executionInput: SandboxCandidateExperimentExecution
+      declare const verification: VerifyCandidateExecutionEvidenceOptions
       declare const storedEvidence: unknown
 
-      const executor = createSandboxApprovedCandidateExecutor({
+      const executor = createSandboxCandidateExperimentExecutor({
         client,
         ports,
         grader,
         outputArtifacts,
         traceStore,
         claimStore,
-        authorizeReview: async () => true,
-      })
-      const execution: Promise<CandidateExecutionEvidence> = executor.execute({
-        proposal,
-        review,
-        task,
       })
-      const evidence = verifyCandidateExecutionEvidence([storedEvidence], {
-        proposal,
-        review,
-        expectedCount: 1,
-      })[0]
+      const execution: Promise<CandidateExecutionEvidence> = executor.execute(executionInput)
+      const evidence = verifyCandidateExecutionEvidence(storedEvidence, verification)
       const activation: AgentCandidateProfileActivation =
-        parseAgentCandidateProfileActivation(evidence?.profileActivation)
-      const outcome: 'output' = sandboxApprovedCandidateExecutionSupport.outcomes[0]
+        parseAgentCandidateProfileActivation(
+          evidence.materializationReceipt.profileActivation,
+          evidence.materializationReceipt.profileActivation.profilePlan.digest,
+        )
+      const outcome: 'output' = sandboxCandidateExperimentExecutionSupport.outcomes[0]
       void execution
       void activation
       void outcome
@@ -259,9 +254,11 @@ try {
           'composeCertifiedProfile',
           'manifestFromProfile',
           'CapabilityNotAdmittedError',
-          'createSandboxApprovedCandidateExecutor',
+          'createSandboxCandidateExperimentExecutor',
+          'executeAgentCandidateExperimentCell',
           'parseAgentCandidateProfileActivation',
-          'sandboxApprovedCandidateExecutionSupport',
+          'runAgentCandidateExperiment',
+          'sandboxCandidateExperimentExecutionSupport',
           'verifyCandidateExecutionEvidence',
         ]
         for (const name of expectedIntelligence) {
@@ -290,7 +287,7 @@ try {
             kind: 'profile',
             profile: { name: 'packed-consumer', harness: 'codex' },
           },
-          code: { kind: 'disabled', reason: 'control' },
+          code: { kind: 'disabled' },
           execution: {
             harness: 'codex',
             harnessVersion: '1.0.0',
@@ -305,7 +302,6 @@ try {
             },
           },
           memory: { mode: 'disabled' },
-          lineage: { source: 'human' },
         })
         const verified = await candidates.verifyAgentCandidateBundle(bundle, {
           artifacts: { read: async () => { throw new Error('unexpected artifact read') } },
diff --git a/src/candidate-execution/benchmark-grader.ts b/src/candidate-execution/benchmark-grader.ts
index 26981081..6357c74b 100644
--- a/src/candidate-execution/benchmark-grader.ts
+++ b/src/candidate-execution/benchmark-grader.ts
@@ -1,11 +1,12 @@
 import type { BenchmarkEvaluation } from '@tangle-network/agent-eval'
 import type {
-  AgentCandidateArtifactRef,
+  AgentCandidateBenchmarkGraderIdentity,
+  AgentCandidateFixedSpend,
   AgentCandidateTermination,
 } from '@tangle-network/agent-interface'
 
 import { readVerifiedArtifact } from './artifacts'
-import { immutableCandidateValue, sha256Bytes } from './digest'
+import { canonicalCandidateDigest, immutableCandidateValue, sha256Bytes } from './digest'
 import type {
   AgentCandidateBenchmarkGraderPort,
   AgentCandidateOutputArtifactPort,
@@ -13,10 +14,14 @@ import type {
 } from './types'
 
 export interface BoundAgentCandidateBenchmarkRun {
-  readonly grader: {
-    readonly name: string
-    readonly version: string
-    readonly artifact: AgentCandidateArtifactRef
+  readonly grader: AgentCandidateBenchmarkGraderIdentity
+  readonly grading: {
+    readonly usage: AgentCandidateFixedSpend
+    readonly timing: {
+      readonly startedAtMs: number
+      readonly endedAtMs: number
+      readonly durationMs: number
+    }
   }
   readonly evaluation: BenchmarkEvaluation
   readonly evidence: Uint8Array
@@ -30,12 +35,13 @@ export async function runBoundCandidateBenchmarkGrader(input: {
   executionId: string
   termination: AgentCandidateTermination
   outcome: VerifiedAgentCandidateTaskOutcome
+  expectedGrader: AgentCandidateBenchmarkGraderIdentity
   grader: AgentCandidateBenchmarkGraderPort
   artifacts: AgentCandidateOutputArtifactPort
   signal?: AbortSignal
 }): Promise<BoundAgentCandidateBenchmarkRun> {
   input.signal?.throwIfAborted()
-  const descriptor = snapshotGraderDescriptor(input.grader)
+  const descriptor = snapshotGraderDescriptor(input.grader, input.expectedGrader)
   const implementationBytes = await readVerifiedArtifact(descriptor.artifact, input.artifacts)
   if (implementationBytes.byteLength === 0) {
     throw new Error('candidate benchmark grader implementation cannot be empty')
@@ -43,6 +49,7 @@ export async function runBoundCandidateBenchmarkGrader(input: {
   const expectedImplementationDigest = sha256Bytes(implementationBytes)
   const termination = immutableCandidateValue(input.termination)
   const run = input.grader.run
+  const startedAtMs = Date.now()
   const result = await run(
     Object.freeze({
       executionId: input.executionId,
@@ -52,6 +59,7 @@ export async function runBoundCandidateBenchmarkGrader(input: {
       signal: input.signal ?? new AbortController().signal,
     }),
   )
+  const endedAtMs = Math.max(startedAtMs, Date.now())
   input.signal?.throwIfAborted()
   assertExactRunnerResult(result)
 
@@ -73,6 +81,14 @@ export async function runBoundCandidateBenchmarkGrader(input: {
 
   return Object.freeze({
     grader: descriptor,
+    grading: immutableCandidateValue({
+      usage: emptyFixedSpend(),
+      timing: {
+        startedAtMs,
+        endedAtMs,
+        durationMs: endedAtMs - startedAtMs,
+      },
+    }),
     evaluation: result.evaluation,
     evidence,
   })
@@ -80,15 +96,32 @@ export async function runBoundCandidateBenchmarkGrader(input: {
 
 function snapshotGraderDescriptor(
   grader: AgentCandidateBenchmarkGraderPort,
+  expected: AgentCandidateBenchmarkGraderIdentity,
 ): BoundAgentCandidateBenchmarkRun['grader'] {
   if (!grader.name || !grader.version) {
     throw new Error('candidate benchmark grader name and version must be non-empty')
   }
-  return immutableCandidateValue({
+  const actual = {
     name: grader.name,
     version: grader.version,
+    format: 'tangle-grader' as const,
     artifact: grader.artifact,
-  })
+  }
+  if (canonicalCandidateDigest(actual) !== canonicalCandidateDigest(expected)) {
+    throw new Error('candidate benchmark grader does not match the signed task grader')
+  }
+  return immutableCandidateValue(expected)
+}
+
+function emptyFixedSpend(): AgentCandidateFixedSpend {
+  return {
+    inputTokens: 0,
+    outputTokens: 0,
+    cachedInputTokens: 0,
+    reasoningTokens: 0,
+    modelCalls: 0,
+    costUsdNanos: 0,
+  }
 }
 
 function detachedImplementation(bytes: Uint8Array): {
diff --git a/src/candidate-execution/builder.ts b/src/candidate-execution/builder.ts
index a9ed5a59..640ba883 100644
--- a/src/candidate-execution/builder.ts
+++ b/src/candidate-execution/builder.ts
@@ -5,14 +5,14 @@ import type {
   AgentCandidateCodeNoOp,
   AgentCandidateExecution,
   AgentCandidateGitHubRepository,
-  AgentCandidateLineage,
+  AgentCandidateKnowledge,
   AgentCandidateMemoryPolicy,
   AgentCandidateProfile,
   AgentProfile,
   AgentProfileDiff,
 } from '@tangle-network/agent-interface'
 import { type AgentCandidateBundleInput, sealAgentCandidateBundle } from './bundle'
-import { canonicalCandidateDigest, embeddedCandidateArtifact } from './digest'
+import { embeddedCandidateArtifact } from './digest'
 import {
   applyExactAgentProfileDiff,
   freezeGenericAgentCandidateProfile,
@@ -30,7 +30,7 @@ export type AgentCandidateProfileSource =
   | {
       kind: 'profile-diffs'
       base: AgentProfile
-      /** Applied in order. Each exact diff is content-addressed into lineage. */
+      /** Applied in order before the resulting profile is frozen into the bundle. */
       diffs: readonly AgentProfileDiff[]
     }
   | {
@@ -59,9 +59,8 @@ export interface BuildAgentCandidateBundleInput {
   profile: AgentCandidateProfileSource
   code: AgentCandidateCodeSource
   execution: AgentCandidateExecution
+  knowledge?: AgentCandidateKnowledge
   memory: AgentCandidateMemoryPolicy
-  /** `profileDiffIds` is derived from `profile`; callers cannot contradict it. */
-  lineage: Omit<AgentCandidateLineage, 'profileDiffIds'>
 }
 
 /**
@@ -74,39 +73,25 @@ export interface BuildAgentCandidateBundleInput {
 export function buildAgentCandidateBundle(
   input: BuildAgentCandidateBundleInput,
 ): ReturnType<typeof sealAgentCandidateBundle> {
-  if (Object.hasOwn(input.lineage, 'profileDiffIds')) {
-    throw new Error('profileDiffIds is derived from the profile source and cannot be supplied')
-  }
-  const compiledProfile = compileCandidateProfile(input.profile)
-  const profileDiffIds = compiledProfile.profileDiffIds
   const bundle: AgentCandidateBundleInput = {
     kind: 'agent-candidate-bundle',
     digestAlgorithm: 'rfc8785-sha256',
-    profile: compiledProfile.profile,
+    profile: compileCandidateProfile(input.profile),
     code: compileCandidateCode(input.code),
     execution: input.execution,
+    ...(input.knowledge !== undefined ? { knowledge: input.knowledge } : {}),
     memory: input.memory,
-    lineage: {
-      ...input.lineage,
-      ...(profileDiffIds.length > 0 ? { profileDiffIds } : {}),
-    },
   }
   return sealAgentCandidateBundle(bundle)
 }
 
-function compileCandidateProfile(source: AgentCandidateProfileSource): {
-  profile: AgentCandidateProfile
-  profileDiffIds: string[]
-} {
+function compileCandidateProfile(source: AgentCandidateProfileSource): AgentCandidateProfile {
   if (source.kind === 'candidate-profile') {
-    return {
-      profile: parseExactCandidateProfile(source.profile),
-      profileDiffIds: [],
-    }
+    return parseExactCandidateProfile(source.profile)
   }
 
   if (source.kind === 'profile') {
-    return { profile: freezeGenericAgentCandidateProfile(source.profile), profileDiffIds: [] }
+    return freezeGenericAgentCandidateProfile(source.profile)
   }
 
   if (source.kind !== 'profile-diffs') {
@@ -118,13 +103,11 @@ function compileCandidateProfile(source: AgentCandidateProfileSource): {
     throw new Error('profile-diffs source requires at least one AgentProfileDiff')
   }
   let profile = parseExactAgentProfile(source.base, 'base profile')
-  const profileDiffIds: string[] = []
   for (const [index, inputDiff] of source.diffs.entries()) {
     const diff = parseExactAgentProfileDiff(inputDiff, `profile diff ${index}`)
     profile = applyExactAgentProfileDiff(profile, diff, `profile diff ${index}`)
-    profileDiffIds.push(canonicalCandidateDigest(diff))
   }
-  return { profile: freezeGenericAgentCandidateProfile(profile), profileDiffIds }
+  return freezeGenericAgentCandidateProfile(profile)
 }
 
 function compileCandidateCode(source: AgentCandidateCodeSource): AgentCandidateBundleInput['code'] {
diff --git a/src/candidate-execution/claim-plan.ts b/src/candidate-execution/claim-plan.ts
index 18283d83..a2a30425 100644
--- a/src/candidate-execution/claim-plan.ts
+++ b/src/candidate-execution/claim-plan.ts
@@ -16,7 +16,11 @@ export function candidateExecutionClaim(
 ): AgentCandidateExecutionClaim {
   const state = assertPreparedCandidateIntegrity(prepared)
   const material = prepared.executionPlan.value.material
-  const attempt = material.attempt
+  const attempt = {
+    number: material.runCell.attempt,
+    maxAttempts: state.benchmarkTask.attempt.maxAttempts,
+    retryPolicy: state.benchmarkTask.attempt.retryPolicy,
+  } as const
   const nowMs = Date.now()
   if (!Number.isSafeInteger(nowMs) || nowMs < 0) {
     throw new Error('candidate execution claim-store clock returned an invalid timestamp')
@@ -89,7 +93,11 @@ function retryLineageDigest(
     resultTimeoutMs,
     executionPlan: {
       ...material,
-      attempt: { ...material.attempt, number: 0 },
+      runCell: {
+        ...material.runCell,
+        attempt: 0,
+        digest: `sha256:${'0'.repeat(64)}`,
+      },
       model: {
         ...material.model,
         access: {
diff --git a/src/candidate-execution/execute.ts b/src/candidate-execution/execute.ts
index f05874eb..f9935288 100644
--- a/src/candidate-execution/execute.ts
+++ b/src/candidate-execution/execute.ts
@@ -1,5 +1,8 @@
 import type { TraceStore } from '@tangle-network/agent-eval'
-import type { AgentCandidateTermination } from '@tangle-network/agent-interface'
+import type {
+  AgentCandidateTaskOutcomeSpec,
+  AgentCandidateTermination,
+} from '@tangle-network/agent-interface'
 
 import type {
   AgentCandidateExecutionClaim,
@@ -300,6 +303,7 @@ export async function executePreparedAgentCandidate(
   const execution = await runAndStopExecutor(
     options.executor,
     request,
+    state.benchmarkTask.outcome,
     protectedTraceStore,
     deadlineAtMs,
     cleanupTimeoutMs,
@@ -544,6 +548,7 @@ type ExecutorOutcome =
 async function runAndStopExecutor(
   executor: AgentCandidateExecutorPort,
   request: AgentCandidateExecutorRequest,
+  expectedOutcome: AgentCandidateTaskOutcomeSpec,
   traceStore: TraceStore,
   deadlineAtMs: number,
   cleanupTimeoutMs: number,
@@ -629,17 +634,24 @@ async function runAndStopExecutor(
     )
     sealAgentCandidateExecutorStopAcknowledgement(stopped)
   } catch (stopError) {
+    let disposalError: unknown
+    try {
+      await disposeExecutor(executor, request, cleanupDeadlineAtMs)
+    } catch (error) {
+      disposalError = error
+    }
     if (timer) clearTimeout(timer)
     controller.abort(stopError)
     return {
       kind: 'error',
-      error: new Error(joinErrors(executionError, stopError)),
+      error: new Error(joinErrors(executionError, stopError, disposalError)),
       ...(timedOut ? { termination: { kind: 'timeout', timeoutMs } } : {}),
       processStopped: false,
     }
   }
 
-  let finalCapture: SealedAgentCandidateExecutorFinalCapture
+  let finalCapture: SealedAgentCandidateExecutorFinalCapture | undefined
+  let finalCaptureError: unknown
   try {
     finalCapture = sealAgentCandidateExecutorFinalCapture(
       await withinCandidateCleanupDeadline(
@@ -657,14 +669,24 @@ async function runAndStopExecutor(
         cleanupDeadlineAtMs,
         'candidate final evidence capture',
       ),
-      request.executionPlan.value.material.task.outcome,
+      expectedOutcome,
     )
   } catch (captureError) {
+    finalCaptureError = captureError
+  }
+
+  let disposalError: unknown
+  try {
+    await disposeExecutor(executor, request, cleanupDeadlineAtMs)
+  } catch (disposeError) {
+    disposalError = disposeError
+  }
+  if (finalCaptureError || disposalError || !finalCapture) {
     if (timer) clearTimeout(timer)
-    controller.abort(captureError)
+    controller.abort(finalCaptureError ?? disposalError)
     return {
       kind: 'error',
-      error: new Error(joinErrors(executionError, captureError)),
+      error: new Error(joinErrors(executionError, finalCaptureError, disposalError)),
       ...(timedOut ? { termination: { kind: 'timeout', timeoutMs } } : {}),
       processStopped: true,
     }
@@ -703,6 +725,35 @@ async function runAndStopExecutor(
   }
 }
 
+async function disposeExecutor(
+  executor: AgentCandidateExecutorPort,
+  request: AgentCandidateExecutorRequest,
+  cleanupDeadlineAtMs: number,
+): Promise<void> {
+  const dispose = executor.dispose
+  if (!dispose) return
+  const disposed = await withinCandidateCleanupDeadline(
+    (cleanupSignal) =>
+      dispose.call(
+        executor,
+        {
+          executionId: request.executionId,
+          executionPlanDigest: request.executionPlan.value.digest,
+        },
+        { signal: cleanupSignal },
+      ),
+    cleanupDeadlineAtMs,
+    'candidate execution resource disposal',
+  )
+  if (
+    !disposed ||
+    disposed.disposed !== true ||
+    Object.keys(disposed).some((key) => key !== 'disposed')
+  ) {
+    throw new Error('candidate executor did not acknowledge resource disposal')
+  }
+}
+
 async function failBeforeActivation(
   prepared: PreparedAgentCandidateExecution,
   state: PreparedCandidateState,
diff --git a/src/candidate-execution/executor-capture-evidence.ts b/src/candidate-execution/executor-capture-evidence.ts
index 1e899c98..26c8dd91 100644
--- a/src/candidate-execution/executor-capture-evidence.ts
+++ b/src/candidate-execution/executor-capture-evidence.ts
@@ -47,7 +47,7 @@ export async function persistAgentCandidateExecutorCapture(
   const executorEvidence = capture.evidence
     ? await persistCandidateOutputArtifact(outputArtifacts, {
         executionId: identity.executionId,
-        purpose: 'executor-capture',
+        purpose: 'executor-native-evidence',
         bytes: capture.evidence,
         signal,
       })
diff --git a/src/candidate-execution/finalize.ts b/src/candidate-execution/finalize.ts
index 4f23b3ef..f517dcaa 100644
--- a/src/candidate-execution/finalize.ts
+++ b/src/candidate-execution/finalize.ts
@@ -156,8 +156,14 @@ export async function finalizeAgentCandidateRun(
       kind: 'agent-candidate-run',
       digestAlgorithm: 'rfc8785-sha256',
       bundleDigest: state.bundle.digest,
+      runCellDigest: state.executionPlan.value.material.runCell.digest,
       materializationReceiptDigest: state.materializationReceipt.digest,
       executionPlanDigest: state.executionPlan.value.digest,
+      timing: {
+        startedAtMs: run.startedAt,
+        endedAtMs: run.endedAt,
+        durationMs: run.endedAt - run.startedAt,
+      },
       memory,
       trace,
       termination,
diff --git a/src/candidate-execution/index.ts b/src/candidate-execution/index.ts
index 7ff08ba1..a75b7185 100644
--- a/src/candidate-execution/index.ts
+++ b/src/candidate-execution/index.ts
@@ -28,6 +28,7 @@ export {
   type AgentCandidateRetryRejection,
   InMemoryAgentCandidateExecutionClaimStore,
 } from './claim'
+export type { AgentCandidatePreparationEvidence } from './claim-file-formats'
 export {
   FileAgentCandidateExecutionClaimStore,
   type FileAgentCandidateExecutionClaimStoreOptions,
@@ -78,6 +79,7 @@ export {
   type AgentCandidateExecutorFinalCapture,
   type AgentCandidateExecutorMemoryCapture,
   type AgentCandidateExecutorPort,
+  type AgentCandidateExecutorProfileFile,
   type AgentCandidateExecutorRequest,
   type AgentCandidateExecutorStopRequest,
   type AgentCandidateExecutorTaskOutcomeCapture,
diff --git a/src/candidate-execution/outcome-evidence.ts b/src/candidate-execution/outcome-evidence.ts
index d7178686..d9066f88 100644
--- a/src/candidate-execution/outcome-evidence.ts
+++ b/src/candidate-execution/outcome-evidence.ts
@@ -138,7 +138,8 @@ export async function persistVerifiedAgentCandidateExecutorCapture(
   signal?.throwIfAborted()
   const taskCapture = capture.taskOutcome
   if (!taskCapture) throw new Error('candidate final capture is missing the task outcome')
-  const expected = state.executionPlan.value.material.task.outcome
+  if (capture.evidence) assertNoProtectedBytes(capture.evidence, protectedValues)
+  const expected = state.benchmarkTask.outcome
   if (taskCapture.kind !== expected.kind) {
     throw new Error(
       `candidate captured ${taskCapture.kind} outcome does not match expected ${expected.kind} outcome`,
@@ -203,7 +204,7 @@ async function verifyWorkspaceTaskCapture(
   protectedValues: readonly string[],
   signal?: AbortSignal,
 ): Promise<VerifiedWorkspaceTaskCapture> {
-  const repository = state.executionPlan.value.material.task.repository
+  const repository = state.benchmarkTask.repository
   if (!repository) throw new Error('workspace task outcome is missing repository identity')
   const patch = Uint8Array.from(capture.gitDiff)
   const archive = Uint8Array.from(capture.archive)
@@ -401,6 +402,7 @@ export async function persistCandidateBenchmarkResult(
     executionId: state.executionId,
     termination: frozenTermination,
     outcome,
+    expectedGrader: state.benchmarkTask.grader,
     grader,
     artifacts: outputArtifacts,
     signal,
@@ -416,23 +418,13 @@ export async function persistCandidateBenchmarkResult(
     bytes: rawEvidence,
     signal,
   })
-  const task = state.executionPlan.value.material.task
   const material = {
     kind: 'agent-candidate-benchmark-result-material' as const,
     executionPlanDigest: state.executionPlan.value.digest,
     taskOutcomeDigest: outcome.evidence.digest,
-    benchmark: {
-      name: task.benchmark,
-      version: task.benchmarkVersion,
-      taskId: task.taskId,
-      splitDigest: task.splitDigest,
-    },
-    grader: {
-      name: graded.grader.name,
-      version: graded.grader.version,
-      artifact: graded.grader.artifact,
-    },
+    grader: graded.grader,
     evidence: evidenceRef,
+    grading: graded.grading,
     score: evaluation.score,
     passed: evaluation.passed,
     dimensions: evaluation.dimensions,
diff --git a/src/candidate-execution/prepare.ts b/src/candidate-execution/prepare.ts
index cf142437..1b9bc081 100644
--- a/src/candidate-execution/prepare.ts
+++ b/src/candidate-execution/prepare.ts
@@ -13,12 +13,15 @@ import type {
   AgentCandidateResolvedModel,
 } from '@tangle-network/agent-interface'
 import {
+  agentCandidateBenchmarkSuiteSchema,
+  agentCandidateBenchmarkTaskSchema,
   agentCandidateContainerSchema,
   agentCandidateExecutionLimitsSchema,
   agentCandidateExecutionPlanEvidenceSchema,
   agentCandidateExecutionPlanMaterialSchema,
   agentCandidateMaterializationReceiptSchema,
   agentCandidateModelAccessNetworkSchema,
+  agentCandidateRunCellSchema,
   agentCandidateTaskOutcomeSpecSchema,
   agentCandidateWorkspaceSnapshotEvidenceSchema,
   sha256DigestSchema,
@@ -47,6 +50,7 @@ import {
   canonicalCandidateDigest,
   canonicalCandidateDocument,
   embeddedCandidateArtifact,
+  omitTopLevelDigest,
   sha256Bytes,
 } from './digest'
 import { candidateExecutionOwnerWindowMs } from './execution-window'
@@ -59,7 +63,7 @@ import {
 } from './knowledge'
 import { sealAgentCandidateModelSettlement, usdToNanos } from './model-settlement'
 import { createPreparedCandidateExecution } from './prepared-state'
-import { candidateMaterializerHarness, createAgentCandidateProfileActivation } from './profile'
+import { candidateMaterializerHarness } from './profile'
 import {
   type AgentCandidateExecutionPorts,
   type AgentCandidateTaskExecution,
@@ -95,11 +99,23 @@ export async function prepareAgentCandidateExecution(
   const verifiedState = getVerifiedCandidateState(candidate)
   assertSameVerificationPorts(verifiedState.ports, ports)
   const bundle = candidate.bundle
+  if (task.runCell.bundleDigest !== bundle.digest) {
+    throw new Error('candidate experiment cell does not bind the verified bundle')
+  }
+  const benchmarkTask = task.task
+  const attempt = {
+    number: task.runCell.attempt,
+    maxAttempts: benchmarkTask.attempt.maxAttempts,
+    retryPolicy: benchmarkTask.attempt.retryPolicy,
+  } as const
   const harness = candidateMaterializerHarness(bundle.execution.harness)
   assertTaskInput(task, bundle.execution.instructionDelivery)
-  const resultTimeoutMs = candidateResultTimeout(options.resultTimeoutMs, task.limits.timeoutMs)
+  const resultTimeoutMs = candidateResultTimeout(
+    options.resultTimeoutMs,
+    benchmarkTask.limits.timeoutMs,
+  )
   const ownerWindowMs = candidateExecutionOwnerWindowMs(
-    task.limits.timeoutMs,
+    benchmarkTask.limits.timeoutMs,
     cleanupTimeoutMs,
     resultTimeoutMs,
   )
@@ -109,25 +125,27 @@ export async function prepareAgentCandidateExecution(
   }
   assertDisjointHostStagingRoots(task)
 
-  const instructionBytes = Buffer.from(task.instruction, 'utf8')
-  const instructionDigest = sha256Bytes(instructionBytes)
+  const instructionBytes = Buffer.from(benchmarkTask.instruction, 'utf8')
 
-  const taskArtifacts = await verifyWorkspaceSnapshotArtifacts(task.workspace, ports.artifacts)
+  const taskArtifacts = await verifyWorkspaceSnapshotArtifacts(
+    benchmarkTask.workspace,
+    ports.artifacts,
+  )
   await ports.workspaces.materialize({
     role: 'task',
-    snapshot: task.workspace,
+    snapshot: benchmarkTask.workspace,
     archive: taskArtifacts.archive,
     destination: task.stagingRoots.taskRoot,
   })
-  await verifyMaterializedWorkspace(task.stagingRoots.taskRoot, task.workspace.material, {
+  await verifyMaterializedWorkspace(task.stagingRoots.taskRoot, benchmarkTask.workspace.material, {
     ignoredProtectedRootEntries: ['.git', '.sidecar'],
   })
-  if (task.repository) {
-    await verifyTaskCheckout(task.stagingRoots.taskRoot, task.repository)
+  if (benchmarkTask.repository) {
+    await verifyTaskCheckout(task.stagingRoots.taskRoot, benchmarkTask.repository)
   }
   const taskExecutorFiles = await readMaterializedWorkspaceFiles(
     task.stagingRoots.taskRoot,
-    task.workspace.material,
+    benchmarkTask.workspace.material,
     { ignoredProtectedRootEntries: ['.git', '.sidecar'] },
   )
 
@@ -169,23 +187,22 @@ export async function prepareAgentCandidateExecution(
   )
   await verifyMaterializedProfileWorkspace(
     task.stagingRoots.profileRoot,
-    profileApplication.profilePlan.material,
+    profileApplication.profileActivation.profilePlan.material,
   )
   const profilePlanBytes = await readVerifiedArtifact(
-    profileApplication.profilePlan.artifact,
+    profileApplication.profileActivation.profilePlan.artifact,
     ports.artifacts,
   )
   if (
     !Buffer.from(profilePlanBytes).equals(
-      Buffer.from(canonicalCandidateBytes(profileApplication.profilePlan.material)),
+      Buffer.from(
+        canonicalCandidateBytes(profileApplication.profileActivation.profilePlan.material),
+      ),
     )
   ) {
     throw new Error('profile materializer did not capture exact canonical plan bytes')
   }
-  const profileActivation = createAgentCandidateProfileActivation(
-    profileWorkspacePlan,
-    profileApplication.profilePlan,
-  )
+  const profileActivation = profileApplication.profileActivation
 
   const container = await resolveContainer(candidate, task, ports)
   const resolvedModel = await resolveModel(candidate, task, ports)
@@ -197,10 +214,10 @@ export async function prepareAgentCandidateExecution(
         executionId: task.executionId,
         preparationId,
         expiresAtMs: reservationExpiresAtMs,
-        attempt: task.attempt,
+        attempt,
         bundleDigest: bundle.digest,
         resolved: resolvedModel,
-        limits: modelLimits(task.limits),
+        limits: modelLimits(benchmarkTask.limits),
       }),
     candidateCleanupDeadline(cleanupTimeoutMs),
     'protected model reservation',
@@ -209,7 +226,7 @@ export async function prepareAgentCandidateExecution(
   try {
     validateProtectedModelReservation(
       modelReservation,
-      task.limits,
+      benchmarkTask.limits,
       preparationId,
       reservationExpiresAtMs,
     )
@@ -258,27 +275,11 @@ export async function prepareAgentCandidateExecution(
           }
         : {},
     )
-    const routes = modelRoutes(bundle.profile, task.model.requested)
+    const routes = modelRoutes(bundle.profile, benchmarkTask.model.requested)
     const executionMaterial: AgentCandidateExecutionPlanMaterial = {
       kind: 'agent-candidate-execution-plan-material',
-      bundleDigest: bundle.digest,
+      runCell: task.runCell,
       executionId: task.executionId,
-      attempt: task.attempt,
-      task: {
-        benchmark: task.benchmark,
-        benchmarkVersion: task.benchmarkVersion,
-        taskId: task.taskId,
-        splitDigest: task.splitDigest,
-        instruction: {
-          encoding: 'utf8',
-          sha256: instructionDigest,
-          byteLength: instructionBytes.byteLength,
-          delivery: bundle.execution.instructionDelivery,
-        },
-        ...(task.repository ? { repository: task.repository } : {}),
-        outcome: task.outcome,
-        workspace: task.workspace,
-      },
       workspaces: {
         taskRoot: task.executionRoots.taskRoot,
         ...(task.executionRoots.candidateRoot
@@ -290,6 +291,8 @@ export async function prepareAgentCandidateExecution(
       profile: profileApplication.application,
       harness: bundle.execution.harness,
       harnessVersion: bundle.execution.harnessVersion,
+      instructionDelivery: bundle.execution.instructionDelivery,
+      limits: benchmarkTask.limits,
       container,
       model: {
         policy: 'single',
@@ -301,7 +304,6 @@ export async function prepareAgentCandidateExecution(
         },
         routes,
       },
-      grader: task.grader,
       launch: {
         executable: baseLaunch.executable,
         args: baseLaunch.args,
@@ -310,7 +312,6 @@ export async function prepareAgentCandidateExecution(
       },
       ...(bundle.knowledge ? { knowledgeManifestDigest: bundle.knowledge.snapshot.digest } : {}),
       memory,
-      limits: task.limits,
       network: { mode: 'disabled' },
     }
     agentCandidateExecutionPlanMaterialSchema.parse(executionMaterial)
@@ -333,7 +334,21 @@ export async function prepareAgentCandidateExecution(
         kind: 'agent-candidate-materialization',
         digestAlgorithm: 'rfc8785-sha256',
         bundleDigest: bundle.digest,
-        profilePlan: profileApplication.profilePlan,
+        benchmark: {
+          suite: {
+            digest: task.benchmarkSuite.digest,
+            material: embeddedCandidateArtifact(
+              canonicalCandidateBytes(omitTopLevelDigest(task.benchmarkSuite)),
+            ),
+          },
+          task: {
+            digest: benchmarkTask.digest,
+            material: embeddedCandidateArtifact(
+              canonicalCandidateBytes(omitTopLevelDigest(benchmarkTask)),
+            ),
+          },
+        },
+        profileActivation,
         executionPlan,
         ...(bundle.execution.workspace ? { candidateWorkspace: bundle.execution.workspace } : {}),
         codeKind: bundle.code.kind,
@@ -348,7 +363,7 @@ export async function prepareAgentCandidateExecution(
     )
     agentCandidateMaterializationReceiptSchema.parse(materializationReceipt.value)
 
-    const traceRunId = `${task.executionId}:attempt-${task.attempt.number}:${canonicalCandidateDigest({ preparationId }).slice(7, 23)}`
+    const traceRunId = `${task.executionId}:attempt-${attempt.number}:${canonicalCandidateDigest({ preparationId }).slice(7, 23)}`
     const traceTags = {
       [CANDIDATE_TRACE_TAGS.executionId]: task.executionId,
       [CANDIDATE_TRACE_TAGS.bundleDigest]: bundle.digest,
@@ -367,13 +382,15 @@ export async function prepareAgentCandidateExecution(
     return createPreparedCandidateExecution({
       ports,
       bundle,
+      benchmarkSuite: task.benchmarkSuite,
+      benchmarkTask,
       executionId: task.executionId,
       roots: {
         execution: { ...task.executionRoots },
         staging: { ...task.stagingRoots },
       },
       profilePlan: {
-        value: profileApplication.profilePlan,
+        value: profileApplication.profileActivation.profilePlan,
         bytes: profilePlanBytes,
         written: [...profileApplication.application.mountPaths],
       },
@@ -408,7 +425,7 @@ export async function prepareAgentCandidateExecution(
         ...(candidateExecutorFiles ? { candidateFiles: candidateExecutorFiles } : {}),
         profileFiles: exactProfileExecutorFiles(
           profileWorkspacePlan.files,
-          profileApplication.profilePlan.material.files,
+          profileApplication.profileActivation.profilePlan.material.files,
         ),
       },
       ...(preparedMemory.accessDigest && preparedMemory.value.mode === 'isolated'
@@ -508,19 +525,22 @@ function assertTaskInput(
   task: AgentCandidateTaskExecution,
   delivery: VerifiedAgentCandidate['bundle']['execution']['instructionDelivery'],
 ): void {
+  const benchmarkTask = agentCandidateBenchmarkTaskSchema.parse(task.task)
+  const benchmarkSuite = agentCandidateBenchmarkSuiteSchema.parse(task.benchmarkSuite)
+  const runCell = agentCandidateRunCellSchema.parse(task.runCell)
   const requiredStrings: Array<[string, string]> = [
     ['executionId', task.executionId],
-    ['benchmark', task.benchmark],
-    ['benchmarkVersion', task.benchmarkVersion],
-    ['taskId', task.taskId],
+    ['benchmark', benchmarkTask.benchmark.name],
+    ['benchmarkVersion', benchmarkTask.benchmark.version],
+    ['taskId', benchmarkTask.scenario.id],
   ]
-  if (task.outcome.kind === 'workspace' && !task.repository) {
+  if (benchmarkTask.outcome.kind === 'workspace' && !benchmarkTask.repository) {
     throw new Error('workspace task outcome requires repository identity')
   }
-  if (task.repository) {
+  if (benchmarkTask.repository) {
     requiredStrings.push(
-      ['repository identity', task.repository.identity],
-      ['repository root identity', task.repository.rootIdentity],
+      ['repository identity', benchmarkTask.repository.identity],
+      ['repository root identity', benchmarkTask.repository.rootIdentity],
     )
   }
   for (const [name, value] of requiredStrings) {
@@ -529,21 +549,39 @@ function assertTaskInput(
   if (!/^[A-Za-z0-9._:-]{1,200}$/.test(task.executionId)) {
     throw new Error('executionId must be a stable filesystem-neutral identifier')
   }
-  if (!task.instruction || !isWellFormedUnicode(task.instruction)) {
+  if (!benchmarkTask.instruction || !isWellFormedUnicode(benchmarkTask.instruction)) {
     throw new Error('task instruction must be non-empty well-formed Unicode')
   }
-  sha256DigestSchema.parse(task.splitDigest)
-  agentCandidateTaskOutcomeSpecSchema.parse(task.outcome)
-  agentCandidateWorkspaceSnapshotEvidenceSchema.parse(task.workspace)
-  agentCandidateExecutionLimitsSchema.parse(task.limits)
-  if (task.repository) {
-    if (!/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(task.repository.baseCommit)) {
+  if (
+    canonicalCandidateDigest(omitTopLevelDigest(benchmarkTask)) !== benchmarkTask.digest ||
+    canonicalCandidateDigest(omitTopLevelDigest(benchmarkSuite)) !== benchmarkSuite.digest ||
+    canonicalCandidateDigest(omitTopLevelDigest(runCell)) !== runCell.digest
+  ) {
+    throw new Error('candidate experiment cell contains an invalid content digest')
+  }
+  const cellIndex = runCell.taskIndex * benchmarkSuite.reps + runCell.repetition
+  if (
+    runCell.suiteDigest !== benchmarkSuite.digest ||
+    runCell.taskDigest !== benchmarkTask.digest ||
+    benchmarkSuite.taskDigests[runCell.taskIndex] !== benchmarkTask.digest ||
+    runCell.repetition >= benchmarkSuite.reps ||
+    benchmarkSuite.seeds[cellIndex] !== runCell.seed ||
+    runCell.attempt > benchmarkTask.attempt.maxAttempts
+  ) {
+    throw new Error('candidate experiment cell does not match its signed suite and task')
+  }
+  sha256DigestSchema.parse(benchmarkTask.benchmark.splitDigest)
+  agentCandidateTaskOutcomeSpecSchema.parse(benchmarkTask.outcome)
+  agentCandidateWorkspaceSnapshotEvidenceSchema.parse(benchmarkTask.workspace)
+  agentCandidateExecutionLimitsSchema.parse(benchmarkTask.limits)
+  if (benchmarkTask.repository) {
+    if (!/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(benchmarkTask.repository.baseCommit)) {
       throw new Error('task repository base commit is not a full Git object id')
     }
-    if (!/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(task.repository.baseTree)) {
+    if (!/^(?:[a-f0-9]{40}|[a-f0-9]{64})$/.test(benchmarkTask.repository.baseTree)) {
       throw new Error('task repository base tree is not a full Git object id')
     }
-    if (task.repository.baseCommit.length !== task.repository.baseTree.length) {
+    if (benchmarkTask.repository.baseCommit.length !== benchmarkTask.repository.baseTree.length) {
       throw new Error('task repository Git object formats disagree')
     }
   }
@@ -561,15 +599,15 @@ function assertTaskInput(
     if (!canonical) throw new Error(`${name} must be a canonical absolute path`)
   }
   if (
-    !Number.isInteger(task.attempt.number) ||
-    !Number.isInteger(task.attempt.maxAttempts) ||
-    task.attempt.number < 1 ||
-    task.attempt.number > task.attempt.maxAttempts ||
-    (task.attempt.retryPolicy === 'none' && task.attempt.maxAttempts !== 1)
+    !Number.isInteger(runCell.attempt) ||
+    !Number.isInteger(benchmarkTask.attempt.maxAttempts) ||
+    runCell.attempt < 1 ||
+    runCell.attempt > benchmarkTask.attempt.maxAttempts ||
+    (benchmarkTask.attempt.retryPolicy === 'none' && benchmarkTask.attempt.maxAttempts !== 1)
   ) {
     throw new Error('task attempt policy is invalid')
   }
-  const limits = task.limits
+  const limits = benchmarkTask.limits
   if (
     !Number.isInteger(limits.timeoutMs) ||
     limits.timeoutMs <= 0 ||
@@ -588,28 +626,33 @@ function assertTaskInput(
     throw new Error('task execution limits are invalid')
   }
   usdToNanos(limits.maxCostUsd, 'task maxCostUsd')
-  if (!task.model.requested.trim()) throw new Error('evaluator model request must be non-empty')
-  if (!task.grader.name.trim() || !task.grader.version.trim()) {
+  if (!benchmarkTask.model.requested.trim()) {
+    throw new Error('evaluator model request must be non-empty')
+  }
+  if (!benchmarkTask.grader.name.trim() || !benchmarkTask.grader.version.trim()) {
     throw new Error('evaluator benchmark grader identity must be non-empty')
   }
-  if (!Number.isInteger(task.grader.artifact.byteLength) || task.grader.artifact.byteLength <= 0) {
+  if (
+    !Number.isInteger(benchmarkTask.grader.artifact.byteLength) ||
+    benchmarkTask.grader.artifact.byteLength <= 0
+  ) {
     throw new Error('evaluator benchmark grader artifact must be non-empty')
   }
-  sha256DigestSchema.parse(task.grader.artifact.sha256)
-  if (task.evaluatorTaskContainer) {
+  sha256DigestSchema.parse(benchmarkTask.grader.artifact.sha256)
+  if (benchmarkTask.evaluatorTaskContainer) {
     if (
-      task.evaluatorTaskContainer.source !== 'evaluator-task-container' ||
-      !task.evaluatorTaskContainer.image.trim() ||
-      !task.evaluatorTaskContainer.platform.os.trim() ||
-      !task.evaluatorTaskContainer.platform.architecture.trim()
+      benchmarkTask.evaluatorTaskContainer.source !== 'evaluator-task-container' ||
+      !benchmarkTask.evaluatorTaskContainer.image.trim() ||
+      !benchmarkTask.evaluatorTaskContainer.platform.os.trim() ||
+      !benchmarkTask.evaluatorTaskContainer.platform.architecture.trim()
     ) {
       throw new Error('evaluator task container evidence is incomplete')
     }
-    sha256DigestSchema.parse(task.evaluatorTaskContainer.indexDigest)
-    sha256DigestSchema.parse(task.evaluatorTaskContainer.manifestDigest)
+    sha256DigestSchema.parse(benchmarkTask.evaluatorTaskContainer.indexDigest)
+    sha256DigestSchema.parse(benchmarkTask.evaluatorTaskContainer.manifestDigest)
     agentCandidateContainerSchema.parse({
-      image: task.evaluatorTaskContainer.image,
-      indexDigest: task.evaluatorTaskContainer.indexDigest,
+      image: benchmarkTask.evaluatorTaskContainer.image,
+      indexDigest: benchmarkTask.evaluatorTaskContainer.indexDigest,
     })
   }
   if (delivery.kind === 'utf8-file') {
@@ -671,24 +714,25 @@ async function resolveContainer(
   ports: AgentCandidateExecutionPorts,
 ): Promise<ResolvedAgentCandidateContainer> {
   const environment = candidate.bundle.execution.environment
+  const evaluatorTaskContainer = task.task.evaluatorTaskContainer
   const pinned = environment.kind === 'pinned-container' ? environment.container : undefined
-  if (environment.kind === 'evaluator-task-container' && !task.evaluatorTaskContainer) {
+  if (environment.kind === 'evaluator-task-container' && !evaluatorTaskContainer) {
     throw new Error('evaluator-task-container candidate requires an evaluator-owned task image')
   }
-  if (environment.kind === 'pinned-container' && task.evaluatorTaskContainer) {
+  if (environment.kind === 'pinned-container' && evaluatorTaskContainer) {
     throw new Error('pinned candidate containers cannot be replaced by a task image')
   }
   const resolved = await ports.containers.resolve({
     candidate: pinned,
-    evaluatorTaskContainer: task.evaluatorTaskContainer,
+    evaluatorTaskContainer,
   })
   if (resolved.source !== environment.kind) throw new Error('resolved container source drifted')
   if (pinned && (resolved.image !== pinned.image || resolved.indexDigest !== pinned.indexDigest)) {
     throw new Error('resolved pinned container does not match the candidate image index')
   }
   if (
-    task.evaluatorTaskContainer &&
-    JSON.stringify(resolved) !== JSON.stringify(task.evaluatorTaskContainer)
+    evaluatorTaskContainer &&
+    canonicalCandidateDigest(resolved) !== canonicalCandidateDigest(evaluatorTaskContainer)
   ) {
     throw new Error('resolved task container does not match evaluator-owned image evidence')
   }
@@ -701,25 +745,20 @@ async function resolveModel(
   ports: AgentCandidateExecutionPorts,
 ): Promise<AgentCandidateResolvedModel> {
   const hints = candidate.bundle.profile.model
-  if (hints?.default !== undefined && hints.default !== task.model.requested) {
+  const expected = task.task.model
+  if (hints?.default !== undefined && hints.default !== expected.requested) {
     throw new Error('candidate model preference conflicts with the evaluator-owned model')
   }
-  if (
-    hints?.reasoningEffort !== undefined &&
-    hints.reasoningEffort !== task.model.reasoningEffort
-  ) {
+  if (hints?.reasoningEffort !== undefined && hints.reasoningEffort !== expected.reasoningEffort) {
     throw new Error('candidate reasoning effort conflicts with the evaluator-owned effort')
   }
   const resolved = await ports.models.resolve({
-    requested: task.model.requested,
+    requested: expected.requested,
     harness: candidate.bundle.execution.harness,
-    reasoningEffort: task.model.reasoningEffort,
+    reasoningEffort: expected.reasoningEffort,
   })
-  if (
-    resolved.requested !== task.model.requested ||
-    resolved.reasoningEffort !== task.model.reasoningEffort
-  ) {
-    throw new Error('model resolver drifted from the evaluator-owned request')
+  if (canonicalCandidateDigest(resolved) !== canonicalCandidateDigest(expected)) {
+    throw new Error('model resolver drifted from the signed benchmark model')
   }
   return resolved
 }
@@ -739,7 +778,7 @@ async function prepareMemory(
   if (policy.mode === 'disabled') return { value: { mode: 'disabled' } }
   const seed = policy.seed ? await verifiedArtifactBytes(candidate, policy.seed) : undefined
   const executionSegment = canonicalCandidateDigest({ executionId: task.executionId }).slice(7)
-  const taskSegment = canonicalCandidateDigest({ taskId: task.taskId }).slice(7)
+  const taskSegment = canonicalCandidateDigest({ taskDigest: task.task.digest }).slice(7)
   const preparationSegment = canonicalCandidateDigest({ preparationId }).slice(7, 23)
   const effectiveNamespace = `candidate/${candidate.bundle.digest.slice(7, 23)}/${executionSegment}/${taskSegment}/${preparationSegment}`
   const reset = await withinCandidateCleanupDeadline(
diff --git a/src/candidate-execution/prepared-state.ts b/src/candidate-execution/prepared-state.ts
index 67211313..aa63aef4 100644
--- a/src/candidate-execution/prepared-state.ts
+++ b/src/candidate-execution/prepared-state.ts
@@ -1,4 +1,9 @@
 import {
+  type AgentCandidateBenchmarkSuite,
+  type AgentCandidateBenchmarkTask,
+  type AgentCandidateWorkspaceManifestMaterial,
+  agentCandidateBenchmarkSuiteSchema,
+  agentCandidateBenchmarkTaskSchema,
   agentCandidateExecutionPlanEvidenceSchema,
   agentCandidateMaterializationReceiptSchema,
   agentCandidateProfilePlanEvidenceSchema,
@@ -28,6 +33,8 @@ import { CANDIDATE_TRACE_ENV, CANDIDATE_TRACE_TAGS, preparedCandidateBrand } fro
 export interface PreparedCandidateState {
   ports: AgentCandidateExecutionPorts
   bundle: PreparedAgentCandidateExecution['bundle']
+  benchmarkSuite: AgentCandidateBenchmarkSuite
+  benchmarkTask: AgentCandidateBenchmarkTask
   executionId: string
   roots: PreparedAgentCandidateExecution['roots']
   profilePlan: PreparedAgentCandidateExecution['profilePlan']
@@ -96,6 +103,7 @@ export function createPreparedCandidateExecution(
   assertPrivateCandidateIntegrity(state)
   const prepared = Object.freeze({
     bundle: state.bundle,
+    benchmark: Object.freeze({ suite: state.benchmarkSuite, task: state.benchmarkTask }),
     executionId: state.executionId,
     roots: state.roots,
     profilePlan: evidenceView(state.profilePlan),
@@ -174,8 +182,9 @@ export function beginPreparedCandidateRun(
   const material = state.executionPlan.value.material
   const request = Object.freeze({
     executionId: state.executionId,
+    benchmark: preparedBenchmark(state),
     inputs: Object.freeze({
-      task: workspaceInputView(material.task.workspace, state.executorInputs.taskFiles),
+      task: workspaceInputView(state.benchmarkTask.workspace, state.executorInputs.taskFiles),
       ...(material.candidateWorkspace && state.executorInputs.candidateFiles
         ? {
             candidate: workspaceInputView(
@@ -207,6 +216,12 @@ export function beginPreparedCandidateRun(
   return { state, request }
 }
 
+function preparedBenchmark(
+  state: PreparedCandidateState,
+): PreparedAgentCandidateExecution['benchmark'] {
+  return Object.freeze({ suite: state.benchmarkSuite, task: state.benchmarkTask })
+}
+
 export function consumePreparedCandidateExecution(
   prepared: PreparedAgentCandidateExecution,
   outcome: 'succeeded' | 'failed' | 'disposed' | 'disposal-failed' | 'cleanup-failed',
@@ -227,11 +242,15 @@ export async function assertPreparedCandidateWorkspaces(
   state: PreparedCandidateState,
 ): Promise<void> {
   const plan = state.executionPlan.value.material
-  await verifyMaterializedWorkspace(state.roots.staging.taskRoot, plan.task.workspace.material, {
-    ignoredProtectedRootEntries: ['.git', '.sidecar'],
-  })
-  if (plan.task.repository) {
-    await verifyTaskCheckout(state.roots.staging.taskRoot, plan.task.repository)
+  await verifyMaterializedWorkspace(
+    state.roots.staging.taskRoot,
+    state.benchmarkTask.workspace.material,
+    {
+      ignoredProtectedRootEntries: ['.git', '.sidecar'],
+    },
+  )
+  if (state.benchmarkTask.repository) {
+    await verifyTaskCheckout(state.roots.staging.taskRoot, state.benchmarkTask.repository)
   }
   await verifyMaterializedProfileWorkspace(
     state.roots.staging.profileRoot,
@@ -247,6 +266,12 @@ export async function assertPreparedCandidateWorkspaces(
 }
 
 function detachPreparedCandidateState(input: PreparedCandidateState): PreparedCandidateState {
+  const benchmarkSuite = immutableCandidateValue(
+    agentCandidateBenchmarkSuiteSchema.parse(input.benchmarkSuite),
+  )
+  const benchmarkTask = immutableCandidateValue(
+    agentCandidateBenchmarkTaskSchema.parse(input.benchmarkTask),
+  )
   const profilePlan = immutableCandidateValue(
     agentCandidateProfilePlanEvidenceSchema.parse(input.profilePlan.value),
   )
@@ -269,6 +294,8 @@ function detachPreparedCandidateState(input: PreparedCandidateState): PreparedCa
   return Object.freeze({
     ports: input.ports,
     bundle: input.bundle,
+    benchmarkSuite,
+    benchmarkTask,
     executionId: input.executionId,
     roots: immutableCandidateValue(input.roots),
     profilePlan: Object.freeze({
@@ -341,11 +368,12 @@ function assertPrivateCandidateIntegrity(state: PreparedCandidateState): void {
     throw new Error('prepared materialization receipt no longer matches its canonical bytes')
   }
 
-  const instruction = state.executionPlan.value.material.task.instruction
+  assertSignedBenchmarkInput(state)
+  const instruction = Buffer.from(state.benchmarkTask.instruction, 'utf8')
   if (
-    sha256Bytes(state.instruction.bytes) !== instruction.sha256 ||
-    state.instruction.bytes.byteLength !== instruction.byteLength ||
-    JSON.stringify(state.instruction.delivery) !== JSON.stringify(instruction.delivery)
+    !Buffer.from(state.instruction.bytes).equals(instruction) ||
+    JSON.stringify(state.instruction.delivery) !==
+      JSON.stringify(state.executionPlan.value.material.instructionDelivery)
   ) {
     throw new Error('prepared instruction no longer matches the signed execution plan')
   }
@@ -405,7 +433,7 @@ function assertPrivateCandidateIntegrity(state: PreparedCandidateState): void {
   }
   if (
     !state.trace.runId.startsWith(
-      `${state.executionId}:attempt-${state.executionPlan.value.material.attempt.number}:`,
+      `${state.executionId}:attempt-${state.executionPlan.value.material.runCell.attempt}:`,
     ) ||
     canonicalCandidateDigest(state.trace.tags) !== canonicalCandidateDigest(expectedTags) ||
     canonicalCandidateDigest(state.trace.env) !== canonicalCandidateDigest(expectedTraceEnvironment)
@@ -511,7 +539,10 @@ function executorFileView(
 
 function assertExecutorInputs(state: PreparedCandidateState): void {
   const material = state.executionPlan.value.material
-  assertWorkspaceExecutorFiles(state.executorInputs.taskFiles, material.task.workspace.material)
+  assertWorkspaceExecutorFiles(
+    state.executorInputs.taskFiles,
+    state.benchmarkTask.workspace.material,
+  )
   if (material.candidateWorkspace) {
     if (!state.executorInputs.candidateFiles) {
       throw new Error('prepared candidate executor files are missing')
@@ -575,7 +606,7 @@ function assertPreparedKnowledge(state: PreparedCandidateState): void {
 
 function assertWorkspaceExecutorFiles(
   actualFiles: PreparedCandidateState['executorInputs']['taskFiles'],
-  expected: PreparedAgentCandidateExecution['executionPlan']['value']['material']['task']['workspace']['material'],
+  expected: AgentCandidateWorkspaceManifestMaterial,
 ): void {
   if (actualFiles.length !== expected.files.length) {
     throw new Error('prepared workspace executor files do not match the signed manifest')
@@ -596,6 +627,42 @@ function assertWorkspaceExecutorFiles(
   }
 }
 
+function assertSignedBenchmarkInput(state: PreparedCandidateState): void {
+  const plan = state.executionPlan.value.material
+  const cell = plan.runCell
+  const suite = state.benchmarkSuite
+  const task = state.benchmarkTask
+  const receipt = state.materializationReceipt.value
+  const cellIndex = cell.taskIndex * suite.reps + cell.repetition
+  if (
+    canonicalCandidateDigest(omitTopLevelDigest(suite)) !== suite.digest ||
+    canonicalCandidateDigest(omitTopLevelDigest(task)) !== task.digest ||
+    canonicalCandidateDigest(omitTopLevelDigest(cell)) !== cell.digest ||
+    cell.bundleDigest !== state.bundle.digest ||
+    cell.suiteDigest !== suite.digest ||
+    cell.taskDigest !== task.digest ||
+    suite.taskDigests[cell.taskIndex] !== task.digest ||
+    suite.seeds[cellIndex] !== cell.seed ||
+    cell.repetition >= suite.reps ||
+    cell.attempt < 1 ||
+    cell.attempt > task.attempt.maxAttempts
+  ) {
+    throw new Error('prepared candidate no longer matches its signed experiment cell')
+  }
+  const suiteBytes = canonicalCandidateBytes(omitTopLevelDigest(suite))
+  const taskBytes = canonicalCandidateBytes(omitTopLevelDigest(task))
+  if (
+    receipt.benchmark.suite.digest !== suite.digest ||
+    receipt.benchmark.suite.material.sha256 !== sha256Bytes(suiteBytes) ||
+    receipt.benchmark.suite.material.byteLength !== suiteBytes.byteLength ||
+    receipt.benchmark.task.digest !== task.digest ||
+    receipt.benchmark.task.material.sha256 !== sha256Bytes(taskBytes) ||
+    receipt.benchmark.task.material.byteLength !== taskBytes.byteLength
+  ) {
+    throw new Error('materialization receipt no longer matches its signed benchmark input')
+  }
+}
+
 function immutableExecutorFiles(
   files: PreparedCandidateState['executorInputs']['taskFiles'],
 ): ReadonlyArray<{ path: string; mode: number; bytes: Uint8Array }> {
diff --git a/src/candidate-execution/profile.ts b/src/candidate-execution/profile.ts
index a702202a..c4d187f0 100644
--- a/src/candidate-execution/profile.ts
+++ b/src/candidate-execution/profile.ts
@@ -96,10 +96,13 @@ export function parseAgentCandidateProfileActivation(
 ): AgentCandidateProfileActivation {
   const activation = agentCandidateProfileActivationSchema.parse(input)
   const planBytes = canonicalCandidateBytes(activation.profilePlan.material)
+  const profilePlanArtifact = activation.profilePlan.artifact
   if (
     sha256Bytes(planBytes) !== activation.profilePlan.digest ||
-    activation.profilePlan.artifact.sha256 !== activation.profilePlan.digest ||
-    activation.profilePlan.artifact.byteLength !== planBytes.byteLength ||
+    profilePlanArtifact.sha256 !== activation.profilePlan.digest ||
+    profilePlanArtifact.byteLength !== planBytes.byteLength ||
+    ('content' in profilePlanArtifact &&
+      !Buffer.from(profilePlanArtifact.content, 'base64').equals(Buffer.from(planBytes))) ||
     (expectedProfilePlanDigest !== undefined &&
       activation.profilePlan.digest !== expectedProfilePlanDigest)
   ) {
diff --git a/src/candidate-execution/recover.ts b/src/candidate-execution/recover.ts
index d3ce1f12..86e5576c 100644
--- a/src/candidate-execution/recover.ts
+++ b/src/candidate-execution/recover.ts
@@ -21,13 +21,10 @@ import {
   withinCandidateCleanupDeadline,
 } from './cleanup'
 import { canonicalCandidateBytes } from './digest'
-import {
-  sealAgentCandidateExecutorFinalCapture,
-  sealAgentCandidateExecutorStopAcknowledgement,
-} from './executor-capture'
-import { persistAgentCandidateExecutorCapture } from './executor-capture-evidence'
+import { sealAgentCandidateExecutorStopAcknowledgement } from './executor-capture'
 import { sealAgentCandidateModelSettlement } from './model-settlement'
 import { persistCandidateModelSettlementEvidence } from './outcome-evidence'
+import { persistCandidateOutputArtifact } from './output-artifacts'
 import { RecoveryAgentCandidateTraceStore } from './protected-trace-store'
 import type {
   AgentCandidateExecutionPorts,
@@ -90,6 +87,18 @@ export async function recoverExpiredAgentCandidateExecution(
         },
       )
       sealAgentCandidateExecutorStopAcknowledgement(stopped)
+      if (options.executor.dispose) {
+        const disposed = await options.executor.dispose(
+          {
+            executionId: record.claim.executionId,
+            executionPlanDigest: record.claim.executionPlanDigest,
+          },
+          { signal: cleanupSignal },
+        )
+        if (disposed.disposed !== true || Object.keys(disposed).some((key) => key !== 'disposed')) {
+          throw new Error('expired candidate executor did not acknowledge resource disposal')
+        }
+      }
       return { stopped: true as const }
     },
     cleanupDeadlineAtMs,
@@ -149,48 +158,6 @@ export async function recoverExpiredAgentCandidateExecution(
   const failures: unknown[] = [preparationOutcome, processOutcome]
     .filter((outcome): outcome is PromiseRejectedResult => outcome.status === 'rejected')
     .map((outcome) => outcome.reason)
-  const recoveredPreparation =
-    preparationOutcome.status === 'fulfilled' ? preparationOutcome.value : undefined
-  const processStopped = processOutcome.status === 'fulfilled'
-  let persistedCapture: Awaited<ReturnType<typeof persistAgentCandidateExecutorCapture>> | undefined
-  if (recoveredPreparation && processStopped) {
-    try {
-      const capture = sealAgentCandidateExecutorFinalCapture(
-        await withinCandidateCleanupDeadline(
-          (cleanupSignal) =>
-            options.executor.capture(
-              {
-                executionId: record.claim.executionId,
-                executionPlanDigest: record.claim.executionPlanDigest,
-              },
-              {
-                traceStore: recoveryTraceStore,
-                signal: cleanupSignal,
-              },
-            ),
-          cleanupDeadlineAtMs,
-          'expired candidate evidence capture',
-        ),
-        recoveredPreparation.plan.task.outcome,
-      )
-      persistedCapture = await withinCandidateCleanupDeadline(
-        () =>
-          persistAgentCandidateExecutorCapture(
-            {
-              executionId: record.claim.executionId,
-              executionPlanDigest: record.claim.executionPlanDigest,
-            },
-            recoveredPreparation.plan.task.outcome,
-            capture,
-            options.outputArtifacts,
-          ),
-        cleanupDeadlineAtMs,
-        'expired candidate capture persistence',
-      )
-    } catch (error) {
-      failures.push(error)
-    }
-  }
   const [modelOutcome, memoryOutcome] = await closureOutcomes
   for (const outcome of [modelOutcome, memoryOutcome]) {
     if (outcome.status === 'rejected') failures.push(outcome.reason)
@@ -198,7 +165,6 @@ export async function recoverExpiredAgentCandidateExecution(
   if (failures.length > 0) {
     throw new AggregateError(failures, 'expired candidate cleanup could not be proven')
   }
-  if (!persistedCapture) throw new Error('expired candidate capture was not persisted')
 
   if (modelOutcome.status !== 'fulfilled') {
     throw new Error('expired candidate model settlement was not recovered')
@@ -218,15 +184,38 @@ export async function recoverExpiredAgentCandidateExecution(
     cleanupDeadlineAtMs,
     'expired candidate model-settlement persistence',
   )
+  const failureClass =
+    record.phase === 'claimed' && model.usage.modelCalls === 0
+      ? 'pre-model-infrastructure'
+      : 'unknown'
+  const failureEvidence = await withinCandidateCleanupDeadline(
+    () =>
+      persistCandidateOutputArtifact(options.outputArtifacts, {
+        executionId: record.claim.executionId,
+        purpose: 'failure-evidence',
+        bytes: canonicalCandidateBytes({
+          kind: 'agent-candidate-recovery-failure',
+          executionId: record.claim.executionId,
+          attempt: record.claim.attempt,
+          bundleDigest: record.claim.bundleDigest,
+          executionPlanDigest: record.claim.executionPlanDigest,
+          materializationReceiptDigest:
+            record.claim.preparationEvidence.materializationReceipt.sha256,
+          failureClass,
+          processStopped: true,
+          modelSettlementDigest: modelSettlement.digest,
+          memoryClosed: record.claim.cleanup.memory !== undefined,
+        }),
+      }),
+    cleanupDeadlineAtMs,
+    'expired candidate failure-evidence persistence',
+  )
   const memory = record.claim.cleanup.memory
   return await options.claimStore.recoverExpired(options.attempt, {
-    failureClass:
-      record.phase === 'claimed' && model.usage.modelCalls === 0
-        ? 'pre-model-infrastructure'
-        : 'unknown',
+    failureClass,
     usage: model.usage,
     modelSettlement: modelSettlement.artifact,
-    failureEvidence: persistedCapture.evidence,
+    failureEvidence,
     process: {
       stopped: true,
       executionPlanDigest: record.claim.executionPlanDigest,
@@ -274,7 +263,7 @@ async function readRecoveryPreparation(
   if (
     receipt.bundleDigest !== claim.bundleDigest ||
     receipt.executionPlan.digest !== claim.executionPlanDigest ||
-    plan.bundleDigest !== claim.bundleDigest ||
+    plan.runCell.bundleDigest !== claim.bundleDigest ||
     plan.executionId !== claim.executionId
   ) {
     throw new Error('recovery preparation evidence does not match the durable claim')
diff --git a/src/candidate-execution/types.ts b/src/candidate-execution/types.ts
index ed251e31..2e98c7a1 100644
--- a/src/candidate-execution/types.ts
+++ b/src/candidate-execution/types.ts
@@ -2,6 +2,8 @@ import type { BenchmarkEvaluation, TraceStore } from '@tangle-network/agent-eval
 import type {
   AgentCandidateArtifactRef,
   AgentCandidateAttemptPolicy,
+  AgentCandidateBenchmarkSuite,
+  AgentCandidateBenchmarkTask,
   AgentCandidateBundle,
   AgentCandidateCapturedArtifact,
   AgentCandidateContainer,
@@ -20,16 +22,14 @@ import type {
   AgentCandidateProfileActivation,
   AgentCandidateProfilePlanEvidence,
   AgentCandidateResolvedModel,
+  AgentCandidateRunCell,
   AgentCandidateRunReceipt,
   AgentCandidateTaskOutcomeEvidence,
   AgentCandidateTaskOutcomeMaterial,
-  AgentCandidateTaskOutcomeSpec,
   AgentCandidateTaskOutputSpec,
-  AgentCandidateTaskRepository,
   AgentCandidateTermination,
   AgentCandidateWorkspaceManifestMaterial,
   AgentCandidateWorkspaceSnapshotEvidence,
-  ReasoningEffort,
   Sha256Digest,
 } from '@tangle-network/agent-interface'
 
@@ -58,6 +58,7 @@ export type AgentCandidateOutputPurpose =
   | 'benchmark-result'
   | 'model-settlement'
   | 'trace'
+  | 'executor-native-evidence'
   | 'executor-capture'
   | 'run-receipt'
   | 'knowledge-retrieval-config'
@@ -166,12 +167,6 @@ export type AgentCandidateModelLimits = Pick<
   'maxModelCalls' | 'maxInputTokens' | 'maxOutputTokens' | 'maxCostUsd'
 >
 
-export interface AgentCandidateBenchmarkGraderIdentity {
-  name: string
-  version: string
-  artifact: AgentCandidateArtifactRef
-}
-
 export interface AgentCandidateProtectedModelReservation {
   preparationId: string
   digest: Sha256Digest
@@ -250,24 +245,12 @@ export interface AgentCandidateExecutionPorts extends AgentCandidateVerification
   memory: AgentCandidateMemoryPort
 }
 
-/** One signed benchmark task and the exact result shape its executor must capture. */
+/** Runtime placement for one exact cell from a signed candidate experiment. */
 export interface AgentCandidateTaskExecution {
   executionId: string
-  benchmark: string
-  benchmarkVersion: string
-  taskId: string
-  splitDigest: Sha256Digest
-  /** Exact agent-visible task instruction. The runtime rejects malformed Unicode. */
-  instruction: string
-  /** Optional source identity, required when the expected outcome is a workspace. */
-  repository?: AgentCandidateTaskRepository
-  outcome: AgentCandidateTaskOutcomeSpec
-  attempt: AgentCandidateAttemptPolicy
-  model: {
-    requested: string
-    reasoningEffort: ReasoningEffort
-  }
-  grader: AgentCandidateBenchmarkGraderIdentity
+  runCell: AgentCandidateRunCell
+  benchmarkSuite: AgentCandidateBenchmarkSuite
+  task: AgentCandidateBenchmarkTask
   /** Absolute paths inside the evaluator-owned execution environment. */
   executionRoots: {
     taskRoot: string
@@ -279,9 +262,6 @@ export interface AgentCandidateTaskExecution {
     candidateRoot?: string
     profileRoot: string
   }
-  workspace: AgentCandidateWorkspaceSnapshotEvidence
-  evaluatorTaskContainer?: ResolvedAgentCandidateContainer
-  limits: AgentCandidateExecutionLimits
 }
 
 export interface VerifiedAgentCandidate {
@@ -328,6 +308,10 @@ export interface PreparedAgentCandidateTrace {
 
 export interface PreparedAgentCandidateExecution {
   readonly bundle: AgentCandidateBundle
+  readonly benchmark: {
+    readonly suite: AgentCandidateBenchmarkSuite
+    readonly task: AgentCandidateBenchmarkTask
+  }
   readonly executionId: string
   readonly roots: {
     execution: {
@@ -360,6 +344,8 @@ export interface PreparedAgentCandidateExecution {
   readonly [preparedCandidateBrand]: true
 }
 
+export type { AgentCandidateBenchmarkGraderIdentity } from '@tangle-network/agent-interface'
+
 export interface AgentCandidateProtectedRunCapture {
   executionId: string
   termination: AgentCandidateTermination
@@ -467,6 +453,7 @@ export interface AgentCandidateBenchmarkGraderPort {
 /** One detached request passed to the trusted environment-specific executor. */
 export interface AgentCandidateExecutorRequest {
   readonly executionId: string
+  readonly benchmark: PreparedAgentCandidateExecution['benchmark']
   /** Immutable bytes from which the executor creates fresh isolated workspaces. */
   readonly inputs: {
     readonly task: AgentCandidateExecutorWorkspaceInput
@@ -532,6 +519,13 @@ export interface AgentCandidateExecutorPort {
       signal: AbortSignal
     },
   ): Promise<AgentCandidateExecutorFinalCapture>
+  /** Remove evaluator-owned execution resources after final capture. Must be idempotent. */
+  dispose?(
+    request: AgentCandidateExecutorStopRequest,
+    context: {
+      signal: AbortSignal
+    },
+  ): Promise<{ readonly disposed: true }>
 }
 
 /** Opaque process identity used for termination without re-exposing launch credentials. */
@@ -551,6 +545,7 @@ export interface AgentCandidateExecutorWorkspaceFile {
   readonly bytes: Uint8Array
 }
 
+/** One exact profile file supplied to an evaluator-owned executor. */
 export interface AgentCandidateExecutorProfileFile {
   readonly path: string
   readonly mode: number
diff --git a/src/candidate-execution/workspace-task.ts b/src/candidate-execution/workspace-task.ts
deleted file mode 100644
index 0f378e60..00000000
--- a/src/candidate-execution/workspace-task.ts
+++ /dev/null
@@ -1,14 +0,0 @@
-import type {
-  AgentCandidateExecutionPlanMaterial,
-  AgentCandidateTaskRepository,
-} from '@tangle-network/agent-interface'
-
-/** Narrow the runtime's workspace-only task contract after schema validation. */
-export function workspaceTaskRepository(
-  task: AgentCandidateExecutionPlanMaterial['task'],
-): AgentCandidateTaskRepository {
-  if (task.outcome.kind !== 'workspace' || !task.repository) {
-    throw new Error('candidate runtime requires a workspace outcome with repository provenance')
-  }
-  return task.repository
-}
diff --git a/src/intelligence/improvement-cycle.ts b/src/intelligence/improvement-cycle.ts
index 7dc4374f..a9065cbc 100644
--- a/src/intelligence/improvement-cycle.ts
+++ b/src/intelligence/improvement-cycle.ts
@@ -1,13 +1,21 @@
-import { type PairedArmRow, pairArms } from '@tangle-network/agent-eval'
 import { type AnalystFinding, assertNoJudgeVerdict } from '@tangle-network/agent-eval/analyst'
 import {
-  assertCodeSurfaceIdentity,
-  codeSurfaceIdentityMaterial,
-} from '@tangle-network/agent-eval/campaign'
-import type { CodeSurface, Scenario, SelfImproveResult } from '@tangle-network/agent-eval/contract'
-import { pairedBootstrap } from '@tangle-network/agent-eval/reporting'
+  type CandidateExperimentExecutionInput,
+  type CompareCandidateExperimentOptions,
+  measuredComparisonFromCandidateExperiment,
+  runCandidateExperiment,
+  type Scenario,
+  verifyCandidateExperiment,
+  verifyCandidateExperimentComparison,
+} from '@tangle-network/agent-eval/contract'
 import type {
+  AgentCandidateBenchmarkCellRef,
   AgentCandidateBundle,
+  AgentCandidateExperiment,
+  AgentCandidateExperimentMeasurement,
+  AgentCandidateRunCell,
+  AgentImprovementActivation,
+  AgentImprovementActivationTarget,
   AgentImprovementMeasuredComparison,
   AgentImprovementProposal,
   AgentImprovementReview,
@@ -18,21 +26,17 @@ import type {
   Sha256Digest,
 } from '@tangle-network/agent-interface'
 import {
-  agentCandidateBundleSchema,
   agentCandidateMaterializationReceiptSchema,
   agentCandidateRunReceiptSchema,
-  agentImprovementMeasuredComparisonSchema,
+  agentImprovementActivationSchema,
   agentImprovementProposalSchema,
   agentImprovementReviewSchema,
   candidateExecutionEvidenceSchema,
 } from '@tangle-network/agent-interface'
 import { materializeCandidateProfile } from '@tangle-network/agent-profile-materialize'
+
 import { runAnalystLoop } from '../analyst-loop'
 import type { RunAnalystLoopOpts, RunAnalystLoopResult } from '../analyst-loop/types'
-import {
-  type AgentCandidateBundleInput,
-  sealAgentCandidateBundle,
-} from '../candidate-execution/bundle'
 import {
   canonicalCandidateBytes,
   canonicalCandidateDigest,
@@ -54,7 +58,6 @@ import {
   candidateMaterializerHarness,
   createAgentCandidateProfileActivation,
   parseAgentCandidateProfileActivation,
-  parseExactAgentProfile,
 } from '../candidate-execution/profile'
 import type {
   AgentCandidateExecutionPorts,
@@ -66,15 +69,10 @@ import {
   verifyAgentCandidateBundle,
 } from '../candidate-execution/verify'
 import { rethrowAfterCleanup } from '../improvement/cleanup'
-import {
-  applyImprovementWinnerToProfile,
-  type ImproveOptions,
-  type ImproveResult,
-  type ImproveSurface,
-  improve,
-} from '../improvement/improve'
+import { type ImproveOptions, type ImproveResult, improve } from '../improvement/improve'
 
 export type {
+  AgentImprovementActivation,
   AgentImprovementMeasuredComparison,
   AgentImprovementProposal,
   AgentImprovementReview,
@@ -82,47 +80,64 @@ export type {
   CandidateExecutionEvidence,
 } from '@tangle-network/agent-interface'
 
-export interface ProposeAgentImprovementOptions<TScenario extends Scenario, TArtifact> {
-  runId: string
-  profile: AgentProfile
-  analysis: Omit<RunAnalystLoopOpts, 'runId' | 'improvementAdapter' | 'autoApply'>
-  improvement: ImproveOptions<TScenario, TArtifact>
-  /** Freezes the measured winner into the exact bundle reviewed for execution. */
-  buildCandidate: (input: {
-    analysis: RunAnalystLoopResult
-    improvement: ImproveResult<TScenario, TArtifact>
-  }) =>
-    | AgentCandidateBundleInput
-    | AgentCandidateBundle
-    | Promise<AgentCandidateBundleInput | AgentCandidateBundle>
-  now?: () => Date
+export interface AgentCandidateExperimentCellPlacement {
+  executionId: string
+  attempt?: number
+  executionRoots: AgentCandidateTaskExecution['executionRoots']
+  stagingRoots: AgentCandidateTaskExecution['stagingRoots']
+  ports: AgentCandidateExecutionPorts
+  preparation?: PrepareAgentCandidateExecutionOptions
+  execution: ExecutePreparedAgentCandidateOptions
 }
 
-export interface ProposeAgentImprovementResult<TScenario extends Scenario, TArtifact> {
-  analysis: RunAnalystLoopResult
-  improvement: ImproveResult<TScenario, TArtifact>
-  proposal: AgentImprovementProposal
+export interface RunAgentCandidateExperimentOptions
+  extends Omit<CompareCandidateExperimentOptions, 'experiment' | 'measurements'> {
+  experiment: AgentCandidateExperiment
+  placeCell: (
+    input: CandidateExperimentExecutionInput,
+  ) => AgentCandidateExperimentCellPlacement | Promise<AgentCandidateExperimentCellPlacement>
+  maxConcurrency?: number
+  signal?: AbortSignal
+}
+
+export interface RunAgentCandidateExperimentResult {
+  experiment: AgentCandidateExperiment
+  measurements: AgentCandidateExperimentMeasurement[]
+  evaluation: AgentImprovementMeasuredComparison
+}
+
+export interface ExecuteAgentCandidateExperimentCellOptions
+  extends CandidateExperimentExecutionInput,
+    AgentCandidateExperimentCellPlacement {}
+
+export interface VerifyCandidateExecutionEvidenceOptions {
+  experiment: AgentCandidateExperiment
+  arm: 'baseline' | 'candidate'
+  benchmarkCell: AgentCandidateBenchmarkCellRef
+  seed: number
+  attempt?: number
+  resolvedResources?: ReadonlyMap<Sha256Digest, string>
+}
+
+/** A failed baseline or candidate cell with its complete Runtime failure result. */
+export class AgentCandidateExperimentCellExecutionError extends Error {
+  readonly finalization: Extract<AgentCandidateRunFinalization, { succeeded: false }>
+
+  constructor(finalization: Extract<AgentCandidateRunFinalization, { succeeded: false }>) {
+    super(`candidate experiment cell failed: ${finalization.reason}`)
+    this.name = 'AgentCandidateExperimentCellExecutionError'
+    this.finalization = finalization
+  }
 }
 
 export interface CreateAgentImprovementProposalOptions {
   runId: string
-  baselineProfile: AgentProfile
   findings: readonly AnalystFinding[]
   evaluation: AgentImprovementMeasuredComparison
-  candidateBundle: AgentCandidateBundleInput | AgentCandidateBundle
   now?: () => Date
 }
 
-export interface CreateAgentImprovementMeasuredComparisonOptions<
-  TScenario extends Scenario,
-  TArtifact,
-> {
-  result: SelfImproveResult<TScenario, TArtifact>
-  measuredSurface: ImproveSurface
-  baselineProfile: AgentProfile
-  candidateBundle: AgentCandidateBundle
-  metadata?: AgentImprovementMeasuredComparison['metadata']
-}
+export type CreateAgentImprovementMeasuredComparisonOptions = CompareCandidateExperimentOptions
 
 export interface ReviewAgentImprovementInput {
   decision: AgentImprovementReviewDecision
@@ -132,38 +147,137 @@ export interface ReviewAgentImprovementInput {
   now?: () => Date
 }
 
-export interface ExecuteApprovedAgentCandidateOptions {
-  proposal: AgentImprovementProposal
-  review: AgentImprovementReview
-  /** Product-owned authentication check for the persisted approval record. */
-  authorizeReview: (
-    review: AgentImprovementReview,
-    proposal: AgentImprovementProposal,
-  ) => boolean | Promise<boolean>
-  task: AgentCandidateTaskExecution
-  ports: AgentCandidateExecutionPorts
-  preparation?: PrepareAgentCandidateExecutionOptions
-  execution: ExecutePreparedAgentCandidateOptions
+export interface CreateAgentImprovementActivationOptions {
+  targets: [AgentImprovementActivationTarget, ...AgentImprovementActivationTarget[]]
+  fundingOwner: string
+  authorizedBy: string
+  now?: () => Date
 }
 
-export type ExecuteApprovedAgentCandidateResult =
-  | {
-      finalization: Extract<AgentCandidateRunFinalization, { succeeded: false }>
-      evidence?: never
-    }
-  | {
-      finalization: Extract<AgentCandidateRunFinalization, { succeeded: true }>
-      evidence: CandidateExecutionEvidence
-    }
+export interface ProposeAgentImprovementOptions<TScenario extends Scenario, TArtifact> {
+  runId: string
+  profile: AgentProfile
+  analysis: Omit<RunAnalystLoopOpts, 'runId' | 'improvementAdapter' | 'autoApply'>
+  improvement: ImproveOptions<TScenario, TArtifact>
+  buildExperiment: (input: {
+    analysis: RunAnalystLoopResult
+    improvement: ImproveResult<TScenario, TArtifact>
+  }) => AgentCandidateExperiment | Promise<AgentCandidateExperiment>
+  placeCell: RunAgentCandidateExperimentOptions['placeCell']
+  maxConcurrency?: number
+  signal?: AbortSignal
+  candidate?: AgentImprovementMeasuredComparison['candidate']
+  metadata?: AgentImprovementMeasuredComparison['metadata']
+  now?: () => Date
+}
 
-export interface VerifyCandidateExecutionEvidenceOptions {
+export interface ProposeAgentImprovementResult<TScenario extends Scenario, TArtifact> {
+  analysis: RunAnalystLoopResult
+  improvement: ImproveResult<TScenario, TArtifact>
+  experiment: AgentCandidateExperiment
+  measurements: AgentCandidateExperimentMeasurement[]
   proposal: AgentImprovementProposal
-  review: AgentImprovementReview
-  expectedCount: number
-  resolvedResources?: ReadonlyMap<Sha256Digest, string>
 }
 
-/** Analyze one run and freeze its measured winner into one exact proposal. */
+type AnalystImprovementProposal = Omit<AgentImprovementProposal, 'findings'> & {
+  findings: AnalystFinding[]
+}
+
+/** Execute both arms of one immutable experiment and derive its paired result. */
+export async function runAgentCandidateExperiment(
+  options: RunAgentCandidateExperimentOptions,
+): Promise<RunAgentCandidateExperimentResult> {
+  const experiment = verifyCandidateExperiment(options.experiment)
+  const measurements = await runCandidateExperiment({
+    experiment,
+    ...(options.maxConcurrency === undefined ? {} : { maxConcurrency: options.maxConcurrency }),
+    ...(options.signal ? { signal: options.signal } : {}),
+    execute: async (input) => {
+      const placement = await options.placeCell(input)
+      return await executeAgentCandidateExperimentCell({ ...input, ...placement })
+    },
+  })
+  const evaluation = createAgentImprovementMeasuredComparison({
+    experiment,
+    measurements,
+    runId: options.runId,
+    ...(options.candidate ? { candidate: options.candidate } : {}),
+    ...(options.generationsExplored === undefined
+      ? {}
+      : { generationsExplored: options.generationsExplored }),
+    ...(options.searchDurationMs === undefined
+      ? {}
+      : { searchDurationMs: options.searchDurationMs }),
+    ...(options.searchCostUsd === undefined ? {} : { searchCostUsd: options.searchCostUsd }),
+    ...(options.metadata ? { metadata: options.metadata } : {}),
+  })
+  return { experiment, measurements, evaluation }
+}
+
+/** Execute one exact arm, task, repetition, seed, and attempt through Runtime. */
+export async function executeAgentCandidateExperimentCell(
+  options: ExecuteAgentCandidateExperimentCellOptions,
+): Promise<CandidateExecutionEvidence> {
+  const experiment = verifyCandidateExperiment(options.experiment)
+  const bundle = experiment[options.arm]
+  assertExactExperimentInput(options, experiment, bundle)
+  const attempt = options.attempt ?? 1
+  if (attempt > options.task.attempt.maxAttempts) {
+    throw new Error('candidate experiment attempt exceeds the signed task policy')
+  }
+  const runCell = canonicalCandidateDocument<AgentCandidateRunCell>({
+    kind: 'agent-candidate-run-cell',
+    experimentDigest: experiment.digest,
+    arm: options.arm,
+    bundleDigest: bundle.digest,
+    suiteDigest: options.benchmarkCell.suiteDigest,
+    taskDigest: options.task.digest,
+    taskIndex: options.benchmarkCell.taskIndex,
+    repetition: options.benchmarkCell.repetition,
+    seed: options.seed,
+    attempt,
+  }).value
+  const verified = await verifyAgentCandidateBundle(bundle, options.ports)
+  const prepared = await prepareAgentCandidateExecution(
+    verified,
+    {
+      executionId: options.executionId,
+      runCell,
+      benchmarkSuite: experiment.benchmark.suite,
+      task: options.task,
+      executionRoots: options.executionRoots,
+      stagingRoots: options.stagingRoots,
+    },
+    options.ports,
+    options.preparation,
+  )
+  const finalization = await executePreparedAgentCandidate(prepared, options.execution)
+  if (!finalization.succeeded) {
+    throw new AgentCandidateExperimentCellExecutionError(finalization)
+  }
+  const evidence = canonicalCandidateDocument<CandidateExecutionEvidence>({
+    kind: 'agent-candidate-execution-evidence',
+    materializationReceipt: prepared.materializationReceipt.value,
+    receipt: finalization.receipt.value,
+  }).value
+  return verifyCandidateExecutionEvidence(evidence, {
+    experiment,
+    arm: options.arm,
+    benchmarkCell: options.benchmarkCell,
+    seed: options.seed,
+    attempt,
+    resolvedResources: verifiedResourceTextByDigest(verified),
+  })
+}
+
+/** Delegate all statistics and promotion checks to agent-eval's receipt-based comparison. */
+export function createAgentImprovementMeasuredComparison(
+  options: CreateAgentImprovementMeasuredComparisonOptions,
+): AgentImprovementMeasuredComparison {
+  return verifyCandidateExperimentComparison(measuredComparisonFromCandidateExperiment(options))
+}
+
+/** Analyze, search, then remeasure the resulting exact candidate before proposing it. */
 export async function proposeAgentImprovement<TScenario extends Scenario, TArtifact>(
   options: ProposeAgentImprovementOptions<TScenario, TArtifact>,
 ): Promise<ProposeAgentImprovementResult<TScenario, TArtifact>> {
@@ -173,13 +287,9 @@ export async function proposeAgentImprovement<TScenario extends Scenario, TArtif
       ? 'memory'
       : null
   if (writeBackSurface) {
-    throw new Error(
-      `proposeAgentImprovement cannot write ${writeBackSurface} before human approval`,
-    )
+    throw new Error(`proposeAgentImprovement cannot write ${writeBackSurface} before approval`)
   }
-  const analysis = await runAnalystLoop(
-    Object.assign({}, options.analysis, { runId: options.runId }),
-  )
+  const analysis = await runAnalystLoop({ ...options.analysis, runId: options.runId })
   const findings = assertNoJudgeVerdict(
     analysis.analystResult.findings,
     'proposeAgentImprovement findings',
@@ -187,35 +297,48 @@ export async function proposeAgentImprovement<TScenario extends Scenario, TArtif
   const improvement = await improve(options.profile, [...findings], options.improvement)
   try {
     if (!improvement.shipped) {
-      throw new Error('agent improvement proposal requires a passing measured comparison')
+      throw new Error('agent improvement search did not produce a promotable candidate')
     }
-    const candidateBundle = sealBuiltCandidate(
-      await options.buildCandidate({ analysis, improvement }),
+    const experiment = verifyCandidateExperiment(
+      await options.buildExperiment({ analysis, improvement }),
     )
+    if (
+      canonicalCandidateDigest(experiment.baseline.profile) !==
+      canonicalCandidateDigest(options.profile)
+    ) {
+      throw new Error('candidate experiment baseline does not match the analyzed agent profile')
+    }
+    const measured = await runAgentCandidateExperiment({
+      experiment,
+      runId: options.runId,
+      placeCell: options.placeCell,
+      ...(options.maxConcurrency === undefined ? {} : { maxConcurrency: options.maxConcurrency }),
+      ...(options.signal ? { signal: options.signal } : {}),
+      ...(options.candidate ? { candidate: options.candidate } : {}),
+      ...(options.metadata ? { metadata: options.metadata } : {}),
+      generationsExplored: improvement.raw.generationsExplored,
+      searchDurationMs: improvement.raw.durationMs,
+      searchCostUsd: improvement.raw.totalCostUsd,
+    })
     const proposal = createAgentImprovementProposal({
       runId: options.runId,
-      baselineProfile: options.profile,
       findings,
-      evaluation: createAgentImprovementMeasuredComparison({
-        result: improvement.raw,
-        measuredSurface: options.improvement.surface ?? 'prompt',
-        baselineProfile: options.profile,
-        candidateBundle,
-      }),
-      candidateBundle,
+      evaluation: measured.evaluation,
       ...(options.now ? { now: options.now } : {}),
     })
-    return { analysis, improvement, proposal }
+    return {
+      analysis,
+      improvement,
+      experiment,
+      measurements: measured.measurements,
+      proposal,
+    }
   } catch (cause) {
     return rethrowAfterCleanup(cause, () => improvement.dispose(), 'proposeAgentImprovement failed')
   }
 }
 
-/**
- * Freeze an already-measured improvement into the one reviewable proposal
- * contract. Products that run analysis or evaluation in separate workers use
- * this constructor instead of rerunning either phase or rebuilding digests.
- */
+/** Create the reviewable record only from a complete, recomputable experiment result. */
 export function createAgentImprovementProposal(
   options: CreateAgentImprovementProposalOptions,
 ): AgentImprovementProposal {
@@ -223,37 +346,30 @@ export function createAgentImprovementProposal(
     [...options.findings],
     'createAgentImprovementProposal findings',
   )
-  const candidateBundle = sealBuiltCandidate(options.candidateBundle)
-  const baselineProfile = parseExactAgentProfile(
-    options.baselineProfile,
-    'proposal baseline profile',
-  )
-  const candidateProfile = agentCandidateProfileAsAgentProfile(candidateBundle.profile)
-  const evaluation = agentImprovementMeasuredComparisonSchema.parse(options.evaluation)
+  const evaluation = verifyCandidateExperimentComparison(options.evaluation)
   if (evaluation.decision.outcome !== 'ship') {
-    throw new Error('agent improvement proposal requires a passing measured comparison')
+    throw new Error('agent improvement proposal requires a passing experiment')
   }
-  assertSupportedCandidateBundle(candidateBundle)
-  assertMeasuredCandidateBinding(evaluation, baselineProfile, candidateBundle)
-  const changedSurfaces = deriveChangedSurfaces(baselineProfile, candidateProfile, candidateBundle)
-  const withoutDigest = {
-    kind: 'agent-improvement-proposal' as const,
-    runId: options.runId,
-    changedSurfaces,
-    proposedAt: (options.now ?? (() => new Date()))().toISOString(),
-    baselineProfile,
-    findings: immutableCandidateValue([
-      ...findings,
-    ]) as unknown as AgentImprovementProposal['findings'],
-    evaluation,
-    candidateBundle,
+  if (options.runId !== evaluation.provenance.runId) {
+    throw new Error('proposal runId does not match its measured experiment')
   }
+  const changedSurfaces = deriveChangedSurfaces(
+    evaluation.experiment.baseline,
+    evaluation.experiment.candidate,
+  )
   return agentImprovementProposalSchema.parse(
-    canonicalCandidateDocument<AgentImprovementProposal>(withoutDigest).value,
+    canonicalCandidateDocument<AnalystImprovementProposal>({
+      kind: 'agent-improvement-proposal',
+      runId: options.runId,
+      changedSurfaces,
+      proposedAt: (options.now ?? (() => new Date()))().toISOString(),
+      findings: [...findings],
+      evaluation,
+    }).value,
   )
 }
 
-/** Persist an approve/reject/change-request decision bound to one exact proposal. */
+/** Persist a human or tenant-policy decision bound to one exact proposal. */
 export function reviewAgentImprovementProposal(
   inputProposal: AgentImprovementProposal,
   input: ReviewAgentImprovementInput,
@@ -261,158 +377,234 @@ export function reviewAgentImprovementProposal(
   const proposal = verifyAgentImprovementProposal(inputProposal)
   if (!input.reviewedBy.trim()) throw new Error('candidate review requires reviewedBy')
   if (!input.reason.trim()) throw new Error('candidate review requires a reason')
-  if (input.decision === 'approve') {
-    if (proposal.evaluation.decision.outcome !== 'ship') {
-      throw new Error('candidate cannot be approved without a passing measured comparison')
-    }
-  }
-  const withoutDigest = {
-    kind: 'agent-improvement-review' as const,
-    proposalDigest: proposal.digest,
-    candidateBundleDigest: proposal.candidateBundle.digest,
-    decision: input.decision,
-    reviewedBy: input.reviewedBy,
-    reviewedAt: (input.now ?? (() => new Date()))().toISOString(),
-    reason: input.reason,
-    ...(input.feedback === undefined ? {} : { feedback: input.feedback }),
+  if (input.decision === 'approve' && proposal.evaluation.decision.outcome !== 'ship') {
+    throw new Error('candidate cannot be approved without a passing experiment')
   }
   return agentImprovementReviewSchema.parse(
-    canonicalCandidateDocument<AgentImprovementReview>(withoutDigest).value,
+    canonicalCandidateDocument<AgentImprovementReview>({
+      kind: 'agent-improvement-review',
+      proposalDigest: proposal.digest,
+      decision: input.decision,
+      reviewedBy: input.reviewedBy,
+      reviewedAt: (input.now ?? (() => new Date()))().toISOString(),
+      reason: input.reason,
+      ...(input.feedback === undefined ? {} : { feedback: input.feedback }),
+    }).value,
   )
 }
 
-/** Verify, materialize, run, grade, and receipt only the exact approved bundle. */
-export async function executeApprovedAgentCandidate(
-  options: ExecuteApprovedAgentCandidateOptions,
-): Promise<ExecuteApprovedAgentCandidateResult> {
-  const proposal = verifyAgentImprovementProposal(options.proposal)
-  const review = verifyAgentImprovementReview(options.review)
-  if (review.decision !== 'approve') throw new Error('candidate review is not an approval')
-  if (review.proposalDigest !== proposal.digest) {
-    throw new Error('candidate approval does not match the proposed improvement')
+/** Authorize product-owned writes only after the exact candidate was measured and approved. */
+export function createAgentImprovementActivation(
+  inputProposal: AgentImprovementProposal,
+  inputReview: AgentImprovementReview,
+  options: CreateAgentImprovementActivationOptions,
+): AgentImprovementActivation {
+  const proposal = verifyAgentImprovementProposal(inputProposal)
+  const review = verifyAgentImprovementReview(inputReview)
+  if (review.decision !== 'approve' || review.proposalDigest !== proposal.digest) {
+    throw new Error('candidate activation requires an approval for the exact proposal')
+  }
+  if (!options.fundingOwner.trim() || !options.authorizedBy.trim()) {
+    throw new Error('candidate activation authority must be non-empty')
+  }
+  const experiment = proposal.evaluation.experiment
+  assertActivationTargets(proposal.changedSurfaces, experiment, options.targets)
+  return agentImprovementActivationSchema.parse(
+    canonicalCandidateDocument<AgentImprovementActivation>({
+      kind: 'agent-improvement-activation',
+      proposalDigest: proposal.digest,
+      reviewDigest: review.digest,
+      experimentDigest: experiment.digest,
+      candidateBundleDigest: experiment.candidate.digest,
+      targets: options.targets,
+      fundingOwner: options.fundingOwner,
+      authorizedBy: options.authorizedBy,
+      authorizedAt: (options.now ?? (() => new Date()))().toISOString(),
+    }).value,
+  )
+}
+
+/** Validate a proposal and recompute every binding to its measured experiment. */
+export function verifyAgentImprovementProposal(input: unknown): AgentImprovementProposal {
+  const proposal = verifyCanonicalCandidateDocument(
+    agentImprovementProposalSchema.parse(input),
+    'agent improvement proposal',
+  )
+  const evaluation = verifyCandidateExperimentComparison(proposal.evaluation)
+  if (evaluation.decision.outcome !== 'ship') {
+    throw new Error('agent improvement proposal does not contain a passing experiment')
   }
-  const bundle = proposal.candidateBundle
-  if (review.candidateBundleDigest !== bundle.digest) {
-    throw new Error('candidate approval does not match the executable bundle')
+  if (proposal.runId !== evaluation.provenance.runId) {
+    throw new Error('proposal runId does not match its measured experiment')
   }
-  if (!(await options.authorizeReview(review, proposal))) {
-    throw new Error('candidate approval was not authorized by the configured authority')
+  const changedSurfaces = deriveChangedSurfaces(
+    evaluation.experiment.baseline,
+    evaluation.experiment.candidate,
+  )
+  if (!sameOrderedValues(proposal.changedSurfaces, changedSurfaces)) {
+    throw new Error('proposal changed surfaces do not match its exact experiment')
   }
+  assertNoJudgeDerivedProposalFindings(proposal.findings)
+  return proposal
+}
 
-  const verified = await verifyAgentCandidateBundle(bundle, options.ports)
-  const prepared = await prepareAgentCandidateExecution(
-    verified,
-    options.task,
-    options.ports,
-    options.preparation,
+/** Validate the canonical identity and wire shape of an improvement review. */
+export function verifyAgentImprovementReview(input: unknown): AgentImprovementReview {
+  return verifyCanonicalCandidateDocument(
+    agentImprovementReviewSchema.parse(input),
+    'agent improvement review',
   )
-  const finalization = await executePreparedAgentCandidate(prepared, options.execution)
-  if (!finalization.succeeded) return { finalization }
-  const evidence = canonicalCandidateDocument<CandidateExecutionEvidence>({
-    kind: 'agent-candidate-execution-evidence',
-    proposalDigest: proposal.digest,
-    reviewDigest: review.digest,
-    executionId: prepared.executionId,
-    succeeded: true,
-    materializationReceipt: prepared.materializationReceipt.value,
-    profileActivation: prepared.profileActivation,
-    receipt: finalization.receipt.value,
-  }).value
-  const verifiedEvidence = verifyCandidateExecutionEvidence([evidence], {
-    proposal,
-    review,
-    expectedCount: 1,
-    resolvedResources: verifiedResourceTextByDigest(verified),
-  })[0]!
-  return {
-    finalization,
-    evidence: verifiedEvidence,
+}
+
+/** Validate activation authority against the exact proposal, review, experiment, and base state. */
+export function verifyAgentImprovementActivation(input: {
+  proposal: unknown
+  review: unknown
+  activation: unknown
+}): AgentImprovementActivation {
+  const proposal = verifyAgentImprovementProposal(input.proposal)
+  const review = verifyAgentImprovementReview(input.review)
+  const activation = verifyCanonicalCandidateDocument(
+    agentImprovementActivationSchema.parse(input.activation),
+    'agent improvement activation',
+  )
+  const experiment = proposal.evaluation.experiment
+  if (
+    review.decision !== 'approve' ||
+    review.proposalDigest !== proposal.digest ||
+    activation.proposalDigest !== proposal.digest ||
+    activation.reviewDigest !== review.digest ||
+    activation.experimentDigest !== experiment.digest ||
+    activation.candidateBundleDigest !== experiment.candidate.digest
+  ) {
+    throw new Error('candidate activation does not bind the measured and approved candidate')
   }
+  assertActivationTargets(proposal.changedSurfaces, experiment, activation.targets)
+  return activation
 }
 
-/** Verify approval, receipts, uniqueness, and the exact native profile files executed. */
+/** Recheck one Runtime receipt against its exact signed experiment cell. */
 export function verifyCandidateExecutionEvidence(
   input: unknown,
   options: VerifyCandidateExecutionEvidenceOptions,
-): CandidateExecutionEvidence[] {
-  const proposal = verifyAgentImprovementProposal(options.proposal)
-  const review = verifyAgentImprovementReview(options.review)
-  if (review.decision !== 'approve') throw new Error('candidate review is not an approval')
+): CandidateExecutionEvidence {
+  const experiment = verifyCandidateExperiment(options.experiment)
+  const bundle = experiment[options.arm]
+  const task = experiment.benchmark.tasks[options.benchmarkCell.taskIndex]
+  const index =
+    options.benchmarkCell.taskIndex * experiment.benchmark.suite.reps +
+    options.benchmarkCell.repetition
   if (
-    review.proposalDigest !== proposal.digest ||
-    review.candidateBundleDigest !== proposal.candidateBundle.digest
+    !task ||
+    options.benchmarkCell.suiteDigest !== experiment.benchmark.suite.digest ||
+    options.seed !== experiment.benchmark.suite.seeds[index]
   ) {
-    throw new Error('candidate review does not bind the proposed bundle')
+    throw new Error('candidate execution evidence points outside its signed experiment')
   }
-  if (!Number.isSafeInteger(options.expectedCount) || options.expectedCount < 1) {
-    throw new Error('candidate execution evidence expectedCount must be a positive integer')
+  const evidence = verifyCanonicalCandidateDocument(
+    candidateExecutionEvidenceSchema.parse(input),
+    'candidate execution evidence',
+  )
+  const materialization = verifyCanonicalCandidateDocument(
+    agentCandidateMaterializationReceiptSchema.parse(evidence.materializationReceipt),
+    'candidate materialization receipt',
+  )
+  const receipt = verifyCanonicalCandidateDocument(
+    agentCandidateRunReceiptSchema.parse(evidence.receipt),
+    'candidate run receipt',
+  )
+  const plan = materialization.executionPlan
+  const cell = plan.material.runCell
+  const attempt = options.attempt ?? 1
+  if (
+    cell.experimentDigest !== experiment.digest ||
+    cell.arm !== options.arm ||
+    cell.bundleDigest !== bundle.digest ||
+    cell.suiteDigest !== experiment.benchmark.suite.digest ||
+    cell.taskDigest !== task.digest ||
+    cell.taskIndex !== options.benchmarkCell.taskIndex ||
+    cell.repetition !== options.benchmarkCell.repetition ||
+    cell.seed !== options.seed ||
+    cell.attempt !== attempt ||
+    canonicalCandidateDigest(omitTopLevelDigest(cell)) !== cell.digest
+  ) {
+    throw new Error('candidate execution receipt substituted its signed experiment cell')
   }
-  if (!Array.isArray(input) || input.length !== options.expectedCount) {
-    throw new Error(
-      `candidate execution evidence must contain exactly ${options.expectedCount} rows`,
-    )
+  assertCapturedInput(
+    materialization.benchmark.suite,
+    experiment.benchmark.suite,
+    'benchmark suite',
+  )
+  assertCapturedInput(materialization.benchmark.task, task, 'benchmark task')
+  assertEvidenceMaterialDigest(plan, 'candidate execution plan')
+  assertEvidenceMaterialDigest(
+    materialization.profileActivation.profilePlan,
+    'candidate profile plan',
+  )
+  const expectedProfilePlan = materializeCandidateProfile(
+    bundle.profile,
+    candidateMaterializerHarness(materialization.harness),
+    { resolvedResources: options.resolvedResources },
+  )
+  const activation = parseAgentCandidateProfileActivation(
+    materialization.profileActivation,
+    materialization.profileActivation.profilePlan.digest,
+  )
+  const regeneratedActivation = createAgentCandidateProfileActivation(
+    expectedProfilePlan,
+    materialization.profileActivation.profilePlan,
+  )
+  if (activation.digest !== regeneratedActivation.digest) {
+    throw new Error('candidate profile activation does not match the experiment bundle')
+  }
+  if (
+    materialization.bundleDigest !== bundle.digest ||
+    receipt.bundleDigest !== bundle.digest ||
+    receipt.runCellDigest !== cell.digest ||
+    receipt.materializationReceiptDigest !== materialization.digest ||
+    receipt.executionPlanDigest !== plan.digest
+  ) {
+    throw new Error('candidate execution evidence does not bind one exact Runtime run')
+  }
+  assertEvidenceMaterialDigest(receipt.modelSettlement, 'candidate model settlement')
+  assertEvidenceMaterialDigest(receipt.taskOutcome, 'candidate task outcome')
+  assertEvidenceMaterialDigest(receipt.benchmarkResult, 'candidate benchmark result')
+  return immutableCandidateValue(evidence)
+}
+
+function assertExactExperimentInput(
+  input: CandidateExperimentExecutionInput,
+  experiment: AgentCandidateExperiment,
+  bundle: AgentCandidateBundle,
+): void {
+  const task = experiment.benchmark.tasks[input.benchmarkCell.taskIndex]
+  const index =
+    input.benchmarkCell.taskIndex * experiment.benchmark.suite.reps + input.benchmarkCell.repetition
+  if (
+    input.experiment.digest !== experiment.digest ||
+    input.bundle.digest !== bundle.digest ||
+    !task ||
+    input.task.digest !== task.digest ||
+    input.benchmarkCell.suiteDigest !== experiment.benchmark.suite.digest ||
+    input.seed !== experiment.benchmark.suite.seeds[index]
+  ) {
+    throw new Error('Runtime received a substituted candidate experiment cell')
+  }
+}
+
+function assertCapturedInput(
+  captured: { digest: Sha256Digest; material: { sha256: Sha256Digest; byteLength: number } },
+  expected: { digest: Sha256Digest },
+  label: string,
+): void {
+  const bytes = canonicalCandidateBytes(omitTopLevelDigest(expected))
+  if (
+    captured.digest !== expected.digest ||
+    captured.material.sha256 !== expected.digest ||
+    captured.material.byteLength !== bytes.byteLength
+  ) {
+    throw new Error(`candidate materialization substituted its ${label}`)
   }
-  const executionIds = new Set<string>()
-  const runReceiptDigests = new Set<Sha256Digest>()
-  const verified = input.map((entry) => {
-    const evidence = verifyCanonicalCandidateDocument(
-      candidateExecutionEvidenceSchema.parse(entry),
-      'candidate execution evidence',
-    )
-    const receipt = verifyCanonicalCandidateDocument(
-      agentCandidateRunReceiptSchema.parse(evidence.receipt),
-      'candidate execution run receipt',
-    )
-    const materializationReceipt = verifyCanonicalCandidateDocument(
-      agentCandidateMaterializationReceiptSchema.parse(evidence.materializationReceipt),
-      'candidate materialization receipt',
-    )
-    assertEvidenceMaterialDigest(materializationReceipt.profilePlan, 'candidate profile plan')
-    const profilePlan = materializeCandidateProfile(
-      proposal.candidateBundle.profile,
-      candidateMaterializerHarness(materializationReceipt.harness),
-      { resolvedResources: options.resolvedResources },
-    )
-    const profileActivation = parseAgentCandidateProfileActivation(
-      evidence.profileActivation,
-      materializationReceipt.profilePlan.digest,
-    )
-    const regenerated = createAgentCandidateProfileActivation(
-      profilePlan,
-      materializationReceipt.profilePlan,
-    )
-    if (regenerated.digest !== profileActivation.digest) {
-      throw new Error('candidate profile plan files do not match the executed materialization')
-    }
-    assertEvidenceMaterialDigest(materializationReceipt.executionPlan, 'candidate execution plan')
-    assertEvidenceMaterialDigest(receipt.modelSettlement, 'candidate model settlement')
-    assertEvidenceMaterialDigest(receipt.taskOutcome, 'candidate task outcome')
-    assertEvidenceMaterialDigest(receipt.benchmarkResult, 'candidate benchmark result')
-    if (
-      evidence.proposalDigest !== proposal.digest ||
-      evidence.reviewDigest !== review.digest ||
-      receipt.bundleDigest !== proposal.candidateBundle.digest ||
-      materializationReceipt.bundleDigest !== proposal.candidateBundle.digest ||
-      receipt.materializationReceiptDigest !== materializationReceipt.digest ||
-      receipt.executionPlanDigest !== materializationReceipt.executionPlan.digest ||
-      evidence.executionId !== materializationReceipt.executionPlan.material.executionId
-    ) {
-      throw new Error('candidate execution evidence does not bind the approved candidate')
-    }
-    if (receipt.termination.kind !== 'exit' || receipt.termination.exitCode !== 0) {
-      throw new Error('candidate execution evidence is not a successful process execution')
-    }
-    if (executionIds.has(evidence.executionId)) {
-      throw new Error('candidate execution evidence reuses an execution id')
-    }
-    if (runReceiptDigests.has(receipt.digest)) {
-      throw new Error('candidate execution evidence reuses a run receipt')
-    }
-    executionIds.add(evidence.executionId)
-    runReceiptDigests.add(receipt.digest)
-    return immutableCandidateValue(evidence)
-  })
-  return verified
 }
 
 function assertEvidenceMaterialDigest(
@@ -433,102 +625,6 @@ function assertEvidenceMaterialDigest(
   }
 }
 
-/** Validate a proposal's schema, profile, sealed bundle, and canonical digest. */
-export function verifyAgentImprovementProposal(input: unknown): AgentImprovementProposal {
-  const proposal = agentImprovementProposalSchema.parse(input)
-  const parsedBaselineProfile = parseExactAgentProfile(
-    proposal.baselineProfile,
-    'proposal baseline profile',
-  )
-  const parsedBundle = agentCandidateBundleSchema.parse(proposal.candidateBundle)
-  const candidateBundle = sealAgentCandidateBundle(omitTopLevelDigest(parsedBundle))
-  if (candidateBundle.digest !== parsedBundle.digest)
-    throw new Error('proposal candidate bundle digest is invalid')
-  const candidateProfile = agentCandidateProfileAsAgentProfile(candidateBundle.profile)
-  if (proposal.evaluation.decision.outcome !== 'ship') {
-    throw new Error('agent improvement proposal requires a passing measured comparison')
-  }
-  assertSupportedCandidateBundle(candidateBundle)
-  assertMeasuredCandidateBinding(proposal.evaluation, parsedBaselineProfile, candidateBundle)
-  const changedSurfaces = deriveChangedSurfaces(
-    parsedBaselineProfile,
-    candidateProfile,
-    candidateBundle,
-  )
-  if (!sameOrderedValues(proposal.changedSurfaces, changedSurfaces)) {
-    throw new Error(
-      'proposal changed surfaces do not match the exact baseline and candidate bundle',
-    )
-  }
-  return verifyCanonicalCandidateDocument(proposal, 'agent improvement proposal')
-}
-
-function assertMeasuredCandidateBinding(
-  evaluation: AgentImprovementMeasuredComparison,
-  baselineProfile: AgentProfile,
-  candidateBundle: AgentCandidateBundle,
-): void {
-  if (evaluation.baselineProfileDigest !== canonicalCandidateDigest(baselineProfile)) {
-    throw new Error('measured comparison does not bind the included baseline profile')
-  }
-  if (evaluation.candidateBundleDigest !== candidateBundle.digest) {
-    throw new Error('measured comparison does not bind the exact candidate bundle')
-  }
-}
-
-function assertRawMeasuredWinnerBinding(input: {
-  surface: ImproveSurface
-  baselineProfile: AgentProfile
-  winner: unknown
-  candidateBundle: AgentCandidateBundle
-}): void {
-  const candidateProfile = agentCandidateProfileAsAgentProfile(input.candidateBundle.profile)
-  const baselineDigest = canonicalCandidateDigest(input.baselineProfile)
-  if (input.surface === 'code') {
-    if (canonicalCandidateDigest(candidateProfile) !== baselineDigest) {
-      throw new Error('code improvement must retain the measured profile')
-    }
-    assertMeasuredCodeBundle(input.winner, input.candidateBundle)
-  } else if (input.surface === 'memory') {
-    throw new Error(
-      'memory improvement proposals require a content-addressed bundle binding, which is unsupported',
-    )
-  } else {
-    const measured = measuredWinnerProfile(input.baselineProfile, input.surface, input.winner)
-    if (!measured) {
-      throw new Error(
-        'skill-document improvement proposals require a content-addressed bundle binding, which is unsupported',
-      )
-    }
-    if (canonicalCandidateDigest(candidateProfile) !== canonicalCandidateDigest(measured)) {
-      throw new Error('proposal candidate profile does not match the measured winner surface')
-    }
-  }
-
-  const changedSurfaces = deriveChangedSurfaces(
-    input.baselineProfile,
-    candidateProfile,
-    input.candidateBundle,
-  )
-  if (input.surface === 'agent-profile') {
-    if (changedSurfaces.includes('code') || changedSurfaces.includes('knowledge')) {
-      throw new Error('agent-profile measurement does not cover code or knowledge bundle changes')
-    }
-  } else if (changedSurfaces.length !== 1 || changedSurfaces[0] !== input.surface) {
-    throw new Error(
-      `measured '${input.surface}' winner does not cover candidate changes: ${changedSurfaces.join(', ')}`,
-    )
-  }
-}
-
-function assertSupportedCandidateBundle(candidateBundle: AgentCandidateBundle): void {
-  if (candidateBundle.memory.mode !== 'disabled') {
-    throw new Error(
-      'memory improvement proposals require a content-addressed policy binding, which is unsupported',
-    )
-  }
-}
-
 const CHANGED_SURFACE_ORDER: readonly AgentImprovementSurface[] = [
   'prompt',
   'skills',
@@ -543,45 +639,51 @@ const CHANGED_SURFACE_ORDER: readonly AgentImprovementSurface[] = [
 ]
 
 function deriveChangedSurfaces(
-  baseline: AgentProfile,
-  candidate: AgentProfile,
-  bundle: AgentCandidateBundle,
+  baselineBundle: AgentCandidateBundle,
+  candidateBundle: AgentCandidateBundle,
 ): [AgentImprovementSurface, ...AgentImprovementSurface[]] {
+  const baseline = improvementSurfaceValues(baselineBundle)
+  const candidate = improvementSurfaceValues(candidateBundle)
   const changed = new Set<AgentImprovementSurface>()
-  const pairs: Array<[AgentImprovementSurface, unknown, unknown]> = [
-    [
-      'prompt',
-      { prompt: baseline.prompt ?? null, instructions: baseline.resources?.instructions ?? null },
-      { prompt: candidate.prompt ?? null, instructions: candidate.resources?.instructions ?? null },
-    ],
-    ['skills', baseline.resources?.skills ?? null, candidate.resources?.skills ?? null],
-    [
-      'tools',
-      { tools: baseline.tools ?? null, resources: baseline.resources?.tools ?? null },
-      { tools: candidate.tools ?? null, resources: candidate.resources?.tools ?? null },
-    ],
-    ['mcp', baseline.mcp ?? null, candidate.mcp ?? null],
-    ['hooks', baseline.hooks ?? null, candidate.hooks ?? null],
-    [
-      'subagents',
-      { subagents: baseline.subagents ?? null, resources: baseline.resources?.agents ?? null },
-      { subagents: candidate.subagents ?? null, resources: candidate.resources?.agents ?? null },
-    ],
-    ['agent-profile', opaqueProfileSlice(baseline), opaqueProfileSlice(candidate)],
-  ]
-  for (const [surface, before, after] of pairs) {
-    if (!sameCanonicalValue(before, after)) changed.add(surface)
+  for (const surface of CHANGED_SURFACE_ORDER) {
+    if (
+      canonicalCandidateDigest(baseline[surface]) !== canonicalCandidateDigest(candidate[surface])
+    ) {
+      changed.add(surface)
+    }
   }
-  if (bundle.memory.mode !== 'disabled') changed.add('memory')
-  if (bundle.code.kind !== 'disabled') changed.add('code')
-  if (bundle.knowledge) changed.add('knowledge')
   const ordered = CHANGED_SURFACE_ORDER.filter((surface) => changed.has(surface))
-  if (ordered.length === 0) {
-    throw new Error('candidate bundle does not contain a change from the included baseline profile')
-  }
+  if (ordered.length === 0) throw new Error('candidate experiment does not change an agent surface')
   return ordered as [AgentImprovementSurface, ...AgentImprovementSurface[]]
 }
 
+function improvementSurfaceValues(
+  bundle: AgentCandidateBundle,
+): Record<AgentImprovementSurface, unknown> {
+  const profile = agentCandidateProfileAsAgentProfile(bundle.profile)
+  return {
+    prompt: {
+      prompt: profile.prompt ?? null,
+      instructions: profile.resources?.instructions ?? null,
+    },
+    skills: profile.resources?.skills ?? null,
+    tools: {
+      tools: profile.tools ?? null,
+      resources: profile.resources?.tools ?? null,
+    },
+    mcp: profile.mcp ?? null,
+    hooks: profile.hooks ?? null,
+    subagents: {
+      subagents: profile.subagents ?? null,
+      resources: profile.resources?.agents ?? null,
+    },
+    'agent-profile': { profile: opaqueProfileSlice(profile), execution: bundle.execution },
+    memory: bundle.memory,
+    code: bundle.code,
+    knowledge: bundle.knowledge ?? null,
+  }
+}
+
 function opaqueProfileSlice(profile: AgentProfile): unknown {
   const {
     prompt: _prompt,
@@ -599,448 +701,59 @@ function opaqueProfileSlice(profile: AgentProfile): unknown {
     agents: _agents,
     ...opaqueResources
   } = resources ?? {}
-  return immutableCandidateValue({
+  return {
     ...opaqueProfile,
     ...(Object.keys(opaqueResources).length > 0 ? { resources: opaqueResources } : {}),
-  })
-}
-
-function sameCanonicalValue(left: unknown, right: unknown): boolean {
-  return canonicalCandidateDigest(left) === canonicalCandidateDigest(right)
-}
-
-function sameOrderedValues<T>(left: readonly T[], right: readonly T[]): boolean {
-  return left.length === right.length && left.every((value, index) => value === right[index])
-}
-
-function measuredWinnerProfile(
-  baselineProfile: AgentProfile,
-  surface: ImproveSurface,
-  winner: unknown,
-): AgentProfile | null {
-  if (typeof winner !== 'string') {
-    throw new Error(`measured '${surface}' winner is not a serializable profile surface`)
-  }
-  if (surface !== 'skills') {
-    return applyImprovementWinnerToProfile(baselineProfile, surface, winner)
-  }
-  try {
-    return applyImprovementWinnerToProfile(baselineProfile, surface, winner)
-  } catch {
-    // A skill document is text stored outside the profile; a skill-ref array is
-    // JSON stored on the profile. Only the latter can bind to a profile bundle.
-    return null
-  }
-}
-
-function assertMeasuredCodeBundle(winner: unknown, candidateBundle: AgentCandidateBundle): void {
-  assertCodeSurfaceIdentity(winner)
-  if (candidateBundle.code.kind !== 'git-patch') {
-    throw new Error('code improvement bundle must contain the measured git patch')
   }
-  const bundledSurface: CodeSurface = {
-    ...winner,
-    baseCommit: candidateBundle.code.baseCommit,
-    baseTree: candidateBundle.code.baseTree,
-    candidateTree: candidateBundle.code.candidateTree,
-    patch: {
-      format: candidateBundle.code.patch.format,
-      sha256: candidateBundle.code.patch.artifact.sha256,
-      byteLength: candidateBundle.code.patch.artifact.byteLength,
-    },
-  }
-  assertCodeSurfaceIdentity(bundledSurface)
-  if (codeSurfaceIdentityMaterial(winner) !== codeSurfaceIdentityMaterial(bundledSurface)) {
-    throw new Error('proposal candidate bundle does not match the measured code winner')
-  }
-}
-
-function sealBuiltCandidate(
-  input: AgentCandidateBundleInput | AgentCandidateBundle,
-): AgentCandidateBundle {
-  if (!('digest' in input)) return sealAgentCandidateBundle(input)
-  const parsed = agentCandidateBundleSchema.parse(input)
-  const sealed = sealAgentCandidateBundle(omitTopLevelDigest(parsed))
-  if (sealed.digest !== parsed.digest) {
-    throw new Error('built candidate bundle digest is invalid')
-  }
-  return sealed
 }
 
-/** Validate a review's decision fields and canonical digest. */
-export function verifyAgentImprovementReview(input: unknown): AgentImprovementReview {
-  const review = agentImprovementReviewSchema.parse(input)
-  return verifyCanonicalCandidateDocument(review, 'agent improvement review')
-}
-
-/** Convert agent-eval's paired result into the portable Interface comparison. */
-export function createAgentImprovementMeasuredComparison<TScenario extends Scenario, TArtifact>(
-  options: CreateAgentImprovementMeasuredComparisonOptions<TScenario, TArtifact>,
-): AgentImprovementMeasuredComparison {
-  const { result } = options
-  const baselineProfile = parseExactAgentProfile(
-    options.baselineProfile,
-    'measured baseline profile',
-  )
-  const candidateBundle = sealBuiltCandidate(options.candidateBundle)
-  assertSupportedCandidateBundle(candidateBundle)
-  assertRawMeasuredWinnerBinding({
-    surface: options.measuredSurface,
-    baselineProfile,
-    winner: result.winner.surface,
-    candidateBundle,
-  })
-  const benchmark = candidateBundle.lineage.benchmark
-  if (!benchmark) {
-    throw new Error('agent improvement proposal requires a benchmark and heldout split identity')
-  }
-  const power = result.power
-  if (!power) throw new Error('agent improvement proposal requires heldout power analysis')
-  if (result.provenance.gate.reasons.length === 0) {
-    throw new Error('agent improvement proposal requires measured decision reasons')
-  }
-  const pairs = pairMeasuredCells(
-    result.raw.baselineOnHoldout.cells,
-    result.raw.winnerOnHoldout.cells,
-  )
-  const composite = measuredObjective(
-    {
-      kind: 'objective',
-      name: 'composite',
-      direction: 'higher-is-better',
-      unit: 'score',
-    },
-    pairs,
-    measuredComposite,
-  )
-  assertMeasuredNumber(result.lift, composite.delta, 'heldout lift')
-  assertMeasuredNumber(result.baseline.compositeMean, composite.baseline, 'heldout baseline')
-  assertMeasuredNumber(result.winner.compositeMean, composite.candidate, 'heldout candidate')
-  assertMeasuredNumber(result.provenance.heldOutLift, composite.delta, 'provenance heldout lift')
+function assertActivationTargets(
+  surfaces: readonly AgentImprovementSurface[],
+  experiment: AgentCandidateExperiment,
+  targets: readonly AgentImprovementActivationTarget[],
+): void {
+  const expected = new Set(surfaces)
+  const actual = new Set(targets.map((target) => target.surface))
+  const baselineValues = improvementSurfaceValues(experiment.baseline)
   if (
-    result.gateDecision !== result.provenance.gate.decision ||
-    power.n !== composite.n ||
-    power.confidence !== 0.95
+    targets.some((target) => !target.identity.trim()) ||
+    targets.some(
+      (target) =>
+        target.expectedBaseDigest !==
+        expectedActivationBaseDigest(experiment, target.surface, baselineValues),
+    ) ||
+    expected.size !== actual.size ||
+    [...expected].some((surface) => !actual.has(surface))
   ) {
-    throw new Error('agent improvement measurement sources do not agree')
-  }
-
-  const comparison = immutableCandidateValue({
-    kind: 'agent-improvement-measured-comparison' as const,
-    benchmark,
-    baselineProfileDigest: canonicalCandidateDigest(baselineProfile),
-    candidateBundleDigest: candidateBundle.digest,
-    overall: {
-      name: 'composite' as const,
-      baseline: composite.baseline,
-      candidate: composite.candidate,
-      delta: composite.delta,
-      confidenceInterval: composite.confidenceInterval,
-      n: composite.n,
-      direction: 'higher-is-better' as const,
-      unit: 'score' as const,
-    },
-    objectives: measuredObjectives(pairs),
-    ...(result.winner.label || result.winner.rationale
-      ? {
-          candidate: {
-            ...(result.winner.label ? { label: result.winner.label } : {}),
-            ...(result.winner.rationale ? { rationale: result.winner.rationale } : {}),
-          },
-        }
-      : {}),
-    decision: {
-      outcome: result.gateDecision,
-      reasons: result.provenance.gate.reasons,
-      contributingChecks: result.provenance.gate.contributingGates.map((check) => ({
-        name: check.name,
-        passed: check.passed,
-      })),
-    },
-    power: {
-      sufficient: power.scaleAssumed && !power.underpowered,
-      n: power.n,
-      minimumDetectableDelta: power.mde,
-      confidenceLevel: power.confidence,
-      scaleAssumed: power.scaleAssumed,
-      sharedScorerChannel: power.sharedChannelCaveat !== undefined,
-      reason: power.recommendation,
-    },
-    provenance: {
-      kind: 'agent-eval-loop' as const,
-      schema: result.provenance.schema,
-      runId: result.provenance.runId,
-      recordDigest: canonicalCandidateDigest(result.provenance),
-      baselineContentHash: result.provenance.baselineContentHash,
-      candidateContentHash: result.provenance.winnerContentHash,
-    },
-    diff: result.diff,
-    evaluation: {
-      generationsExplored: result.generationsExplored,
-      durationMs: result.durationMs,
-      totalCostUsd: result.totalCostUsd,
-    },
-    ...(options.metadata ? { metadata: options.metadata } : {}),
-  })
-  return agentImprovementMeasuredComparisonSchema.parse(comparison)
-}
-
-interface MeasuredEvaluationCell {
-  scenarioId: string
-  rep: number
-  judgeScores: Record<
-    string,
-    { composite: number; dimensions: Record<string, number>; failed?: true }
-  >
-  costUsd: number
-  tokenUsage?: { input: number; output: number }
-  durationMs: number
-  error?: string
-}
-
-function measuredObjectives(
-  pairs: ReadonlyArray<readonly [MeasuredEvaluationCell, MeasuredEvaluationCell]>,
-): AgentImprovementMeasuredComparison['objectives'] {
-  const qualityColumns = new Map<string, MeasuredQualityColumn>()
-  for (const [baseline, candidate] of pairs) {
-    for (const cell of [baseline, candidate]) {
-      for (const [objective, score] of Object.entries(cell.judgeScores)) {
-        if (score.failed) continue
-        qualityColumns.set(`objective:${objective}`, {
-          kind: 'objective',
-          name: objective,
-          direction: 'higher-is-better',
-          unit: 'score',
-        })
-        for (const name of Object.keys(score.dimensions)) {
-          qualityColumns.set(`dimension:${objective}:${name}`, {
-            kind: 'dimension',
-            objective,
-            name,
-            direction: 'higher-is-better',
-            unit: 'score',
-          })
-        }
-      }
-    }
+    throw new Error('candidate activation targets must cover exactly the changed surfaces')
   }
-  const objectives: AgentImprovementMeasuredComparison['objectives'] = [
-    ...[...qualityColumns.entries()]
-      .sort(([left], [right]) => left.localeCompare(right))
-      .map(([, column]) =>
-        measuredObjective(column, pairs, (cell) => measuredQuality(cell, column)),
-      ),
-    measuredCostObjective(pairs),
-    measuredObjective(
-      {
-        kind: 'latency',
-        name: 'latency',
-        direction: 'lower-is-better',
-        unit: 'milliseconds',
-      },
-      pairs,
-      (cell) => cell.durationMs,
-    ),
-  ]
-  return objectives
 }
 
-function measuredCostObjective(
-  pairs: ReadonlyArray<readonly [MeasuredEvaluationCell, MeasuredEvaluationCell]>,
-): AgentImprovementMeasuredComparison['objectives'][number] {
-  const cells = pairs.flat()
-  for (const cell of cells) finiteMeasuredValue(cell.costUsd, 'cost:cost')
-  const reported = cells.some(
-    (cell) =>
-      cell.costUsd !== 0 ||
-      (cell.tokenUsage !== undefined && (cell.tokenUsage.input > 0 || cell.tokenUsage.output > 0)),
-  )
-  if (!reported) {
-    return {
-      kind: 'cost',
-      name: 'cost',
-      availability: 'unavailable',
-      reason: 'heldout cells did not report model usage or cost',
-      direction: 'lower-is-better',
-      unit: 'usd',
-    }
+function expectedActivationBaseDigest(
+  experiment: AgentCandidateExperiment,
+  surface: AgentImprovementSurface,
+  baselineValues: Record<AgentImprovementSurface, unknown>,
+): Sha256Digest {
+  if (surface === 'knowledge' && experiment.candidate.knowledge) {
+    return experiment.candidate.knowledge.candidate.baseHash
   }
-  return measuredObjective(
-    {
-      kind: 'cost',
-      name: 'cost',
-      direction: 'lower-is-better',
-      unit: 'usd',
-    },
-    pairs,
-    (cell) => cell.costUsd,
-  )
+  return canonicalCandidateDigest(baselineValues[surface])
 }
 
-function measuredComposite(cell: MeasuredEvaluationCell): number {
-  const values = Object.values(cell.judgeScores)
-    .filter((score) => !score.failed)
-    .map((score) => score.composite)
-    .filter(Number.isFinite)
-  if (values.length === 0) {
-    throw new Error(`heldout cell ${measuredCellKey(cell)} has no successful composite score`)
-  }
-  return measuredMean(values)
+function sameOrderedValues<T>(left: readonly T[], right: readonly T[]): boolean {
+  return left.length === right.length && left.every((value, index) => value === right[index])
 }
 
-function pairMeasuredCells(
-  baselineCells: readonly MeasuredEvaluationCell[],
-  candidateCells: readonly MeasuredEvaluationCell[],
-): Array<readonly [MeasuredEvaluationCell, MeasuredEvaluationCell]> {
-  type MeasuredArmRow = PairedArmRow & { cell: MeasuredEvaluationCell }
-  const rows: MeasuredArmRow[] = [
-    ...baselineCells
-      .filter((cell) => !cell.error)
-      .map((cell) => ({
-        pairKey: cell.scenarioId,
-        repKey: String(cell.rep),
-        arm: 'baseline',
-        cell,
-      })),
-    ...candidateCells
-      .filter((cell) => !cell.error)
-      .map((cell) => ({
-        pairKey: cell.scenarioId,
-        repKey: String(cell.rep),
-        arm: 'candidate',
-        cell,
-      })),
-  ]
-  const paired = pairArms(rows, { baselineArm: 'baseline', treatmentArm: 'candidate' })
-  if (
-    paired.pairs.length === 0 ||
-    paired.unpairedBaseline.length > 0 ||
-    paired.unpairedTreatment.length > 0
-  ) {
-    throw new Error('measured objectives require the same non-empty paired heldout cells')
-  }
-  return paired.pairs.map(
-    (pair) =>
-      [(pair.baseline as MeasuredArmRow).cell, (pair.treatment as MeasuredArmRow).cell] as const,
+function assertNoJudgeDerivedProposalFindings(
+  findings: AgentImprovementProposal['findings'],
+): void {
+  const leaked = findings.filter((finding) => finding.derived_from_judge === true)
+  if (leaked.length === 0) return
+  const identifiers = leaked.map((finding) =>
+    typeof finding.finding_id === 'string' ? finding.finding_id : '<unknown>',
+  )
+  throw new Error(
+    'agent improvement proposal findings: judge-derived findings cannot steer an improvement: ' +
+      `[${identifiers.join(', ')}]`,
   )
-}
-
-type MeasuredObjectiveIdentity =
-  | {
-      kind: 'objective'
-      name: string
-      direction: 'higher-is-better'
-      unit: 'score'
-    }
-  | {
-      kind: 'dimension'
-      objective: string
-      name: string
-      direction: 'higher-is-better'
-      unit: 'score'
-    }
-  | {
-      kind: 'cost'
-      name: 'cost'
-      direction: 'lower-is-better'
-      unit: 'usd'
-    }
-  | {
-      kind: 'latency'
-      name: 'latency'
-      direction: 'lower-is-better'
-      unit: 'milliseconds'
-    }
-
-type MeasuredQualityColumn = Extract<MeasuredObjectiveIdentity, { kind: 'objective' | 'dimension' }>
-
-function measuredObjective(
-  identity: MeasuredObjectiveIdentity,
-  pairs: ReadonlyArray<readonly [MeasuredEvaluationCell, MeasuredEvaluationCell]>,
-  value: (cell: MeasuredEvaluationCell) => number,
-): Extract<AgentImprovementMeasuredComparison['objectives'][number], { availability: 'measured' }> {
-  const key = measuredObjectiveKey(identity)
-  const baseline = pairs.map(([cell]) => finiteMeasuredValue(value(cell), key))
-  const candidate = pairs.map(([, cell]) => finiteMeasuredValue(value(cell), key))
-  const interval = pairedBootstrap(baseline, candidate, {
-    confidence: 0.95,
-    resamples: 2_000,
-    statistic: 'mean',
-    seed: measuredSeed(key),
-  })
-  const baselineMean = measuredMean(baseline)
-  const candidateMean = measuredMean(candidate)
-  const delta = interval.mean
-  const estimate = {
-    availability: 'measured',
-    baseline: baselineMean,
-    candidate: candidateMean,
-    delta,
-    confidenceInterval: {
-      level: interval.confidence,
-      lower: Math.min(interval.low, delta),
-      upper: Math.max(interval.high, delta),
-      method: 'paired-bootstrap',
-      statistic: 'mean',
-      resamples: interval.resamples,
-    },
-    n: interval.n,
-  } as const
-  return { ...identity, ...estimate }
-}
-
-function measuredQuality(cell: MeasuredEvaluationCell, column: MeasuredQualityColumn): number {
-  const objective = column.kind === 'objective' ? column.name : column.objective
-  const score = cell.judgeScores[objective]
-  if (!score || score.failed) {
-    throw new Error(
-      `heldout cell ${measuredCellKey(cell)} is missing measured objective '${objective}'`,
-    )
-  }
-  const value = column.kind === 'objective' ? score.composite : score.dimensions[column.name]
-  if (value === undefined) {
-    throw new Error(
-      `heldout cell ${measuredCellKey(cell)} is missing '${objective}' dimension '${column.name}'`,
-    )
-  }
-  return finiteMeasuredValue(value, measuredObjectiveKey(column))
-}
-
-function measuredObjectiveKey(identity: MeasuredObjectiveIdentity): string {
-  return identity.kind === 'dimension'
-    ? `${identity.kind}:${identity.objective}:${identity.name}`
-    : `${identity.kind}:${identity.name}`
-}
-
-function measuredCellKey(cell: Pick<MeasuredEvaluationCell, 'scenarioId' | 'rep'>): string {
-  return `${cell.scenarioId}:${cell.rep}`
-}
-
-function finiteMeasuredValue(value: number, name: string): number {
-  if (!Number.isFinite(value)) throw new Error(`measured objective '${name}' is not finite`)
-  return value
-}
-
-function measuredMean(values: readonly number[]): number {
-  if (values.length === 0) throw new Error('measured objective has no paired values')
-  return values.reduce((sum, value) => sum + value, 0) / values.length
-}
-
-function measuredSeed(value: string): number {
-  let seed = 0x811c9dc5
-  for (const byte of Buffer.from(value, 'utf8')) {
-    seed = Math.imul(seed ^ byte, 0x01000193) >>> 0
-  }
-  return seed
-}
-
-function assertMeasuredNumber(actual: number, expected: number, name: string): void {
-  const tolerance = Number.EPSILON * Math.max(1, Math.abs(actual), Math.abs(expected)) * 8
-  if (
-    !Number.isFinite(actual) ||
-    !Number.isFinite(expected) ||
-    Math.abs(actual - expected) > tolerance
-  ) {
-    throw new Error(`${name} does not agree across the measured comparison`)
-  }
 }
diff --git a/src/intelligence/index.ts b/src/intelligence/index.ts
index 9d4122f8..e4e0a73e 100644
--- a/src/intelligence/index.ts
+++ b/src/intelligence/index.ts
@@ -3,8 +3,9 @@
  * Tangle Intelligence SDK — trace capture plus reviewable improvement.
  *
  * The client keeps live-agent trace delivery best-effort. The separate
- * improvement-cycle exports analyze completed traces, measure one candidate,
- * bind human review, and execute only an approved immutable bundle.
+ * improvement-cycle exports analyze completed traces, run a signed baseline
+ * versus candidate experiment, bind review to its result, and activate only
+ * the exact measured candidate.
  *
  *   1. OBSERVE — wrap a generic agent and export one trace span per call to
  *      Tangle Intelligence, swallowing every export failure so a live agent
@@ -102,20 +103,27 @@ export {
   resolveEffort,
 } from './effort'
 export type {
+  AgentCandidateExperimentCellPlacement,
+  CreateAgentImprovementActivationOptions,
   CreateAgentImprovementProposalOptions,
-  ExecuteApprovedAgentCandidateOptions,
-  ExecuteApprovedAgentCandidateResult,
+  ExecuteAgentCandidateExperimentCellOptions,
   ProposeAgentImprovementOptions,
   ProposeAgentImprovementResult,
   ReviewAgentImprovementInput,
+  RunAgentCandidateExperimentOptions,
+  RunAgentCandidateExperimentResult,
   VerifyCandidateExecutionEvidenceOptions,
 } from './improvement-cycle'
 export {
+  AgentCandidateExperimentCellExecutionError,
+  createAgentImprovementActivation,
   createAgentImprovementMeasuredComparison,
   createAgentImprovementProposal,
-  executeApprovedAgentCandidate,
+  executeAgentCandidateExperimentCell,
   proposeAgentImprovement,
   reviewAgentImprovementProposal,
+  runAgentCandidateExperiment,
+  verifyAgentImprovementActivation,
   verifyAgentImprovementProposal,
   verifyAgentImprovementReview,
   verifyCandidateExecutionEvidence,
@@ -128,13 +136,13 @@ export {
   composeCertifiedProfileFromWire,
 } from './resolver'
 export type {
-  CreateSandboxApprovedCandidateExecutorOptions,
-  SandboxApprovedCandidateExecution,
-  SandboxApprovedCandidateExecutor,
+  CreateSandboxCandidateExperimentExecutorOptions,
+  SandboxCandidateExperimentExecution,
+  SandboxCandidateExperimentExecutor,
 } from './sandbox-approved-candidate'
 export {
-  createSandboxApprovedCandidateExecutor,
-  sandboxApprovedCandidateExecutionSupport,
+  createSandboxCandidateExperimentExecutor,
+  sandboxCandidateExperimentExecutionSupport,
 } from './sandbox-approved-candidate'
 export type {
   AppliedIntelligence,
@@ -599,15 +607,20 @@ export function createIntelligenceClient(config: IntelligenceConfig): Intelligen
           : {}),
         ...(record.candidateExecution
           ? {
-              'tangle.candidate.proposal_digest': record.candidateExecution.proposalDigest,
-              'tangle.candidate.review_digest': record.candidateExecution.reviewDigest,
               'tangle.candidate.bundle_digest': record.candidateExecution.receipt.bundleDigest,
-              'tangle.candidate.execution_id': record.candidateExecution.executionId,
+              'tangle.candidate.experiment_digest':
+                record.candidateExecution.materializationReceipt.executionPlan.material.runCell
+                  .experimentDigest,
+              'tangle.candidate.execution_id':
+                record.candidateExecution.materializationReceipt.executionPlan.material.executionId,
               'tangle.candidate.execution_plan_digest':
                 record.candidateExecution.receipt.executionPlanDigest,
               'tangle.candidate.materialization_receipt_digest':
                 record.candidateExecution.receipt.materializationReceiptDigest,
-              'tangle.candidate.succeeded': record.candidateExecution.succeeded,
+              'tangle.candidate.succeeded':
+                record.candidateExecution.receipt.termination.kind === 'exit' &&
+                record.candidateExecution.receipt.termination.exitCode === 0 &&
+                record.candidateExecution.receipt.benchmarkResult.material.passed,
               'tangle.candidate.run_receipt_digest': record.candidateExecution.receipt.digest,
             }
           : {}),
diff --git a/src/intelligence/sandbox-approved-candidate.ts b/src/intelligence/sandbox-approved-candidate.ts
index b0fd4ecd..738fbce7 100644
--- a/src/intelligence/sandbox-approved-candidate.ts
+++ b/src/intelligence/sandbox-approved-candidate.ts
@@ -1,10 +1,8 @@
 import { posix } from 'node:path'
-
 import type { TraceStore } from '@tangle-network/agent-eval'
+import type { CandidateExperimentExecutionInput } from '@tangle-network/agent-eval/contract'
 import type {
   AgentCandidateTermination,
-  AgentImprovementProposal,
-  AgentImprovementReview,
   CandidateExecutionEvidence,
   Sha256Digest,
 } from '@tangle-network/agent-interface'
@@ -35,20 +33,34 @@ import type {
   AgentCandidateOutputArtifactPort,
 } from '../candidate-execution/types'
 import { AGENT_CANDIDATE_EXECUTION_SUPPORT } from '../candidate-execution/verify'
-import { executeApprovedAgentCandidate } from './improvement-cycle'
+import {
+  type AgentCandidateExperimentCellPlacement,
+  executeAgentCandidateExperimentCell,
+} from './improvement-cycle'
 
 type SandboxClientPort = Pick<SandboxClient, 'create' | 'get' | 'list'>
 
-type SandboxExecutorOptions = NonNullable<CreateSandboxApprovedCandidateExecutorOptions['sandbox']>
+type SandboxExecutorOptions = NonNullable<
+  CreateSandboxCandidateExperimentExecutorOptions['sandbox']
+>
 
 interface SandboxRunState {
   sandbox: SandboxInstance
   output?: Uint8Array
   termination?: AgentCandidateTermination
+  trace?: {
+    runId: string
+    scenarioId: string
+    startedAt: number
+    deadlineAtMs: number
+    timeoutMs: number
+    tags: Record<string, string>
+    written: boolean
+  }
 }
 
 /** Declares the exact candidate surfaces the sandbox executor can run. */
-export const sandboxApprovedCandidateExecutionSupport = Object.freeze({
+export const sandboxCandidateExperimentExecutionSupport = Object.freeze({
   outcomes: Object.freeze(['output'] as const),
   outputMediaTypes: Object.freeze(['text/*', 'application/json', '*+json'] as const),
   code: Object.freeze(['disabled'] as const),
@@ -62,17 +74,13 @@ export const sandboxApprovedCandidateExecutionSupport = Object.freeze({
   }),
 })
 
-export interface CreateSandboxApprovedCandidateExecutorOptions {
+export interface CreateSandboxCandidateExperimentExecutorOptions {
   client: SandboxClientPort
   ports: AgentCandidateExecutionPorts
   grader: AgentCandidateBenchmarkGraderPort
   outputArtifacts: AgentCandidateOutputArtifactPort
   traceStore: TraceStore
   claimStore: AgentCandidateExecutionClaimStore
-  authorizeReview: (
-    review: AgentImprovementReview,
-    proposal: AgentImprovementProposal,
-  ) => boolean | Promise<boolean>
   sandbox?: {
     teamId?: string
     resources?: SandboxResources
@@ -83,32 +91,38 @@ export interface CreateSandboxApprovedCandidateExecutorOptions {
   resultTimeoutMs?: number
 }
 
-export interface SandboxApprovedCandidateExecution {
-  proposal: AgentImprovementProposal
-  review: AgentImprovementReview
-  task: Parameters<typeof executeApprovedAgentCandidate>[0]['task']
+export interface SandboxCandidateExperimentExecution extends CandidateExperimentExecutionInput {
+  executionId: string
+  attempt?: number
+  executionRoots: AgentCandidateExperimentCellPlacement['executionRoots']
+  stagingRoots: AgentCandidateExperimentCellPlacement['stagingRoots']
   preparation?: PrepareAgentCandidateExecutionOptions
 }
 
-export interface SandboxApprovedCandidateExecutor {
+export interface SandboxCandidateExperimentExecutor {
   /** The same port is usable by Runtime's expired-claim recovery path. */
   readonly executor: AgentCandidateExecutorPort
-  execute(input: SandboxApprovedCandidateExecution): Promise<CandidateExecutionEvidence>
+  execute(input: SandboxCandidateExperimentExecution): Promise<CandidateExecutionEvidence>
 }
 
-/** Compose approved-candidate execution directly onto fresh Tangle sandboxes. */
-export function createSandboxApprovedCandidateExecutor(
-  options: CreateSandboxApprovedCandidateExecutorOptions,
-): SandboxApprovedCandidateExecutor {
-  const executor = new SandboxAgentCandidateExecutor(options.client, options.sandbox)
+/** Execute one signed experiment cell inside a fresh Tangle sandbox. */
+export function createSandboxCandidateExperimentExecutor(
+  options: CreateSandboxCandidateExperimentExecutorOptions,
+): SandboxCandidateExperimentExecutor {
+  const executor = new SandboxAgentCandidateExecutor(
+    options.client,
+    options.sandbox,
+    options.resultTimeoutMs,
+  )
   return Object.freeze({
     executor,
-    async execute(input: SandboxApprovedCandidateExecution): Promise<CandidateExecutionEvidence> {
-      const result = await executeApprovedAgentCandidate({
-        proposal: input.proposal,
-        review: input.review,
-        authorizeReview: options.authorizeReview,
-        task: input.task,
+    async execute(input: SandboxCandidateExperimentExecution): Promise<CandidateExecutionEvidence> {
+      return await executeAgentCandidateExperimentCell({
+        ...input,
+        attempt: input.attempt ?? 1,
+        executionId: input.executionId,
+        executionRoots: input.executionRoots,
+        stagingRoots: input.stagingRoots,
         ports: options.ports,
         ...(input.preparation ? { preparation: input.preparation } : {}),
         execution: {
@@ -125,25 +139,6 @@ export function createSandboxApprovedCandidateExecutor(
             : { resultTimeoutMs: options.resultTimeoutMs }),
         },
       })
-      if (result.finalization.succeeded) {
-        if (!result.evidence) throw new Error('approved candidate execution returned no evidence')
-        try {
-          return result.evidence
-        } finally {
-          await executor.delete({
-            executionId: result.evidence.executionId,
-            executionPlanDigest: result.evidence.receipt.executionPlanDigest,
-          })
-        }
-      }
-      try {
-        throw new Error(`approved candidate execution failed: ${result.finalization.reason}`)
-      } finally {
-        await executor.delete({
-          executionId: result.finalization.partial.executionId,
-          executionPlanDigest: result.finalization.partial.executionPlanDigest,
-        })
-      }
     },
   })
 }
@@ -154,6 +149,7 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
   constructor(
     private readonly client: SandboxClientPort,
     private readonly options: SandboxExecutorOptions = {},
+    private readonly captureTimeoutMs = 120_000,
   ) {}
 
   async execute(
@@ -161,7 +157,7 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
     context: Parameters<AgentCandidateExecutorPort['execute']>[1],
   ) {
     assertSupportedRequest(request)
-    const outcome = request.executionPlan.value.material.task.outcome
+    const outcome = request.benchmark.task.outcome
     if (outcome.kind !== 'output') throw new Error('sandbox executor requires an output task')
     const key = executionKey(request.executionId, request.executionPlan.value.digest)
     const sandbox = await this.client.create(sandboxCreateOptions(request, this.options), {
@@ -176,6 +172,15 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
     await materializeRequest(sandbox, request)
     const launch = exactLaunch(request)
     const startedAt = Date.now()
+    state.trace = {
+      runId: request.trace.runId,
+      scenarioId: request.benchmark.task.scenario.id,
+      startedAt,
+      deadlineAtMs: context.deadlineAtMs,
+      timeoutMs: request.hardLimits.timeoutMs,
+      tags: { ...request.trace.tags, sandboxId: sandbox.id },
+      written: false,
+    }
     const process = await sandbox.process.spawnExact(launch.executable, launch.args, {
       cwd: request.launch.cwd,
       env: { ...request.launch.env },
@@ -207,14 +212,7 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
       }
       state.output = await awaitOutput(outputPromise, context.signal, context.deadlineAtMs)
       state.termination = cancellationTermination ?? processTermination(status)
-      await context.traceStore.appendRun({
-        runId: request.trace.runId,
-        scenarioId: request.executionPlan.value.material.task.taskId,
-        startedAt,
-        endedAt: Date.now(),
-        status: status.exitCode === 0 ? 'completed' : 'failed',
-        tags: { ...request.trace.tags, sandboxId: sandbox.id },
-      })
+      await appendSandboxTrace(state, context.traceStore, status)
       return { executionId: request.executionId, termination: state.termination }
     } finally {
       context.signal.removeEventListener('abort', cancel)
@@ -242,6 +240,13 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
     if (remaining.some((entry) => entry.running)) {
       throw new Error('candidate sandbox process termination is not proven')
     }
+    const state = this.states.get(executionKey(request.executionId, request.executionPlanDigest))
+    if (state) {
+      if (context.reason === 'timeout') {
+        state.termination = { kind: 'timeout', timeoutMs: state.trace?.timeoutMs ?? 0 }
+      }
+      await appendSandboxTrace(state, context.traceStore, remaining[0])
+    }
     return { stopped: true }
   }
 
@@ -259,16 +264,18 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
     let output = state?.output
     let termination = state?.termination
     const status = statuses[0]
-    if (status && (!output || !termination)) {
+    if (status && !output) {
       const process = await sandbox.process.get(status.pid)
       if (!process) throw new Error('candidate sandbox lost its captured process handle')
       const expected = sandboxOutputSpec(sandbox)
       output = await awaitOutput(
         collectOutput(process, expected.maxBytes),
         context.signal,
-        Date.now() + (this.options.createTimeoutMs ?? 120_000),
+        Date.now() + this.captureTimeoutMs,
       )
-      termination = processTermination(await process.status())
+    }
+    if (status && !termination) {
+      termination = processTermination(status)
     }
     context.signal.throwIfAborted()
     const evidence = canonicalCandidateBytes({
@@ -293,15 +300,21 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
     }
   }
 
-  async delete(request: AgentCandidateExecutorStopRequest): Promise<void> {
+  async dispose(
+    request: AgentCandidateExecutorStopRequest,
+    context: { signal: AbortSignal },
+  ): Promise<{ readonly disposed: true }> {
+    context.signal.throwIfAborted()
     const sandbox = await this.resolve(request, true)
-    if (!sandbox) return
+    if (!sandbox) return { disposed: true }
     try {
       await sandbox.delete()
     } catch (error) {
       if (await this.client.get(sandbox.id)) throw error
     }
+    context.signal.throwIfAborted()
     this.states.delete(executionKey(request.executionId, request.executionPlanDigest))
+    return { disposed: true }
   }
 
   private async resolve(
@@ -337,8 +350,30 @@ class SandboxAgentCandidateExecutor implements AgentCandidateExecutorPort {
   }
 }
 
+async function appendSandboxTrace(
+  state: SandboxRunState,
+  traceStore: TraceStore,
+  status?: ProcessStatus,
+): Promise<void> {
+  const trace = state.trace
+  if (!trace || trace.written) return
+  const endedAt = Math.max(trace.startedAt, Math.min(Date.now(), trace.deadlineAtMs))
+  await traceStore.appendRun({
+    runId: trace.runId,
+    scenarioId: trace.scenarioId,
+    startedAt: trace.startedAt,
+    endedAt,
+    status:
+      state.termination?.kind !== 'timeout' && status?.running === false && status.exitCode === 0
+        ? 'completed'
+        : 'failed',
+    tags: trace.tags,
+  })
+  trace.written = true
+}
+
 function assertSupportedRequest(request: AgentCandidateExecutorRequest): void {
-  if (request.executionPlan.value.material.task.outcome.kind !== 'output') {
+  if (request.benchmark.task.outcome.kind !== 'output') {
     throw new Error('sandbox candidate executor does not support workspace outcomes; use Pier')
   }
   if (request.executionPlan.value.material.codeKind !== 'disabled' || request.inputs.candidate) {
@@ -347,7 +382,7 @@ function assertSupportedRequest(request: AgentCandidateExecutorRequest): void {
   if (request.memory.mode !== 'disabled') {
     throw new Error('sandbox candidate executor does not support isolated memory')
   }
-  const mediaType = request.executionPlan.value.material.task.outcome.mediaType.toLowerCase()
+  const mediaType = request.benchmark.task.outcome.mediaType.toLowerCase()
   if (
     !(
       mediaType.startsWith('text/') ||
@@ -395,8 +430,13 @@ function sandboxCreateOptions(
       executionId: request.executionId,
       executionPlanDigest: request.executionPlan.value.digest,
       outputMediaType:
-        material.task.outcome.kind === 'output' ? material.task.outcome.mediaType : '',
-      outputMaxBytes: material.task.outcome.kind === 'output' ? material.task.outcome.maxBytes : 0,
+        request.benchmark.task.outcome.kind === 'output'
+          ? request.benchmark.task.outcome.mediaType
+          : '',
+      outputMaxBytes:
+        request.benchmark.task.outcome.kind === 'output'
+          ? request.benchmark.task.outcome.maxBytes
+          : 0,
     },
     ...(options.teamId ? { teamId: options.teamId } : {}),
     ...(options.resources ? { resources: options.resources } : {}),
diff --git a/src/intelligence/with-intelligence.test.ts b/src/intelligence/with-intelligence.test.ts
index 4c78d8ce..a0a9d04d 100644
--- a/src/intelligence/with-intelligence.test.ts
+++ b/src/intelligence/with-intelligence.test.ts
@@ -247,20 +247,24 @@ describe('withIntelligence — SEND (a typed RunRecord to /v1/otlp)', () => {
             ],
             candidateExecution: {
               kind: 'agent-candidate-execution-evidence',
-              proposalDigest: `sha256:${'1'.repeat(64)}`,
-              reviewDigest: `sha256:${'2'.repeat(64)}`,
-              executionId: 'candidate-execution-1',
-              succeeded: true,
-              materializationReceipt: {} as CandidateExecutionEvidence['materializationReceipt'],
-              profileActivation: {} as CandidateExecutionEvidence['profileActivation'],
+              materializationReceipt: {
+                executionPlan: {
+                  material: {
+                    executionId: 'candidate-execution-1',
+                    runCell: { experimentDigest: `sha256:${'1'.repeat(64)}` },
+                  },
+                },
+              },
               receipt: {
                 bundleDigest: `sha256:${'3'.repeat(64)}`,
                 executionPlanDigest: `sha256:${'4'.repeat(64)}`,
                 materializationReceiptDigest: `sha256:${'5'.repeat(64)}`,
+                termination: { kind: 'exit', exitCode: 0 },
+                benchmarkResult: { material: { passed: true } },
                 digest: `sha256:${'6'.repeat(64)}`,
-              } as CandidateExecutionEvidence['receipt'],
+              },
               digest: `sha256:${'7'.repeat(64)}`,
-            },
+            } as CandidateExecutionEvidence,
           })
           return 'answer'
         },
@@ -309,7 +313,7 @@ describe('withIntelligence — SEND (a typed RunRecord to /v1/otlp)', () => {
       expect(attrs['tool.name']).toBe('mcp__linear__linear_graphql')
       expect(String(attrs['tool.input'])).toContain(longInput)
       expect(attrs['tangle.candidate.execution_id']).toBe('candidate-execution-1')
-      expect(attrs['tangle.candidate.proposal_digest']).toBe(`sha256:${'1'.repeat(64)}`)
+      expect(attrs['tangle.candidate.experiment_digest']).toBe(`sha256:${'1'.repeat(64)}`)
       expect(attrs['tangle.candidate.run_receipt_digest']).toBe(`sha256:${'6'.repeat(64)}`)
     } finally {
       vi.useRealTimers()
diff --git a/src/knowledge/improvement-job.ts b/src/knowledge/improvement-job.ts
index 13b2e677..80bf8782 100644
--- a/src/knowledge/improvement-job.ts
+++ b/src/knowledge/improvement-job.ts
@@ -2,6 +2,7 @@ import { realpath } from 'node:fs/promises'
 import {
   type AgentCandidateCapturedArtifact,
   type AgentCandidateKnowledge,
+  type AgentImprovementActivation,
   type AgentImprovementProposal,
   type AgentImprovementReview,
   agentCandidateKnowledgeSchema,
@@ -9,6 +10,7 @@ import {
 import {
   type BuildEvalKnowledgeBundleOptions,
   evaluateKnowledgeBaseReadiness,
+  fromAgentCandidateKnowledgeRef,
   hashKnowledgeBase,
   improveKnowledgeBase,
   type KnowledgeBaseQualityOptions,
@@ -18,6 +20,7 @@ import {
   type KnowledgeReadinessSpec,
   knowledgeImprovementCandidateRef,
   promoteKnowledgeCandidate,
+  toAgentCandidateKnowledgeRef,
   withKnowledgeImprovementCandidate,
 } from '@tangle-network/agent-knowledge'
 import { canonicalCandidateBytes, embeddedCandidateArtifact } from '../candidate-execution/digest'
@@ -25,6 +28,7 @@ import { persistCandidateOutputArtifact } from '../candidate-execution/output-ar
 import type { AgentCandidateOutputArtifactPort } from '../candidate-execution/types'
 import { captureAgentCandidateWorkspace } from '../candidate-execution/workspace-archive'
 import {
+  verifyAgentImprovementActivation,
   verifyAgentImprovementProposal,
   verifyAgentImprovementReview,
 } from '../intelligence/improvement-cycle'
@@ -68,9 +72,11 @@ export interface RunKnowledgeImprovementJobOptions
 export interface ApprovedKnowledgeImprovementCandidate {
   proposal: AgentImprovementProposal
   review: AgentImprovementReview
-  authorizeReview: (
-    review: AgentImprovementReview,
+  activation: AgentImprovementActivation
+  authorizeActivation: (
+    activation: AgentImprovementActivation,
     proposal: AgentImprovementProposal,
+    review: AgentImprovementReview,
   ) => boolean | Promise<boolean>
 }
 
@@ -208,14 +214,8 @@ export async function runKnowledgeImprovementJob(
       ...(knowledgeOptions.onState ? { onState: knowledgeOptions.onState } : {}),
     })
     knowledgeOptions.signal?.throwIfAborted()
-    const promotedHash = knowledgeDigest(
-      await hashKnowledgeBase(options.root),
-      'promoted knowledge base',
-    )
-    if (
-      !resolvedImprovement.promoted ||
-      promotedHash !== approvedKnowledge.candidate.candidateHash
-    ) {
+    const promotedHash = await hashKnowledgeBase(options.root)
+    if (!resolvedImprovement.promoted || promotedHash !== candidate.candidateHash) {
       throw new Error('knowledge promotion did not activate the approved snapshot bytes')
     }
     candidateKnowledge = approvedKnowledge
@@ -223,7 +223,7 @@ export async function runKnowledgeImprovementJob(
     resolvedImprovement = await improveKnowledgeBase({
       ...knowledgeOptions,
       updateKnowledge: instrumentedUpdateKnowledge,
-    } as KnowledgeImprovementOptions)
+    })
     if (resolvedImprovement.candidate) {
       candidateKnowledge = await freezeKnowledgeCandidate(
         options.root,
@@ -261,7 +261,7 @@ async function freezeKnowledgeCandidate(
   const candidate = knowledgeImprovementCandidateRef(improvement)
   return withKnowledgeImprovementCandidate({ root, candidate }, async (resolved) => {
     const executionId = `knowledge-${candidate.candidateId}`
-    const candidateRef = interfaceKnowledgeCandidateRef(candidate)
+    const candidateRef = toAgentCandidateKnowledgeRef(candidate)
     const captured = await captureAgentCandidateWorkspace(await realpath(resolved.root), {
       ...(artifacts
         ? {
@@ -313,74 +313,34 @@ async function approvedKnowledgeCandidate(
 ): Promise<AgentCandidateKnowledge> {
   const proposal = verifyAgentImprovementProposal(approval.proposal)
   const review = verifyAgentImprovementReview(approval.review)
+  const activation = verifyAgentImprovementActivation({
+    proposal,
+    review,
+    activation: approval.activation,
+  })
+  const candidateBundle = proposal.evaluation.experiment.candidate
+  const knowledgeTarget = activation.targets.find((target) => target.surface === 'knowledge')
   if (
     proposal.changedSurfaces.length !== 1 ||
     proposal.changedSurfaces[0] !== 'knowledge' ||
-    !proposal.candidateBundle.knowledge ||
+    !candidateBundle.knowledge ||
+    !knowledgeTarget ||
+    knowledgeTarget.expectedBaseDigest !== candidateBundle.knowledge.candidate.baseHash ||
     review.decision !== 'approve' ||
-    review.proposalDigest !== proposal.digest ||
-    review.candidateBundleDigest !== proposal.candidateBundle.digest
+    review.proposalDigest !== proposal.digest
   ) {
     throw new Error('knowledge promotion requires an approved exact knowledge proposal')
   }
-  if (!(await approval.authorizeReview(review, proposal))) {
-    throw new Error('knowledge candidate approval was not authorized')
-  }
-  return proposal.candidateBundle.knowledge
-}
-
-function interfaceKnowledgeCandidateRef(
-  candidate: KnowledgeImprovementCandidateRef,
-): AgentCandidateKnowledge['candidate'] {
-  return {
-    kind: 'knowledge-improvement-candidate',
-    runId: candidate.runId,
-    candidateId: candidate.candidateId,
-    goalHash: knowledgeDigest(candidate.goalHash, 'knowledge candidate goal'),
-    baseHash: knowledgeDigest(candidate.baseHash, 'knowledge candidate base'),
-    candidateHash: knowledgeDigest(candidate.candidateHash, 'knowledge candidate snapshot'),
-    evidenceHash: knowledgeDigest(candidate.evidenceHash, 'knowledge candidate evidence'),
-    promotionPlanHash: knowledgeDigest(
-      candidate.promotionPlanHash,
-      'knowledge candidate promotion plan',
-    ),
+  if (!(await approval.authorizeActivation(activation, proposal, review))) {
+    throw new Error('knowledge candidate activation was not authorized')
   }
+  return candidateBundle.knowledge
 }
 
 function agentKnowledgeCandidateRef(
   knowledge: AgentCandidateKnowledge,
 ): KnowledgeImprovementCandidateRef {
-  return {
-    kind: 'knowledge-improvement-candidate',
-    runId: knowledge.candidate.runId,
-    candidateId: knowledge.candidate.candidateId,
-    goalHash: rawKnowledgeDigest(knowledge.candidate.goalHash, 'knowledge candidate goal'),
-    baseHash: rawKnowledgeDigest(knowledge.candidate.baseHash, 'knowledge candidate base'),
-    candidateHash: rawKnowledgeDigest(
-      knowledge.candidate.candidateHash,
-      'knowledge candidate snapshot',
-    ),
-    evidenceHash: rawKnowledgeDigest(
-      knowledge.candidate.evidenceHash,
-      'knowledge candidate evidence',
-    ),
-    promotionPlanHash: rawKnowledgeDigest(
-      knowledge.candidate.promotionPlanHash,
-      'knowledge candidate promotion plan',
-    ),
-  }
-}
-
-function knowledgeDigest(value: string, label: string): `sha256:${string}` {
-  const digest = value.startsWith('sha256:') ? value : `sha256:${value}`
-  if (!/^sha256:[a-f0-9]{64}$/.test(digest)) {
-    throw new Error(`${label} is not a SHA-256 digest`)
-  }
-  return digest as `sha256:${string}`
-}
-
-function rawKnowledgeDigest(value: string, label: string): string {
-  return knowledgeDigest(value, label).slice('sha256:'.length)
+  return fromAgentCandidateKnowledgeRef(knowledge.candidate)
 }
 
 function emptySpent(): KnowledgeImprovementJobMeasurement['supervisedSpent'] {
diff --git a/src/mcp/worktree-harness.ts b/src/mcp/worktree-harness.ts
index 4a5ab50c..7e1b22a0 100644
--- a/src/mcp/worktree-harness.ts
+++ b/src/mcp/worktree-harness.ts
@@ -218,6 +218,7 @@ export async function runWorktreeHarness(
 
   try {
     assertSupportedWorktreeProfile(opts.profile, opts.harness)
+    assertSafeProfileResourcePaths(opts.profile)
     const resourceInstructions = resolveResourceInstructions(opts.profile)
     const workspaceProfile = materializationOnlyProfile(opts.profile)
     const plan = materializeProfile(workspaceProfile, opts.harness)
@@ -229,7 +230,7 @@ export async function runWorktreeHarness(
       )
     }
     assertSafeMaterializedPaths(plan)
-    const applied = applyWorkspacePlan(plan, worktree.path)
+    const applied = applyWorkspacePlan(plan, worktree.path, { existingFiles: 'reject' })
     if (applied.unsupported.length > 0) {
       throw new Error('runWorktreeHarness: applied profile unexpectedly retained unsupported rows')
     }
@@ -437,6 +438,16 @@ function assertSafeMaterializedPaths(plan: WorkspacePlan): void {
   }
 }
 
+function assertSafeProfileResourcePaths(profile: AgentProfile): void {
+  for (const file of profile.resources?.files ?? []) {
+    if (file.path.split('/').some(isGitMetadataSegment)) {
+      throw new Error(
+        `runWorktreeHarness: profile file cannot target reserved Git metadata: ${file.path}`,
+      )
+    }
+  }
+}
+
 function isGitMetadataSegment(segment: string): boolean {
   const windowsCanonical = segment.replace(/[ .]+$/u, '').toLowerCase()
   return windowsCanonical === '.git' || windowsCanonical.startsWith('.git:')
diff --git a/tests/candidate-bundle-builder.test.ts b/tests/candidate-bundle-builder.test.ts
index b1e5f048..bf1266ad 100644
--- a/tests/candidate-bundle-builder.test.ts
+++ b/tests/candidate-bundle-builder.test.ts
@@ -3,7 +3,6 @@ import { mkdtempSync, rmSync, writeFileSync } from 'node:fs'
 import { tmpdir } from 'node:os'
 import { join } from 'node:path'
 
-import { canonicalJson } from '@tangle-network/agent-eval'
 import { gitWorktreeAdapter } from '@tangle-network/agent-eval/campaign'
 import type {
   AgentCandidateArtifactRef,
@@ -98,16 +97,6 @@ describe('public agent candidate bundle builder', () => {
         evaluation: knowledgeEvaluation,
       },
       memory: { mode: 'isolated', scope: 'task', seed: memorySeed },
-      lineage: {
-        source: 'compound',
-        parentDigests: [
-          fixture.bundle.lineage.parentDigests?.[0] ?? candidateSha('e'),
-          candidateSha('9'),
-        ],
-        runIds: ['optimizer-run-1'],
-        benchmark: fixture.bundle.lineage.benchmark,
-        spend: fixture.bundle.lineage.spend,
-      },
     }
 
     try {
@@ -134,9 +123,7 @@ describe('public agent candidate bundle builder', () => {
         name: 'review/SKILL.md',
         sha256: expect.stringMatching(/^sha256:[a-f0-9]{64}$/),
       })
-      expect(first.lineage.profileDiffIds).toEqual([
-        `sha256:${createHash('sha256').update(canonicalJson(diff)).digest('hex')}`,
-      ])
+      expect(first.knowledge).toEqual(input.knowledge)
 
       const verified = await verifyAgentCandidateBundle(first, {
         repositories: fixture.ports.repositories,
@@ -179,7 +166,6 @@ describe('public agent candidate bundle builder', () => {
           },
           execution: fixture.bundle.execution,
           memory: { mode: 'disabled' },
-          lineage: { source: 'optimizer' },
         }),
       ).toThrow(/CodeSurface|patch/i)
 
@@ -193,10 +179,9 @@ describe('public agent candidate bundle builder', () => {
   it('fails closed when a generic profile would lose behavior or byte identity', () => {
     const fixture = createCandidateExecutionFixture(false)
     const base = {
-      code: { kind: 'disabled', reason: 'control' } as const,
+      code: { kind: 'disabled' } as const,
       execution: fixture.bundle.execution,
       memory: { mode: 'disabled' } as const,
-      lineage: { source: 'human' } as const,
     }
     expect(() =>
       buildAgentCandidateBundle({
@@ -246,16 +231,6 @@ describe('public agent candidate bundle builder', () => {
         code: { kind: 'git-patch' } as unknown as BuildAgentCandidateBundleInput['code'],
       }),
     ).toThrow(/unsupported candidate code source/)
-    expect(() =>
-      buildAgentCandidateBundle({
-        ...base,
-        profile: { kind: 'profile', profile: simpleProfile() },
-        lineage: {
-          source: 'human',
-          profileDiffIds: ['caller-controlled'],
-        } as unknown as BuildAgentCandidateBundleInput['lineage'],
-      }),
-    ).toThrow(/profileDiffIds.*derived/)
   })
 
   it('round-trips every currently representable generic profile surface', () => {
@@ -327,10 +302,9 @@ describe('public agent candidate bundle builder', () => {
 
     const bundle = buildAgentCandidateBundle({
       profile: { kind: 'profile', profile },
-      code: { kind: 'disabled', reason: 'control' },
+      code: { kind: 'disabled' },
       execution: fixture.bundle.execution,
       memory: { mode: 'disabled' },
-      lineage: { source: 'human' },
     })
 
     expect(bundle.profile).toMatchObject({
@@ -361,10 +335,9 @@ describe('public agent candidate bundle builder', () => {
           },
         },
       },
-      code: { kind: 'disabled', reason: 'control' },
+      code: { kind: 'disabled' },
       execution: fixture.bundle.execution,
       memory: { mode: 'disabled' },
-      lineage: { source: 'human' },
     })
     expect(bundle.profile.hooks?.Stop?.[0]).toEqual({
       executable: 'node',
diff --git a/tests/candidate-execution-claim.test.ts b/tests/candidate-execution-claim.test.ts
index 5ad0e9ff..699b6e31 100644
--- a/tests/candidate-execution-claim.test.ts
+++ b/tests/candidate-execution-claim.test.ts
@@ -55,15 +55,15 @@ describe('candidate execution claim lifecycle', () => {
 
     const requested = candidateExecutionClaim(prepared, preparationEvidenceFor(prepared))
     const ownerWindowMs = candidateExecutionOwnerWindowMs(
-      fixture.task.limits.timeoutMs,
+      fixture.task.task.limits.timeoutMs,
       cleanupTimeoutMs,
-      fixture.task.limits.timeoutMs,
-      fixture.task.limits.timeoutMs,
+      fixture.task.task.limits.timeoutMs,
+      fixture.task.task.limits.timeoutMs,
     )
 
     expect(requested.leaseExpiresAtMs).toBe(claimedAtMs + ownerWindowMs)
     expect(requested.cleanup.cleanupTimeoutMs).toBe(cleanupTimeoutMs)
-    expect(requested.resultTimeoutMs).toBe(fixture.task.limits.timeoutMs)
+    expect(requested.resultTimeoutMs).toBe(fixture.task.task.limits.timeoutMs)
     expect(ownerWindowMs).toBeLessThan(15 * 60_000)
   })
 
@@ -83,10 +83,10 @@ describe('candidate execution claim lifecycle', () => {
       { cleanupTimeoutMs },
     )
     const ownerWindowMs = candidateExecutionOwnerWindowMs(
-      fixture.task.limits.timeoutMs,
+      fixture.task.task.limits.timeoutMs,
       cleanupTimeoutMs,
-      fixture.task.limits.timeoutMs,
-      fixture.task.limits.timeoutMs,
+      fixture.task.task.limits.timeoutMs,
+      fixture.task.task.limits.timeoutMs,
     )
     vi.spyOn(Date, 'now').mockReturnValue(reservationExpiresAtMs - ownerWindowMs + 1)
 
diff --git a/tests/candidate-execution-execute.test.ts b/tests/candidate-execution-execute.test.ts
index 644e429b..f06e3c9d 100644
--- a/tests/candidate-execution-execute.test.ts
+++ b/tests/candidate-execution-execute.test.ts
@@ -21,6 +21,7 @@ import {
   createAgentCandidateWorkspacePort,
 } from '../src/candidate-execution/workspace-archive'
 import {
+  bindCandidateFixtureBundle,
   candidateBundle,
   candidateSha,
   cleanupCandidateFixtures,
@@ -29,6 +30,8 @@ import {
   createCandidateOutputFixture,
   emptyCandidateSnapshot,
   redigestCandidateBundle,
+  replaceCandidateFixtureAttempt,
+  replaceCandidateFixtureTask,
   unchangedTaskOutcomeCapture,
 } from './helpers/candidate-execution-fixture'
 
@@ -57,6 +60,7 @@ interface TestExecutor {
   execute: AgentCandidateExecutorPort['execute']
   stop: AgentCandidateExecutorPort['stop']
   capture?: AgentCandidateExecutorPort['capture']
+  dispose?: AgentCandidateExecutorPort['dispose']
 }
 
 function options(
@@ -73,6 +77,7 @@ function options(
         return await executor.execute(...args)
       },
       stop: executor.stop,
+      ...(executor.dispose !== undefined ? { dispose: executor.dispose } : {}),
       capture: async (
         stopRequest: Parameters<AgentCandidateExecutorPort['capture']>[0],
         context: Parameters<AgentCandidateExecutorPort['capture']>[1],
@@ -81,18 +86,18 @@ function options(
         if (capture.taskOutcome) return capture
         const request = requests.get(stopRequest.executionId)
         if (!request) throw new Error('fixture capture has no execution request')
-        const outcome = request.executionPlan.value.material.task.outcome
+        const outcome = request.benchmark.task.outcome
         if (outcome.kind !== 'workspace') {
           throw new Error('fixture fallback requires a workspace outcome')
         }
-        const repository = request.executionPlan.value.material.task.repository
+        const repository = request.benchmark.task.repository
         if (!repository) throw new Error('fixture fallback requires repository identity')
         return {
           ...capture,
           taskOutcome: {
             kind: 'workspace' as const,
             resultTree: repository.baseTree,
-            afterState: request.executionPlan.value.material.task.workspace.material,
+            afterState: request.benchmark.task.workspace.material,
             archive: Buffer.from('fixture unchanged task archive', 'utf8'),
             gitDiff: Buffer.alloc(0),
           },
@@ -108,9 +113,10 @@ function options(
 describe('atomic prepared candidate execution', () => {
   it('reveals credentials only to one trusted executor and returns a durable receipt', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.bundle = candidateBundle({
-      env: { PUBLIC_MODE: { kind: 'public', value: 'fixture' } },
-    })
+    bindCandidateFixtureBundle(
+      fixture,
+      candidateBundle({ env: { PUBLIC_MODE: { kind: 'public', value: 'fixture' } } }),
+    )
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -134,8 +140,8 @@ describe('atomic prepared candidate execution', () => {
           TANGLE_CANDIDATE_EXECUTION_ID: prepared.executionId,
         })
         expect(request.launch.env.PATH).toBeUndefined()
-        expect(request.hardLimits).toEqual({ timeoutMs: fixture.task.limits.timeoutMs })
-        expect(request.observedLimits).toEqual({ maxSteps: fixture.task.limits.maxSteps })
+        expect(request.hardLimits).toEqual({ timeoutMs: fixture.task.task.limits.timeoutMs })
+        expect(request.observedLimits).toEqual({ maxSteps: fixture.task.task.limits.maxSteps })
         expect(request.executionPlan.value.material.model.access.network).toEqual({
           mode: 'gateway-only',
           domains: ['router.tangle.tools'],
@@ -507,7 +513,7 @@ describe('atomic prepared candidate execution', () => {
     const claimStore = {
       tryClaim: async (claim: Parameters<typeof baseStore.tryClaim>[0]) => {
         const result = await baseStore.tryClaim(claim)
-        now += fixture.task.limits.timeoutMs + 1
+        now += fixture.task.task.limits.timeoutMs + 1
         return result
       },
       getAttempt: (attempt: Parameters<typeof baseStore.getAttempt>[0]) =>
@@ -552,7 +558,9 @@ describe('atomic prepared candidate execution', () => {
     let now = 1_000_000
     vi.spyOn(Date, 'now').mockImplementation(() => now)
     const fixture = createCandidateExecutionFixture()
-    fixture.task.limits = { ...fixture.task.limits, timeoutMs: 100 }
+    replaceCandidateFixtureTask(fixture, {
+      limits: { ...fixture.task.task.limits, timeoutMs: 100 },
+    })
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -566,7 +574,7 @@ describe('atomic prepared candidate execution', () => {
         baseStore.getAttempt(attempt),
       markCandidateMayRun: async (lease: Parameters<typeof baseStore.markCandidateMayRun>[0]) => {
         const result = await baseStore.markCandidateMayRun(lease)
-        now += fixture.task.limits.timeoutMs
+        now += fixture.task.task.limits.timeoutMs
         return result
       },
       stageTerminal: (
@@ -614,7 +622,9 @@ describe('atomic prepared candidate execution', () => {
     vi.spyOn(Date, 'now').mockImplementation(() => now)
     const cleanupTimeoutMs = 100
     const fixture = createCandidateExecutionFixture()
-    fixture.task.limits = { ...fixture.task.limits, timeoutMs: 1_000 }
+    replaceCandidateFixtureTask(fixture, {
+      limits: { ...fixture.task.task.limits, timeoutMs: 1_000 },
+    })
     fixture.ports.models.settleGrant = async ({ preparationId }) => {
       now += cleanupTimeoutMs - 1
       return { preparationId, grantDigest: candidateSha('c'), closed: true, calls: [] }
@@ -698,7 +708,9 @@ describe('atomic prepared candidate execution', () => {
 
   it('allows executable grading to exceed cleanup time inside its frozen result budget', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.task.limits = { ...fixture.task.limits, timeoutMs: 1_000 }
+    replaceCandidateFixtureTask(fixture, {
+      limits: { ...fixture.task.task.limits, timeoutMs: 1_000 },
+    })
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -746,7 +758,9 @@ describe('atomic prepared candidate execution', () => {
 
   it('cancels over-budget grading without any output appearing after terminal failure', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.task.limits = { ...fixture.task.limits, timeoutMs: 1_000 }
+    replaceCandidateFixtureTask(fixture, {
+      limits: { ...fixture.task.task.limits, timeoutMs: 1_000 },
+    })
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -1195,7 +1209,9 @@ describe('atomic prepared candidate execution', () => {
 
   it('owns the deadline, aborts, waits for process death, and then settles', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.task.limits = { ...fixture.task.limits, timeoutMs: 15 }
+    replaceCandidateFixtureTask(fixture, {
+      limits: { ...fixture.task.task.limits, timeoutMs: 15 },
+    })
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -1260,7 +1276,9 @@ describe('atomic prepared candidate execution', () => {
 
   it('cannot turn a stop acknowledgement after the frozen deadline into an exit success', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.task.limits = { ...fixture.task.limits, timeoutMs: 15 }
+    replaceCandidateFixtureTask(fixture, {
+      limits: { ...fixture.task.task.limits, timeoutMs: 15 },
+    })
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -1332,6 +1350,7 @@ describe('atomic prepared candidate execution', () => {
   it('withholds a receipt but still revokes model access when process death is unproven', async () => {
     const fixture = createCandidateExecutionFixture()
     let settlements = 0
+    let disposals = 0
     fixture.ports.models.settleGrant = async ({ preparationId }) => {
       settlements++
       return { preparationId, grantDigest: candidateSha('c'), closed: true, calls: [] }
@@ -1353,6 +1372,10 @@ describe('atomic prepared candidate execution', () => {
           stop: async () => {
             throw new Error('container death not observed')
           },
+          dispose: async () => {
+            disposals++
+            return { disposed: true }
+          },
         },
         traceStore,
       ),
@@ -1362,17 +1385,17 @@ describe('atomic prepared candidate execution', () => {
       reason: expect.stringMatching(/death not observed/),
       usage: { modelCalls: 0 },
     })
-    expect(settlements).toBe(1)
+    expect({ settlements, disposals }).toEqual({ settlements: 1, disposals: 1 })
   })
 
   it('leaves unknown settlement unfinished so a retry cannot guess zero spend', async () => {
     const claimStore = new InMemoryAgentCandidateExecutionClaimStore()
     const first = createCandidateExecutionFixture()
-    first.task.attempt = {
+    replaceCandidateFixtureAttempt(first, {
       number: 1,
       maxAttempts: 2,
       retryPolicy: 'pre-model-infrastructure-only',
-    }
+    })
     first.ports.models.settleGrant = async () => {
       throw new Error('gateway settlement unavailable')
     }
@@ -1398,11 +1421,11 @@ describe('atomic prepared candidate execution', () => {
 
     rmSync(first.task.stagingRoots.profileRoot, { recursive: true })
     mkdirSync(first.task.stagingRoots.profileRoot)
-    first.task.attempt = {
+    replaceCandidateFixtureAttempt(first, {
       number: 2,
       maxAttempts: 2,
       retryPolicy: 'pre-model-infrastructure-only',
-    }
+    })
     const retryPrepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(first.bundle, first.ports),
       first.task,
@@ -1432,11 +1455,11 @@ describe('atomic prepared candidate execution', () => {
 
   it('admits a zero-call retry only for an explicit pre-model infrastructure failure', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.task.attempt = {
+    replaceCandidateFixtureAttempt(fixture, {
       number: 1,
       maxAttempts: 2,
       retryPolicy: 'pre-model-infrastructure-only',
-    }
+    })
     const claimStore = new InMemoryAgentCandidateExecutionClaimStore()
     let activations = 0
     fixture.ports.models.activateGrant = async () => {
@@ -1467,11 +1490,11 @@ describe('atomic prepared candidate execution', () => {
 
     rmSync(fixture.task.stagingRoots.profileRoot, { recursive: true })
     mkdirSync(fixture.task.stagingRoots.profileRoot)
-    fixture.task.attempt = {
+    replaceCandidateFixtureAttempt(fixture, {
       number: 2,
       maxAttempts: 2,
       retryPolicy: 'pre-model-infrastructure-only',
-    }
+    })
     const retryPrepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -1497,9 +1520,12 @@ describe('atomic prepared candidate execution', () => {
 
   it('closes isolated memory and persists a large after-state archive by reference', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.bundle = redigestCandidateBundle(fixture.bundle, {
-      memory: { mode: 'isolated', scope: 'task' },
-    })
+    bindCandidateFixtureBundle(
+      fixture,
+      redigestCandidateBundle(fixture.bundle, {
+        memory: { mode: 'isolated', scope: 'task' },
+      }),
+    )
     const before = emptyCandidateSnapshot('before')
     const after = emptyCandidateSnapshot('after')
     const order: string[] = []
@@ -1574,9 +1600,12 @@ describe('atomic prepared candidate execution', () => {
 
   it('rejects a compressed protected value in isolated memory before persisting memory evidence', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.bundle = redigestCandidateBundle(fixture.bundle, {
-      memory: { mode: 'isolated', scope: 'task' },
-    })
+    bindCandidateFixtureBundle(
+      fixture,
+      redigestCandidateBundle(fixture.bundle, {
+        memory: { mode: 'isolated', scope: 'task' },
+      }),
+    )
     const before = emptyCandidateSnapshot('secret-memory-before')
     fixture.ports.memory.reset = async ({ preparationId, expiresAtMs }) => ({
       preparationId,
@@ -1721,9 +1750,10 @@ describe('atomic prepared candidate execution', () => {
 
   it('rejects protected environment collisions before executor access', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.bundle = candidateBundle({
-      env: { PUBLIC_MODE: { kind: 'public', value: 'fixture' } },
-    })
+    bindCandidateFixtureBundle(
+      fixture,
+      candidateBundle({ env: { PUBLIC_MODE: { kind: 'public', value: 'fixture' } } }),
+    )
     fixture.ports.models.activateGrant = async () => ({ env: { PUBLIC_MODE: 'secret-value' } })
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
diff --git a/tests/candidate-execution-finalize.test.ts b/tests/candidate-execution-finalize.test.ts
index 73a5ac06..afbf560a 100644
--- a/tests/candidate-execution-finalize.test.ts
+++ b/tests/candidate-execution-finalize.test.ts
@@ -23,6 +23,7 @@ import {
   cleanupCandidateFixtures,
   createCandidateExecutionFixture,
   createCandidateOutputFixture,
+  replaceCandidateFixtureTask,
 } from './helpers/candidate-execution-fixture'
 
 afterEach(cleanupCandidateFixtures)
@@ -45,7 +46,7 @@ async function prepared(
   },
 ): Promise<PreparedAgentCandidateExecution> {
   const fixture = createCandidateExecutionFixture()
-  fixture.task.limits = limits
+  replaceCandidateFixtureTask(fixture, { limits })
   return await prepareAgentCandidateExecution(
     await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
     fixture.task,
@@ -94,18 +95,18 @@ async function finalizePrepared(
       },
     )
   const outputs = overrides.outputs ?? createCandidateOutputFixture()
-  const expectedOutcome = state.executionPlan.value.material.task.outcome
+  const expectedOutcome = state.benchmarkTask.outcome
   if (expectedOutcome.kind !== 'workspace') {
     throw new Error('finalization fixture requires a workspace outcome')
   }
-  const repository = state.executionPlan.value.material.task.repository
+  const repository = state.benchmarkTask.repository
   if (!repository) throw new Error('finalization fixture requires repository identity')
   const finalCapture = sealAgentCandidateExecutorFinalCapture(
     {
       taskOutcome: {
         kind: 'workspace' as const,
         resultTree: repository.baseTree,
-        afterState: state.executionPlan.value.material.task.workspace.material,
+        afterState: state.benchmarkTask.workspace.material,
         archive: Buffer.from('fixture unchanged task archive', 'utf8'),
         gitDiff: Buffer.alloc(0),
       },
@@ -214,7 +215,7 @@ describe('protected candidate run finalization', () => {
     if (!result.succeeded) return
     expect(result.receipt.value.trace).toMatchObject({ eventCount: 3, modelCallCount: 1 })
     expect(result.receipt.bytes.byteLength).toBeGreaterThan(0)
-    const task = assertPreparedCandidateIntegrity(execution).executionPlan.value.material.task
+    const task = assertPreparedCandidateIntegrity(execution).benchmarkTask
     if (task.outcome.kind !== 'workspace') throw new Error('expected workspace task outcome')
     if (!task.repository) throw new Error('expected task repository identity')
     expect(result.receipt.value).toMatchObject({
diff --git a/tests/candidate-execution-outcome-evidence.test.ts b/tests/candidate-execution-outcome-evidence.test.ts
index 683aa1a2..0abd84de 100644
--- a/tests/candidate-execution-outcome-evidence.test.ts
+++ b/tests/candidate-execution-outcome-evidence.test.ts
@@ -78,7 +78,7 @@ describe('candidate outcome evidence', () => {
         {
           kind: 'workspace',
           resultTree: '0'.repeat(40),
-          afterState: fixture.task.workspace.material,
+          afterState: fixture.task.task.workspace.material,
           archive: Buffer.from('archive'),
           gitDiff: new Uint8Array(),
         },
@@ -109,6 +109,56 @@ describe('candidate outcome evidence', () => {
     expect(puts).toBe(0)
   })
 
+  it('screens native executor evidence and stores it separately from the capture summary', async () => {
+    const fixture = createCandidateOutputExecutionFixture('text/plain', 32)
+    const state = await preparedState(fixture)
+    const outputs = createCandidateOutputFixture()
+    const purposes: string[] = []
+    const outputArtifacts = {
+      read: outputs.outputArtifacts.read,
+      put: async (input: Parameters<typeof outputs.outputArtifacts.put>[0]) => {
+        purposes.push(input.purpose)
+        return await outputs.outputArtifacts.put(input)
+      },
+    }
+    const capture = sealAgentCandidateExecutorFinalCapture(
+      {
+        taskOutcome: { kind: 'output', bytes: Buffer.from('answer', 'utf8') },
+        evidence: Buffer.from('native evidence', 'utf8'),
+      },
+      state.benchmarkTask.outcome,
+    )
+
+    await persistVerifiedAgentCandidateExecutorCapture(state, capture, outputArtifacts, [])
+    expect(purposes).toContain('executor-native-evidence')
+    expect(purposes).toContain('executor-capture')
+
+    let protectedPuts = 0
+    const secret = 'sk-native-evidence-secret-123456789'
+    const protectedCapture = sealAgentCandidateExecutorFinalCapture(
+      {
+        taskOutcome: { kind: 'output', bytes: Buffer.from('answer', 'utf8') },
+        evidence: Buffer.from(secret, 'utf8'),
+      },
+      state.benchmarkTask.outcome,
+    )
+    await expect(
+      persistVerifiedAgentCandidateExecutorCapture(
+        state,
+        protectedCapture,
+        {
+          read: outputs.outputArtifacts.read,
+          put: async (input) => {
+            protectedPuts++
+            return await outputs.outputArtifacts.put(input)
+          },
+        },
+        [secret],
+      ),
+    ).rejects.toThrow(/protected value/)
+    expect(protectedPuts).toBe(0)
+  })
+
   it('validates task manifest material before any output write', async () => {
     const fixture = createCandidateExecutionFixture()
     const state = await preparedState(fixture)
@@ -422,7 +472,7 @@ async function persistTaskOutcome(
 ) {
   const capture = sealAgentCandidateExecutorFinalCapture(
     { taskOutcome },
-    state.executionPlan.value.material.task.outcome,
+    state.benchmarkTask.outcome,
   )
   return (
     await persistVerifiedAgentCandidateExecutorCapture(
diff --git a/tests/candidate-execution-prepare.test.ts b/tests/candidate-execution-prepare.test.ts
index cf19c9d7..7c1a1acd 100644
--- a/tests/candidate-execution-prepare.test.ts
+++ b/tests/candidate-execution-prepare.test.ts
@@ -1,19 +1,11 @@
-import { execFileSync } from 'node:child_process'
-import { chmodSync, mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs'
-import { tmpdir } from 'node:os'
+import { mkdirSync, writeFileSync } from 'node:fs'
 import { join } from 'node:path'
 
-import type {
-  AgentCandidateBundle,
-  AgentCandidateExecution,
-  AgentCandidateWorkspaceSnapshotEvidence,
-  Sha256Digest,
-} from '@tangle-network/agent-interface'
+import type { AgentCandidateBundle } from '@tangle-network/agent-interface'
 import { afterEach, describe, expect, it } from 'vitest'
 import { MAX_CANDIDATE_TIMER_INTERVAL_MS } from '../src/candidate-execution/cleanup'
 import {
-  canonicalCandidateBytes,
-  canonicalCandidateDigest,
+  canonicalCandidateDocument,
   embeddedCandidateArtifact,
 } from '../src/candidate-execution/digest'
 import {
@@ -21,293 +13,24 @@ import {
   CANDIDATE_KNOWLEDGE_ROOT_ENV,
 } from '../src/candidate-execution/knowledge'
 import { prepareAgentCandidateExecution } from '../src/candidate-execution/prepare'
-import type {
-  AgentCandidateExecutionPorts,
-  AgentCandidateTaskExecution,
-  ResolvedAgentCandidateContainer,
-} from '../src/candidate-execution/types'
+import { parseAgentCandidateProfileActivation } from '../src/candidate-execution/profile'
+import type { AgentCandidateExecutionPorts } from '../src/candidate-execution/types'
 import { verifyAgentCandidateBundle } from '../src/candidate-execution/verify'
-
-const roots: string[] = []
-const sha = (character: string): Sha256Digest => `sha256:${character.repeat(64)}`
+import {
+  bindCandidateFixtureBundle,
+  candidateBundle as bundle,
+  cleanupCandidateFixtures,
+  emptyCandidateSnapshot as emptySnapshot,
+  createCandidateExecutionFixture as fixture,
+  redigestCandidateBundle as redigestBundle,
+  replaceCandidateFixtureTask,
+  candidateSha as sha,
+} from './helpers/candidate-execution-fixture'
 
 afterEach(() => {
-  for (const root of roots.splice(0)) rmSync(root, { recursive: true, force: true })
+  cleanupCandidateFixtures()
 })
 
-function temporaryRoot(prefix: string): string {
-  const root = mkdtempSync(join(tmpdir(), prefix))
-  roots.push(root)
-  return root
-}
-
-function git(root: string, args: string[]): string {
-  return execFileSync('git', args, { cwd: root, encoding: 'utf8' }).trim()
-}
-
-function taskRepository(): { root: string; commit: string; tree: string } {
-  const root = temporaryRoot('candidate-task-')
-  git(root, ['init', '-b', 'main'])
-  git(root, ['config', 'user.email', 'test@example.com'])
-  git(root, ['config', 'user.name', 'Test'])
-  git(root, ['config', 'core.hooksPath', '/dev/null'])
-  git(root, ['remote', 'add', 'origin', 'git@github.com:owner/repo.git'])
-  writeFileSync(join(root, 'source.ts'), 'export const value = 1\n')
-  chmodSync(join(root, 'source.ts'), 0o644)
-  git(root, ['add', 'source.ts'])
-  git(root, ['commit', '-m', 'base'])
-  return {
-    root,
-    commit: git(root, ['rev-parse', 'HEAD']),
-    tree: git(root, ['rev-parse', 'HEAD^{tree}']),
-  }
-}
-
-function snapshot(
-  root: string,
-  files: Array<{ path: string; mode: number }>,
-): AgentCandidateWorkspaceSnapshotEvidence {
-  const material = {
-    kind: 'agent-candidate-workspace-manifest' as const,
-    files: files
-      .map((file) => {
-        const bytes = execFileSync('node', [
-          '-e',
-          `process.stdout.write(require('fs').readFileSync(${JSON.stringify(join(root, file.path))}))`,
-        ])
-        return {
-          path: file.path,
-          mode: file.mode,
-          sha256: embeddedCandidateArtifact(bytes).sha256,
-          byteLength: bytes.byteLength,
-        }
-      })
-      .sort((a, b) => a.path.localeCompare(b.path)),
-  }
-  const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
-  return {
-    kind: 'agent-candidate-workspace-snapshot',
-    digest: manifest.sha256,
-    material,
-    manifest,
-    archive: embeddedCandidateArtifact(Buffer.from(`archive:${manifest.sha256}`)),
-  }
-}
-
-function bundle(
-  execution: Partial<AgentCandidateExecution> = {},
-  active?: { commit: string; tree: string; workspace: AgentCandidateWorkspaceSnapshotEvidence },
-): AgentCandidateBundle {
-  const value = {
-    kind: 'agent-candidate-bundle' as const,
-    digestAlgorithm: 'rfc8785-sha256' as const,
-    profile: {
-      name: 'candidate',
-      prompt: { instructions: ['Inspect the repository, implement the fix, and run tests.'] },
-      model: { default: 'provider/model', reasoningEffort: 'high' as const },
-      harness: 'codex' as const,
-      resources: { failOnError: true as const },
-    },
-    code: active
-      ? {
-          kind: 'no-op' as const,
-          reason: 'proposer-no-change' as const,
-          repository: { kind: 'github' as const, owner: 'owner', repo: 'repo' },
-          baseCommit: active.commit,
-          baseTree: active.tree,
-        }
-      : { kind: 'disabled' as const, reason: 'control' as const },
-    execution: {
-      harness: 'codex' as const,
-      harnessVersion: '1.2.3',
-      launch: active
-        ? {
-            kind: 'candidate-entrypoint' as const,
-            entrypoint: 'run.js',
-            interpreter: 'node' as const,
-          }
-        : { kind: 'container-command' as const, executable: 'codex' },
-      instructionDelivery: { kind: 'stdin-utf8' as const },
-      cwd: { workspace: 'task' as const, path: '.' },
-      environment: { kind: 'evaluator-task-container' as const },
-      ...(active ? { workspace: active.workspace } : {}),
-      isolation: {
-        network: 'disabled' as const,
-        remoteIntegrations: 'disabled' as const,
-        candidateSecrets: 'disabled' as const,
-      },
-      ...execution,
-    },
-    memory: { mode: 'disabled' as const },
-    lineage: active
-      ? {
-          source: 'optimizer' as const,
-          parentDigests: [sha('e')],
-          runIds: ['optimizer-run-1'],
-          benchmark: { name: 'development', version: '1', splitDigest: sha('f') },
-          spend: {
-            proposal: { costUsd: 0, inputTokens: 0, outputTokens: 0, modelCalls: 0 },
-            evaluation: { costUsd: 0, inputTokens: 0, outputTokens: 0, modelCalls: 0 },
-          },
-        }
-      : { source: 'human' as const },
-  }
-  return { ...value, digest: canonicalCandidateDigest(value) }
-}
-
-function redigestBundle(
-  source: AgentCandidateBundle,
-  overrides: Partial<Omit<AgentCandidateBundle, 'digest'>>,
-): AgentCandidateBundle {
-  const { digest: _digest, ...withoutDigest } = source
-  const value = { ...withoutDigest, ...overrides }
-  return { ...value, digest: canonicalCandidateDigest(value) }
-}
-
-function emptySnapshot(label: string): AgentCandidateWorkspaceSnapshotEvidence {
-  const material = {
-    kind: 'agent-candidate-workspace-manifest' as const,
-    files: [],
-  }
-  const manifest = embeddedCandidateArtifact(canonicalCandidateBytes(material))
-  return {
-    kind: 'agent-candidate-workspace-snapshot',
-    digest: manifest.sha256,
-    material,
-    manifest,
-    archive: embeddedCandidateArtifact(Buffer.from(`empty:${label}`)),
-  }
-}
-
-function fixture(active = false): {
-  bundle: AgentCandidateBundle
-  task: AgentCandidateTaskExecution
-  ports: AgentCandidateExecutionPorts
-  candidateRoot?: string
-} {
-  const repository = taskRepository()
-  const taskWorkspace = snapshot(repository.root, [{ path: 'source.ts', mode: 0o644 }])
-  let candidateRoot: string | undefined
-  let candidateWorkspace: AgentCandidateWorkspaceSnapshotEvidence | undefined
-  if (active) {
-    candidateRoot = temporaryRoot('candidate-built-')
-    writeFileSync(join(candidateRoot, 'run.js'), '#!/usr/bin/env node\n')
-    chmodSync(join(candidateRoot, 'run.js'), 0o755)
-    candidateWorkspace = snapshot(candidateRoot, [{ path: 'run.js', mode: 0o755 }])
-  }
-  const profileRoot = temporaryRoot('candidate-profile-')
-  const selectedContainer: ResolvedAgentCandidateContainer = {
-    source: 'evaluator-task-container',
-    image: 'ghcr.io/example/task',
-    indexDigest: sha('a'),
-    manifestDigest: sha('b'),
-    platform: { os: 'linux', architecture: 'amd64' },
-  }
-  const ports: AgentCandidateExecutionPorts = {
-    artifacts: {
-      read: async () => {
-        throw new Error('all fixture artifacts are embedded')
-      },
-    },
-    repositories: { resolve: async () => repository.root },
-    workspaces: { materialize: async () => undefined },
-    containers: { resolve: async () => selectedContainer },
-    models: {
-      resolve: async ({ requested, reasoningEffort }) => ({
-        requested,
-        provider: 'provider',
-        model: 'model-snapshot',
-        snapshot: 'model-snapshot-2026-07-01',
-        reasoningEffort,
-      }),
-      reserveGrant: async ({ preparationId, expiresAtMs, limits }) => ({
-        preparationId,
-        digest: sha('c'),
-        expiresAtMs,
-        enforcedLimits: limits,
-        network:
-          limits.maxModelCalls === 0
-            ? { mode: 'disabled' as const }
-            : { mode: 'gateway-only' as const, domains: ['router.tangle.tools'] },
-      }),
-      activateGrant: async () => ({ env: { MODEL_GATEWAY_TOKEN: 'protected' } }),
-      settleGrant: async ({ preparationId }) => ({
-        preparationId,
-        grantDigest: sha('c'),
-        closed: true,
-        calls: [],
-      }),
-    },
-    memory: {
-      reset: async () => {
-        throw new Error('disabled memory must not reset')
-      },
-      activate: async () => {
-        throw new Error('disabled memory must not activate')
-      },
-      close: async () => {
-        throw new Error('disabled memory must not close')
-      },
-    },
-  }
-  const task: AgentCandidateTaskExecution = {
-    executionId: 'execution-1',
-    benchmark: 'repository-disjoint-smoke',
-    benchmarkVersion: '1',
-    taskId: 'owner-repo-1',
-    splitDigest: sha('d'),
-    instruction: 'Fix the failing behavior without changing the public API.',
-    repository: {
-      identity: 'github.com/owner/repo',
-      rootIdentity: 'owner/repo',
-      baseCommit: repository.commit,
-      baseTree: repository.tree,
-    },
-    outcome: { kind: 'workspace' },
-    attempt: { number: 1, maxAttempts: 1, retryPolicy: 'none' },
-    model: { requested: 'provider/model', reasoningEffort: 'high' },
-    grader: {
-      name: 'fixture-executable-grader',
-      version: '1.0.0',
-      artifact: {
-        locator: { kind: 's3', bucket: 'candidate-test-artifacts', key: 'grader/fixture' },
-        sha256: sha('a'),
-        byteLength: 1,
-      },
-    },
-    executionRoots: {
-      taskRoot: '/workspace/task',
-      ...(active ? { candidateRoot: '/opt/candidate' } : {}),
-    },
-    stagingRoots: {
-      taskRoot: repository.root,
-      ...(candidateRoot ? { candidateRoot } : {}),
-      profileRoot,
-    },
-    workspace: taskWorkspace,
-    evaluatorTaskContainer: selectedContainer,
-    limits: {
-      timeoutMs: 60_000,
-      maxSteps: 100,
-      maxModelCalls: 50,
-      maxInputTokens: 100_000,
-      maxOutputTokens: 50_000,
-      maxCostUsd: 5,
-    },
-  }
-  return {
-    bundle: bundle(
-      {},
-      active && candidateWorkspace
-        ? { commit: repository.commit, tree: repository.tree, workspace: candidateWorkspace }
-        : undefined,
-    ),
-    task,
-    ports,
-    ...(candidateRoot ? { candidateRoot } : {}),
-  }
-}
-
 describe('candidate execution preparation', () => {
   it('binds exact instruction, repository, profile, model, image, roots, and limits', async () => {
     const value = fixture()
@@ -318,30 +41,28 @@ describe('candidate execution preparation', () => {
         path: '/tangle/input/task.txt',
       },
     })
+    bindCandidateFixtureBundle(value)
     const verified = await verifyAgentCandidateBundle(value.bundle, value.ports)
     const prepared = await prepareAgentCandidateExecution(verified, value.task, value.ports)
     const plan = prepared.executionPlan.value.material
-    expect(plan.task.instruction).toEqual({
-      encoding: 'utf8',
-      sha256: embeddedCandidateArtifact(Buffer.from(value.task.instruction)).sha256,
-      byteLength: Buffer.byteLength(value.task.instruction),
-      delivery: value.bundle.execution.instructionDelivery,
-    })
-    expect(plan.task.outcome).toEqual(value.task.outcome)
-    expect(plan.task.repository).toEqual(value.task.repository)
-    expect(plan.task.outcome).toEqual({ kind: 'workspace' })
+    expect(plan.instructionDelivery).toEqual(value.bundle.execution.instructionDelivery)
+    expect(prepared.benchmark.task.outcome).toEqual(value.task.task.outcome)
+    expect(prepared.benchmark.task.repository).toEqual(value.task.task.repository)
+    expect(prepared.benchmark.task.outcome).toEqual({ kind: 'workspace' })
     expect(plan.profile).toEqual({
       planDigest: prepared.profilePlan.value.digest,
       targetWorkspace: 'task',
       mountPaths: ['AGENTS.md'],
     })
     expect(plan.launch.env.TANGLE_CANDIDATE_TASK_PATH?.value).toBe('/tangle/input/task.txt')
-    expect(Buffer.from(prepared.instruction.bytes)).toEqual(Buffer.from(value.task.instruction))
+    expect(Buffer.from(prepared.instruction.bytes)).toEqual(
+      Buffer.from(value.task.task.instruction),
+    )
     expect(Buffer.from(prepared.executionPlan.bytes)).toEqual(
       Buffer.from(prepared.executionPlan.value.artifact.content, 'base64'),
     )
     expect(prepared.materializationReceipt.bytes.byteLength).toBeGreaterThan(0)
-    expect(JSON.stringify(plan)).not.toContain(value.task.instruction)
+    expect(JSON.stringify(plan)).not.toContain(value.task.task.instruction)
     expect(JSON.stringify(prepared)).not.toContain('MODEL_GATEWAY_TOKEN')
     expect(JSON.stringify(prepared)).not.toContain('protected')
     expect(plan.model.access.network).toEqual({
@@ -350,15 +71,36 @@ describe('candidate execution preparation', () => {
     })
   })
 
+  it('rejects a self-rehashed profile activation whose native file differs from its plan', async () => {
+    const value = fixture()
+    const prepared = await prepareAgentCandidateExecution(
+      await verifyAgentCandidateBundle(value.bundle, value.ports),
+      value.task,
+      value.ports,
+    )
+    const { digest: _digest, ...activation } = prepared.profileActivation
+    const [first, ...rest] = activation.files
+    if (!first) throw new Error('fixture profile activation has no native files')
+    const forged = canonicalCandidateDocument({
+      ...activation,
+      files: [{ ...first, content: `${first.content}\nforged` }, ...rest],
+    }).value
+
+    expect(() =>
+      parseAgentCandidateProfileActivation(forged, prepared.profilePlan.value.digest),
+    ).toThrow(/must match the canonical plan/)
+  })
+
   it('keeps argv task bytes out of fixed args and exposes deterministic delivery separately', async () => {
     const value = fixture()
     value.bundle = bundle({ instructionDelivery: { kind: 'argv-append' } })
+    bindCandidateFixtureBundle(value)
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(value.bundle, value.ports),
       value.task,
       value.ports,
     )
-    expect(prepared.launch.args).not.toContain(value.task.instruction)
+    expect(prepared.launch.args).not.toContain(value.task.task.instruction)
     expect(prepared.instruction).toMatchObject({ delivery: { kind: 'argv-append' } })
   })
 
@@ -396,6 +138,7 @@ describe('candidate execution preparation', () => {
         [field]: profileValue,
       } as AgentCandidateBundle['profile'],
     })
+    bindCandidateFixtureBundle(value)
     let stagingCalls = 0
     let reservationCalls = 0
     value.ports.workspaces.materialize = async () => {
@@ -421,9 +164,9 @@ describe('candidate execution preparation', () => {
         tools: {},
         permissions: {},
         modes: {},
-        confidential: {},
       },
     })
+    bindCandidateFixtureBundle(value)
 
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(value.bundle, value.ports),
@@ -435,8 +178,10 @@ describe('candidate execution preparation', () => {
 
   it('rejects task Git drift, dirty profile staging, and unenforced model limits', async () => {
     const gitDrift = fixture()
-    if (!gitDrift.task.repository) throw new Error('expected repository identity')
-    gitDrift.task.repository.baseTree = '0'.repeat(40)
+    if (!gitDrift.task.task.repository) throw new Error('expected repository identity')
+    replaceCandidateFixtureTask(gitDrift, {
+      repository: { ...gitDrift.task.task.repository, baseTree: '0'.repeat(40) },
+    })
     await expect(
       prepareAgentCandidateExecution(
         await verifyAgentCandidateBundle(gitDrift.bundle, gitDrift.ports),
@@ -446,13 +191,11 @@ describe('candidate execution preparation', () => {
     ).rejects.toThrow(/base tree/)
 
     const outputGitDrift = fixture()
-    if (!outputGitDrift.task.repository) throw new Error('expected repository identity')
-    outputGitDrift.task.outcome = {
-      kind: 'output',
-      mediaType: 'application/json',
-      maxBytes: 1_024,
-    }
-    outputGitDrift.task.repository.baseTree = '0'.repeat(40)
+    if (!outputGitDrift.task.task.repository) throw new Error('expected repository identity')
+    replaceCandidateFixtureTask(outputGitDrift, {
+      outcome: { kind: 'output', mediaType: 'application/json', maxBytes: 1_024 },
+      repository: { ...outputGitDrift.task.task.repository, baseTree: '0'.repeat(40) },
+    })
     await expect(
       prepareAgentCandidateExecution(
         await verifyAgentCandidateBundle(outputGitDrift.bundle, outputGitDrift.ports),
@@ -518,13 +261,15 @@ describe('candidate execution preparation', () => {
     expect(settledReservations).toBe(1)
 
     const zeroCall = fixture()
-    zeroCall.task.limits = {
-      ...zeroCall.task.limits,
-      maxModelCalls: 0,
-      maxInputTokens: 0,
-      maxOutputTokens: 0,
-      maxCostUsd: 0,
-    }
+    replaceCandidateFixtureTask(zeroCall, {
+      limits: {
+        ...zeroCall.task.task.limits,
+        maxModelCalls: 0,
+        maxInputTokens: 0,
+        maxOutputTokens: 0,
+        maxCostUsd: 0,
+      },
+    })
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(zeroCall.bundle, zeroCall.ports),
       zeroCall.task,
@@ -581,7 +326,9 @@ describe('candidate execution preparation', () => {
 
   it('rejects a cost limit that cannot be represented by the protected integer ledger', async () => {
     const value = fixture()
-    value.task.limits = { ...value.task.limits, maxCostUsd: 10 ** 20 }
+    replaceCandidateFixtureTask(value, {
+      limits: { ...value.task.task.limits, maxCostUsd: 10 ** 20 },
+    })
     let reservations = 0
     value.ports.models.reserveGrant = async () => {
       reservations++
@@ -599,10 +346,12 @@ describe('candidate execution preparation', () => {
 
   it('rejects a wall-time limit that Node would clamp to an immediate timer', async () => {
     const value = fixture()
-    value.task.limits = {
-      ...value.task.limits,
-      timeoutMs: MAX_CANDIDATE_TIMER_INTERVAL_MS + 1,
-    }
+    replaceCandidateFixtureTask(value, {
+      limits: {
+        ...value.task.task.limits,
+        timeoutMs: MAX_CANDIDATE_TIMER_INTERVAL_MS + 1,
+      },
+    })
     let reservations = 0
     value.ports.models.reserveGrant = async () => {
       reservations++
@@ -663,6 +412,7 @@ describe('candidate execution preparation', () => {
         evaluation,
       },
     })
+    bindCandidateFixtureBundle(value)
     value.ports.artifacts.read = async (ref) => {
       if (ref.sha256 === seed.sha256) return seedBytes
       throw new Error(`unexpected artifact ${ref.sha256}`)
@@ -679,6 +429,7 @@ describe('candidate execution preparation', () => {
         beforeState: emptySnapshot('before'),
       }
     }
+    value.ports.memory.close = async () => ({ closed: true })
     const prepared = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(value.bundle, value.ports),
       value.task,
@@ -726,6 +477,7 @@ describe('candidate execution preparation', () => {
     value.bundle = redigestBundle(value.bundle, {
       knowledge: { snapshotId: 'knowledge-1', manifest },
     })
+    bindCandidateFixtureBundle(value)
     let reads = 0
     value.ports.artifacts.read = async () => {
       reads++
@@ -743,6 +495,7 @@ describe('candidate execution preparation', () => {
     value.bundle = redigestBundle(value.bundle, {
       memory: { mode: 'isolated', scope: 'task' },
     })
+    bindCandidateFixtureBundle(value)
     const before = emptySnapshot('malformed-reset')
     const closed: string[] = []
     value.ports.memory.reset = async ({ preparationId, expiresAtMs }) => ({
diff --git a/tests/candidate-execution-recover.test.ts b/tests/candidate-execution-recover.test.ts
index 6eb4d56e..3434b881 100644
--- a/tests/candidate-execution-recover.test.ts
+++ b/tests/candidate-execution-recover.test.ts
@@ -17,11 +17,13 @@ import type {
 } from '../src/candidate-execution/types'
 import { verifyAgentCandidateBundle } from '../src/candidate-execution/verify'
 import {
+  bindCandidateFixtureBundle,
   candidateSha,
   cleanupCandidateFixtures,
   createCandidateExecutionFixture,
   emptyCandidateSnapshot,
   redigestCandidateBundle,
+  replaceCandidateFixtureAttempt,
 } from './helpers/candidate-execution-fixture'
 
 afterEach(cleanupCandidateFixtures)
@@ -42,6 +44,7 @@ describe('expired candidate recovery', () => {
     expect(acquired.acquired).toBe(true)
     now = claim.leaseExpiresAtMs
     let stops = 0
+    let disposals = 0
     let settlements = 0
     const originalSettle = fixture.ports.models.settleGrant
     fixture.ports.models.settleGrant = async (input) => {
@@ -69,6 +72,15 @@ describe('expired candidate recovery', () => {
           expect(context.deadlineAtMs).toBeGreaterThan(Date.now())
           return { stopped: true }
         },
+        dispose: async (request, context) => {
+          disposals++
+          expect(request).toEqual({
+            executionId: claim.executionId,
+            executionPlanDigest: claim.executionPlanDigest,
+          })
+          expect(context.signal.aborted).toBe(false)
+          return { disposed: true }
+        },
         capture: async () => ({ evidence: Buffer.from('official recovery evidence') }),
       },
     })
@@ -113,7 +125,7 @@ describe('expired candidate recovery', () => {
       },
     })
     expect(replay).toMatchObject({ finished: false, exactReplay: true })
-    expect({ stops, settlements }).toEqual({ stops: 1, settlements: 1 })
+    expect({ stops, disposals, settlements }).toEqual({ stops: 1, disposals: 1, settlements: 1 })
   })
 
   it('does no outward cleanup before expiry and records nothing when cleanup is unproven', async () => {
@@ -162,11 +174,11 @@ describe('expired candidate recovery', () => {
 
   it('unlocks attempt two only when a recovered pre-run crash used zero model calls', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.task.attempt = {
+    replaceCandidateFixtureAttempt(fixture, {
       number: 1,
       maxAttempts: 2,
       retryPolicy: 'pre-model-infrastructure-only',
-    }
+    })
     const first = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -200,11 +212,11 @@ describe('expired candidate recovery', () => {
 
     rmSync(fixture.task.stagingRoots.profileRoot, { recursive: true, force: true })
     mkdirSync(fixture.task.stagingRoots.profileRoot)
-    fixture.task.attempt = {
+    replaceCandidateFixtureAttempt(fixture, {
       number: 2,
       maxAttempts: 2,
       retryPolicy: 'pre-model-infrastructure-only',
-    }
+    })
     const second = await prepareAgentCandidateExecution(
       await verifyAgentCandidateBundle(fixture.bundle, fixture.ports),
       fixture.task,
@@ -217,9 +229,12 @@ describe('expired candidate recovery', () => {
 
   it('repeats idempotent cleanup after a partial recovery failure', async () => {
     const fixture = createCandidateExecutionFixture()
-    fixture.bundle = redigestCandidateBundle(fixture.bundle, {
-      memory: { mode: 'isolated', scope: 'task' },
-    })
+    bindCandidateFixtureBundle(
+      fixture,
+      redigestCandidateBundle(fixture.bundle, {
+        memory: { mode: 'isolated', scope: 'task' },
+      }),
+    )
     const before = emptyCandidateSnapshot('recovery-before')
     fixture.ports.memory.reset = async ({ preparationId, expiresAtMs }) => ({
       preparationId,
@@ -286,14 +301,14 @@ describe('expired candidate recovery', () => {
       })
 
     await expect(recover()).rejects.toThrow(/cleanup could not be proven/)
-    expect(persistedPurposes).toContain('executor-capture')
+    expect(persistedPurposes).not.toContain('executor-capture')
     expect(
       await claimStore.getAttempt({ executionId: claim.executionId, attempt: 1 }),
     ).not.toHaveProperty('terminal')
     await expect(recover()).resolves.toMatchObject({ finished: true })
     expect({ stops, captures, settlements, memoryCloses }).toEqual({
       stops: 2,
-      captures: 2,
+      captures: 0,
       settlements: 2,
       memoryCloses: 2,
     })
diff --git a/tests/candidate-execution-task-outcome.test.ts b/tests/candidate-execution-task-outcome.test.ts
index d370f4d0..47b75ea8 100644
--- a/tests/candidate-execution-task-outcome.test.ts
+++ b/tests/candidate-execution-task-outcome.test.ts
@@ -66,8 +66,8 @@ function changedOutcome(root: string, baseTree: string) {
 describe('candidate task outcome verification', () => {
   it('proves a captured binary patch and after-state against the signed base tree', async () => {
     const fixture = createCandidateExecutionFixture()
-    if (!fixture.task.repository) throw new Error('expected repository identity')
-    const repository = fixture.task.repository
+    if (!fixture.task.task.repository) throw new Error('expected repository identity')
+    const repository = fixture.task.task.repository
     const outcome = changedOutcome(fixture.task.stagingRoots.taskRoot, repository.baseTree)
     await expect(
       verifyTaskOutcomePatch({
@@ -84,8 +84,8 @@ describe('candidate task outcome verification', () => {
 
   it('rejects a claimed result tree or after-state that does not match the patch', async () => {
     const fixture = createCandidateExecutionFixture()
-    if (!fixture.task.repository) throw new Error('expected repository identity')
-    const repository = fixture.task.repository
+    if (!fixture.task.task.repository) throw new Error('expected repository identity')
+    const repository = fixture.task.task.repository
     const outcome = changedOutcome(fixture.task.stagingRoots.taskRoot, repository.baseTree)
     await expect(
       verifyTaskOutcomePatch({
diff --git a/tests/helpers/candidate-execution-fixture.ts b/tests/helpers/candidate-execution-fixture.ts
index 619282ef..07ed7533 100644
--- a/tests/helpers/candidate-execution-fixture.ts
+++ b/tests/helpers/candidate-execution-fixture.ts
@@ -2,9 +2,13 @@ import { execFileSync } from 'node:child_process'
 import { chmodSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from 'node:fs'
 import { tmpdir } from 'node:os'
 import { dirname, join } from 'node:path'
-
+import {
+  sealCandidateBenchmarkSuite,
+  sealCandidateBenchmarkTask,
+} from '@tangle-network/agent-eval/contract'
 import type {
   AgentCandidateArtifactRef,
+  AgentCandidateBenchmarkTask,
   AgentCandidateBundle,
   AgentCandidateExecution,
   AgentCandidateWorkspaceSnapshotEvidence,
@@ -13,6 +17,7 @@ import type {
 import {
   canonicalCandidateBytes,
   canonicalCandidateDigest,
+  canonicalCandidateDocument,
   embeddedCandidateArtifact,
   sha256Bytes,
 } from '../../src/candidate-execution/digest'
@@ -120,7 +125,7 @@ export function candidateBundle(
           baseCommit: active.commit,
           baseTree: active.tree,
         }
-      : { kind: 'disabled' as const, reason: 'control' as const },
+      : { kind: 'disabled' as const },
     execution: {
       harness: 'codex' as const,
       harnessVersion: '1.2.3',
@@ -143,29 +148,6 @@ export function candidateBundle(
       ...execution,
     },
     memory: { mode: 'disabled' as const },
-    lineage: active
-      ? {
-          source: 'optimizer' as const,
-          parentDigests: [candidateSha('e')],
-          runIds: ['optimizer-run-1'],
-          benchmark: {
-            name: 'development',
-            version: '1',
-            splitDigest: candidateSha('f'),
-          },
-          spend: {
-            proposal: { costUsd: 0, inputTokens: 0, outputTokens: 0, modelCalls: 0 },
-            evaluation: { costUsd: 0, inputTokens: 0, outputTokens: 0, modelCalls: 0 },
-          },
-        }
-      : {
-          source: 'human' as const,
-          benchmark: {
-            name: 'development',
-            version: '1',
-            splitDigest: candidateSha('f'),
-          },
-        },
   }
   return { ...value, digest: canonicalCandidateDigest(value) }
 }
@@ -179,6 +161,63 @@ export function redigestCandidateBundle(
   return { ...value, digest: canonicalCandidateDigest(value) }
 }
 
+export function bindCandidateFixtureBundle(
+  fixture: CandidateExecutionFixture,
+  bundle: AgentCandidateBundle = fixture.bundle,
+): void {
+  const { digest: _digest, ...cell } = fixture.task.runCell
+  fixture.bundle = bundle
+  fixture.task = {
+    ...fixture.task,
+    runCell: canonicalCandidateDocument({ ...cell, bundleDigest: bundle.digest }).value,
+  }
+}
+
+export function replaceCandidateFixtureTask(
+  fixture: CandidateExecutionFixture,
+  overrides: Partial<Omit<AgentCandidateBenchmarkTask, 'digest'>>,
+): void {
+  const material = { ...omitTaskDigest(fixture.task.task), ...overrides }
+  const task = sealCandidateBenchmarkTask(material)
+  const benchmark = sealCandidateBenchmarkSuite({
+    tasks: [task],
+    reps: fixture.task.benchmarkSuite.reps,
+    seeds: fixture.task.benchmarkSuite.seeds,
+  })
+  const { digest: _digest, ...cell } = fixture.task.runCell
+  fixture.task = {
+    ...fixture.task,
+    task,
+    benchmarkSuite: benchmark.suite,
+    runCell: canonicalCandidateDocument({
+      ...cell,
+      suiteDigest: benchmark.suite.digest,
+      taskDigest: task.digest,
+    }).value,
+  }
+}
+
+export function replaceCandidateFixtureAttempt(
+  fixture: CandidateExecutionFixture,
+  attempt: {
+    number: number
+    maxAttempts: number
+    retryPolicy: 'none' | 'pre-model-infrastructure-only'
+  },
+): void {
+  replaceCandidateFixtureTask(fixture, {
+    attempt: {
+      maxAttempts: attempt.maxAttempts,
+      retryPolicy: attempt.retryPolicy,
+    },
+  })
+  const { digest: _digest, ...cell } = fixture.task.runCell
+  fixture.task = {
+    ...fixture.task,
+    runCell: canonicalCandidateDocument({ ...cell, attempt: attempt.number }).value,
+  }
+}
+
 export function emptyCandidateSnapshot(label: string): AgentCandidateWorkspaceSnapshotEvidence {
   const material = {
     kind: 'agent-candidate-workspace-manifest' as const,
@@ -204,15 +243,17 @@ export interface CandidateExecutionFixture {
 export function unchangedTaskOutcomeCapture(
   fixture: CandidateExecutionFixture,
 ): AgentCandidateExecutorTaskOutcomeCapture {
-  if (fixture.task.outcome.kind !== 'workspace') {
+  if (fixture.task.task.outcome.kind !== 'workspace') {
     throw new Error('unchanged workspace capture requires a workspace outcome')
   }
-  if (!fixture.task.repository) throw new Error('workspace task is missing repository identity')
+  if (!fixture.task.task.repository) {
+    throw new Error('workspace task is missing repository identity')
+  }
   return {
     kind: 'workspace',
-    resultTree: fixture.task.repository.baseTree,
-    afterState: fixture.task.workspace.material,
-    archive: Buffer.from(`task-archive:${fixture.task.workspace.digest}`, 'utf8'),
+    resultTree: fixture.task.task.repository.baseTree,
+    afterState: fixture.task.task.workspace.material,
+    archive: Buffer.from(`task-archive:${fixture.task.task.workspace.digest}`, 'utf8'),
     gitDiff: Buffer.alloc(0),
   }
 }
@@ -301,6 +342,7 @@ export function createCandidateExecutionFixture(active = false): CandidateExecut
           }
           return
         }
+        if (role === 'knowledge' && workspace.material.files.length === 0) return
         const source = role === 'task' ? repository.root : candidateRoot
         if (!source) throw new Error(`fixture ${role} workspace source is missing`)
         if (source === destination) return
@@ -351,12 +393,25 @@ export function createCandidateExecutionFixture(active = false): CandidateExecut
       },
     },
   }
-  const task: AgentCandidateTaskExecution = {
-    executionId: 'execution-1',
-    benchmark: 'repository-disjoint-smoke',
-    benchmarkVersion: '1',
-    taskId: 'owner-repo-1',
-    splitDigest: candidateSha('d'),
+  const bundle = candidateBundle(
+    {},
+    active && candidateWorkspace
+      ? { commit: repository.commit, tree: repository.tree, workspace: candidateWorkspace }
+      : undefined,
+  )
+  const benchmarkTask = sealCandidateBenchmarkTask({
+    kind: 'agent-candidate-benchmark-task',
+    digestAlgorithm: 'rfc8785-sha256',
+    benchmark: {
+      name: 'repository-disjoint-smoke',
+      version: '1',
+      splitDigest: candidateSha('d'),
+    },
+    scenario: {
+      id: 'owner-repo-1',
+      kind: 'coding',
+      scenarioDigest: candidateSha('8'),
+    },
     instruction: 'Fix the failing behavior without changing the public API.',
     repository: {
       identity: 'github.com/owner/repo',
@@ -365,26 +420,28 @@ export function createCandidateExecutionFixture(active = false): CandidateExecut
       baseTree: repository.tree,
     },
     outcome: { kind: 'workspace' },
-    attempt: { number: 1, maxAttempts: 1, retryPolicy: 'none' },
-    model: { requested: 'provider/model', reasoningEffort: 'high' },
+    attempt: { maxAttempts: 1, retryPolicy: 'none' },
+    model: {
+      requested: 'provider/model',
+      provider: 'provider',
+      model: 'model-snapshot',
+      snapshot: 'model-snapshot-2026-07-01',
+      reasoningEffort: 'high',
+    },
     grader: {
       name: 'fixture-executable-grader',
       version: '1.0.0',
+      format: 'tangle-grader',
       artifact: {
-        locator: { kind: 's3', bucket: 'candidate-test-artifacts', key: 'grader/fixture' },
+        locator: {
+          kind: 's3',
+          bucket: 'candidate-test-artifacts',
+          key: `grader/${sha256Bytes(fixtureGraderBytes).slice('sha256:'.length)}`,
+        },
         sha256: sha256Bytes(fixtureGraderBytes),
         byteLength: fixtureGraderBytes.byteLength,
       },
     },
-    executionRoots: {
-      taskRoot: '/workspace/task',
-      ...(active ? { candidateRoot: '/opt/candidate' } : {}),
-    },
-    stagingRoots: {
-      taskRoot: repository.root,
-      ...(candidateRoot ? { candidateRoot } : {}),
-      profileRoot,
-    },
     workspace: taskWorkspace,
     evaluatorTaskContainer: selectedContainer,
     limits: {
@@ -395,14 +452,40 @@ export function createCandidateExecutionFixture(active = false): CandidateExecut
       maxOutputTokens: 50_000,
       maxCostUsd: 5,
     },
+  })
+  const benchmark = sealCandidateBenchmarkSuite({
+    tasks: [benchmarkTask],
+    reps: 1,
+    seeds: [42],
+  })
+  const task: AgentCandidateTaskExecution = {
+    executionId: 'execution-1',
+    runCell: canonicalCandidateDocument({
+      kind: 'agent-candidate-run-cell' as const,
+      experimentDigest: candidateSha('9'),
+      arm: 'candidate' as const,
+      bundleDigest: bundle.digest,
+      suiteDigest: benchmark.suite.digest,
+      taskDigest: benchmarkTask.digest,
+      taskIndex: 0,
+      repetition: 0,
+      seed: 42,
+      attempt: 1,
+    }).value,
+    benchmarkSuite: benchmark.suite,
+    task: benchmarkTask,
+    executionRoots: {
+      taskRoot: '/workspace/task',
+      ...(active ? { candidateRoot: '/opt/candidate' } : {}),
+    },
+    stagingRoots: {
+      taskRoot: repository.root,
+      ...(candidateRoot ? { candidateRoot } : {}),
+      profileRoot,
+    },
   }
   return {
-    bundle: candidateBundle(
-      {},
-      active && candidateWorkspace
-        ? { commit: repository.commit, tree: repository.tree, workspace: candidateWorkspace }
-        : undefined,
-    ),
+    bundle,
     task,
     ports,
     ...(candidateRoot ? { candidateRoot } : {}),
@@ -416,20 +499,59 @@ export function createCandidateOutputExecutionFixture(
 ): CandidateExecutionFixture {
   const fixture = createCandidateExecutionFixture()
   if (withRepository) {
-    return {
-      ...fixture,
-      task: { ...fixture.task, outcome: { kind: 'output', mediaType, maxBytes } },
-    }
+    return withBenchmarkTask(fixture, {
+      ...omitTaskDigest(fixture.task.task),
+      outcome: { kind: 'output', mediaType, maxBytes },
+    })
   }
   const taskRoot = temporaryRoot('candidate-output-task-')
-  const { repository: _repository, ...taskWithoutRepository } = fixture.task
-  return {
-    ...fixture,
-    task: {
+  const { repository: _repository, ...taskWithoutRepository } = omitTaskDigest(fixture.task.task)
+  return withBenchmarkTask(
+    {
+      ...fixture,
+      task: {
+        ...fixture.task,
+        stagingRoots: { ...fixture.task.stagingRoots, taskRoot },
+      },
+    },
+    {
       ...taskWithoutRepository,
       outcome: { kind: 'output', mediaType, maxBytes },
       workspace: emptyCandidateSnapshot('output-task'),
-      stagingRoots: { ...fixture.task.stagingRoots, taskRoot },
+    },
+  )
+}
+
+function omitTaskDigest(
+  task: AgentCandidateBenchmarkTask,
+): Omit<AgentCandidateBenchmarkTask, 'digest'> {
+  const { digest: _digest, ...material } = task
+  return material
+}
+
+function withBenchmarkTask(
+  fixture: CandidateExecutionFixture,
+  material: Omit<AgentCandidateBenchmarkTask, 'digest'>,
+): CandidateExecutionFixture {
+  const task = sealCandidateBenchmarkTask(material)
+  const benchmark = sealCandidateBenchmarkSuite({
+    tasks: [task],
+    reps: fixture.task.benchmarkSuite.reps,
+    seeds: fixture.task.benchmarkSuite.seeds,
+  })
+  const { digest: _cellDigest, ...cell } = fixture.task.runCell
+  const runCell = canonicalCandidateDocument({
+    ...cell,
+    suiteDigest: benchmark.suite.digest,
+    taskDigest: task.digest,
+  }).value
+  return {
+    ...fixture,
+    task: {
+      ...fixture.task,
+      runCell,
+      benchmarkSuite: benchmark.suite,
+      task,
     },
   }
 }
diff --git a/tests/helpers/candidate-experiment-fixture.ts b/tests/helpers/candidate-experiment-fixture.ts
new file mode 100644
index 00000000..ee31da70
--- /dev/null
+++ b/tests/helpers/candidate-experiment-fixture.ts
@@ -0,0 +1,199 @@
+import { mkdtempSync, rmSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+
+import { InMemoryTraceStore } from '@tangle-network/agent-eval'
+import {
+  type CandidateExperimentExecutionInput,
+  sealCandidateBenchmarkSuite,
+  sealCandidateExperiment,
+} from '@tangle-network/agent-eval/contract'
+import type {
+  AgentCandidateBundle,
+  AgentCandidateExperiment,
+  CandidateExecutionEvidence,
+} from '@tangle-network/agent-interface'
+
+import { InMemoryAgentCandidateExecutionClaimStore } from '../../src/candidate-execution/claim'
+import { sha256Bytes } from '../../src/candidate-execution/digest'
+import type { AgentCandidateExperimentCellPlacement } from '../../src/intelligence/improvement-cycle'
+import {
+  type CandidateExecutionFixture,
+  candidateSha,
+  createCandidateOutputExecutionFixture,
+  createCandidateOutputFixture,
+  emptyCandidateSnapshot,
+  redigestCandidateBundle,
+} from './candidate-execution-fixture'
+
+const roots: string[] = []
+
+export interface CandidateExperimentFixture {
+  fixture: CandidateExecutionFixture
+  experiment: AgentCandidateExperiment
+  placeCell(input: CandidateExperimentExecutionInput): AgentCandidateExperimentCellPlacement
+}
+
+export interface CreateCandidateExperimentFixtureOptions {
+  baseline?: AgentCandidateBundle
+  candidate?: AgentCandidateBundle
+  candidateSurface?: 'prompt' | 'memory'
+  reps?: number
+  scoreFor?: (input: CandidateExperimentExecutionInput) => number
+  configureFixture?: (fixture: CandidateExecutionFixture) => void
+}
+
+export function cleanupCandidateExperimentFixtures(): void {
+  for (const root of roots.splice(0)) rmSync(root, { recursive: true, force: true })
+}
+
+export function createCandidateExperimentFixture(
+  options: CreateCandidateExperimentFixtureOptions = {},
+): CandidateExperimentFixture {
+  const fixture = createCandidateOutputExecutionFixture('text/plain', 1_024)
+  options.configureFixture?.(fixture)
+  const baseline = options.baseline ?? fixture.bundle
+  const candidate =
+    options.candidate ??
+    (options.candidateSurface === 'memory'
+      ? redigestCandidateBundle(baseline, { memory: { mode: 'isolated', scope: 'task' } })
+      : redigestCandidateBundle(baseline, {
+          profile: {
+            ...baseline.profile,
+            prompt: {
+              ...baseline.profile.prompt,
+              systemPrompt: 'Return the exact measured answer.',
+            },
+          },
+        }))
+  const reps = options.reps ?? 3
+  const seeds = Array.from({ length: reps }, (_, index) => 101 + index) as [number, ...number[]]
+  const experiment = sealCandidateExperiment({
+    kind: 'agent-candidate-experiment',
+    digestAlgorithm: 'rfc8785-sha256',
+    baseline,
+    candidate,
+    candidateLineage: { source: 'human' },
+    benchmark: sealCandidateBenchmarkSuite({ tasks: [fixture.task.task], reps, seeds }),
+    policy: {
+      confidenceLevel: 0.95,
+      resamples: 500,
+      bootstrapSeed: 1_337,
+      deltaThreshold: 0,
+      minProductiveRuns: 3,
+      budgetUsd: 1,
+      criticalDimensions: ['quality'],
+      regressionTolerance: 0.05,
+    },
+  })
+  const memoryBefore = emptyCandidateSnapshot('experiment-memory-before')
+  fixture.ports.memory.reset = async ({ preparationId, expiresAtMs }) => ({
+    preparationId,
+    accessDigest: candidateSha('8'),
+    expiresAtMs,
+    evidence: memoryBefore.manifest,
+    emptyStateDigest: memoryBefore.digest,
+    beforeState: memoryBefore,
+  })
+  fixture.ports.memory.activate = async ({ effectiveNamespace }) => ({
+    env: { TANGLE_MEMORY_NAMESPACE: effectiveNamespace },
+  })
+  fixture.ports.memory.close = async () => ({ closed: true })
+
+  return {
+    fixture,
+    experiment,
+    placeCell(input) {
+      const taskRoot = temporaryRoot('candidate-experiment-task-')
+      const profileRoot = temporaryRoot('candidate-experiment-profile-')
+      const traceStore = new InMemoryTraceStore()
+      const outputs = createCandidateOutputFixture()
+      const score = options.scoreFor?.(input) ?? (input.arm === 'candidate' ? 1 : 0)
+      const output = Buffer.from(score === 1 ? 'strong' : 'weak', 'utf8')
+      return {
+        executionId: [
+          'experiment',
+          input.arm,
+          input.benchmarkCell.taskIndex,
+          input.benchmarkCell.repetition,
+        ].join('-'),
+        executionRoots: { taskRoot: '/workspace/task' },
+        stagingRoots: { taskRoot, profileRoot },
+        ports: fixture.ports,
+        execution: {
+          executor: {
+            execute: async (request, context) => {
+              const startedAt = 1_000 + input.benchmarkCell.repetition * 100
+              await context.traceStore.appendRun({
+                runId: request.trace.runId,
+                scenarioId: request.benchmark.task.scenario.id,
+                startedAt,
+                endedAt: startedAt + 50,
+                status: 'completed',
+                tags: { ...request.trace.tags },
+              })
+              return {
+                executionId: request.executionId,
+                termination: { kind: 'exit' as const, exitCode: 0 },
+              }
+            },
+            stop: async () => ({ stopped: true }),
+            capture: async () => ({
+              taskOutcome: { kind: 'output' as const, bytes: output },
+              ...(input.bundle.memory.mode === 'isolated'
+                ? {
+                    memoryAfter: {
+                      afterState: memoryBefore.material,
+                      archive: Buffer.from('empty experiment memory', 'utf8'),
+                    },
+                  }
+                : {}),
+            }),
+          },
+          grader: {
+            ...outputs.grader,
+            run: async ({ implementation, outcome, signal }) => {
+              signal.throwIfAborted()
+              if (outcome.kind !== 'output') throw new Error('fixture grader requires output')
+              const measured = Buffer.from(outcome.bytes).toString('utf8') === 'strong' ? 1 : 0
+              const evidence = Buffer.from(JSON.stringify({ score: measured }), 'utf8')
+              return {
+                evaluation: {
+                  score: measured,
+                  passed: measured === 1,
+                  dimensions: { quality: measured },
+                  raw: {},
+                },
+                evidence,
+                binding: {
+                  implementationDigest: sha256Bytes(implementation.bytes),
+                  taskOutcomeDigest: outcome.evidence.digest,
+                  outputDigest: sha256Bytes(evidence),
+                },
+              }
+            },
+          },
+          outputArtifacts: outputs.outputArtifacts,
+          traceStore,
+          claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
+        },
+      }
+    },
+  }
+}
+
+export async function executeCandidateExperimentInput(
+  input: CandidateExperimentExecutionInput,
+  placeCell: CandidateExperimentFixture['placeCell'],
+  execute: (
+    input: CandidateExperimentExecutionInput & AgentCandidateExperimentCellPlacement,
+  ) => Promise<CandidateExecutionEvidence>,
+): Promise<CandidateExecutionEvidence> {
+  return await execute({ ...input, ...placeCell(input) })
+}
+
+function temporaryRoot(prefix: string): string {
+  const root = mkdtempSync(join(tmpdir(), prefix))
+  roots.push(root)
+  return root
+}
diff --git a/tests/improvement-cycle.test.ts b/tests/improvement-cycle.test.ts
index dc00e79e..77d1df13 100644
--- a/tests/improvement-cycle.test.ts
+++ b/tests/improvement-cycle.test.ts
@@ -1,53 +1,25 @@
-import { type AnalystFinding, InMemoryTraceStore } from '@tangle-network/agent-eval'
-import type {
-  CodeSurface,
-  DispatchContext,
-  JudgeConfig,
-  MutableSurface,
-  Scenario,
-  SurfaceProposer,
-} from '@tangle-network/agent-eval/contract'
-import type { CandidateExecutionEvidence } from '@tangle-network/agent-interface'
-import { afterEach, describe, expect, it, vi } from 'vitest'
-import type { AnalystRegistryLike } from '../src/analyst-loop/types'
-import { buildAgentCandidateBundle } from '../src/candidate-execution/builder'
-import { sealAgentCandidateBundle } from '../src/candidate-execution/bundle'
-import { InMemoryAgentCandidateExecutionClaimStore } from '../src/candidate-execution/claim'
-import {
-  canonicalCandidateDocument,
-  embeddedCandidateArtifact,
-} from '../src/candidate-execution/digest'
-import { assertCandidateProfileBinding } from '../src/candidate-execution/profile'
-import type {
-  AgentCandidateExecutorPort,
-  AgentCandidateExecutorRequest,
-} from '../src/candidate-execution/types'
+import type { AnalystFinding } from '@tangle-network/agent-eval'
+import { afterEach, describe, expect, it } from 'vitest'
+
+import { canonicalCandidateDigest } from '../src/candidate-execution/digest'
+import { agentCandidateProfileAsAgentProfile } from '../src/candidate-execution/profile'
 import {
-  createAgentImprovementMeasuredComparison,
+  createAgentImprovementActivation,
   createAgentImprovementProposal,
-  executeApprovedAgentCandidate,
   proposeAgentImprovement,
   reviewAgentImprovementProposal,
+  runAgentCandidateExperiment,
+  verifyAgentImprovementActivation,
   verifyAgentImprovementProposal,
   verifyAgentImprovementReview,
   verifyCandidateExecutionEvidence,
 } from '../src/intelligence/improvement-cycle'
+import { cleanupCandidateFixtures } from './helpers/candidate-execution-fixture'
 import {
-  candidateSha,
-  cleanupCandidateFixtures,
-  createCandidateExecutionFixture,
-  createCandidateOutputFixture,
-  unchangedTaskOutcomeCapture,
-} from './helpers/candidate-execution-fixture'
-
-interface DemoScenario extends Scenario {
-  kind: 'demo'
-}
-
-const scenarios: DemoScenario[] = Array.from({ length: 12 }, (_, index) => ({
-  id: `scenario-${index}`,
-  kind: 'demo' as const,
-}))
+  type CandidateExperimentFixture,
+  cleanupCandidateExperimentFixtures,
+  createCandidateExperimentFixture,
+} from './helpers/candidate-experiment-fixture'
 
 const finding: AnalystFinding = {
   schema_version: '1.0.0',
@@ -56,855 +28,240 @@ const finding: AnalystFinding = {
   produced_at: '2026-07-10T00:00:00.000Z',
   severity: 'high',
   area: 'prompt',
-  claim: 'The agent omits the required marker.',
+  claim: 'The agent omits the required answer.',
   evidence_refs: [{ kind: 'span', id: 'span-1' }],
-  recommended_action: 'Add the measured marker.',
+  recommended_action: 'Return the measured answer.',
   confidence: 0.9,
   subject: 'agent-profile:prompt.systemPrompt',
 }
 
-const registry = (findings: AnalystFinding[]): AnalystRegistryLike => ({
-  list: () => [{ id: 'improvement' }],
-  run: async (runId) => ({
-    run_id: runId,
-    correlation_id: `correlation-${runId}`,
-    started_at: '2026-07-10T00:00:00.000Z',
-    ended_at: '2026-07-10T00:00:01.000Z',
-    findings,
-    per_analyst: [
-      {
-        analyst_id: 'improvement',
-        status: 'ok',
-        findings_count: findings.length,
-        latency_ms: 1,
-        cost_usd: 0,
-      },
-    ],
-    total_cost_usd: 0,
-  }),
-})
-
-const proposer: SurfaceProposer = {
-  kind: 'scripted',
-  propose: async () => [
-    { surface: 'PROMOTED', label: 'measured winner', rationale: 'addresses finding-1' },
-  ],
-}
-
-const judge: JudgeConfig<string, DemoScenario> = {
-  name: 'literal-marker',
-  dimensions: [{ key: 'marker', description: 'Contains the required marker.' }],
-  score: ({ artifact }) => {
-    const composite = artifact.includes('PROMOTED') ? 1 : 0
-    return { dimensions: { marker: composite }, composite, notes: '' }
-  },
-}
-
-async function agent(
-  surface: MutableSurface,
-  _scenario: DemoScenario,
-  context: DispatchContext,
-): Promise<string> {
-  const paid = await context.cost.runPaidCall({
-    channel: 'agent',
-    actor: 'fixture',
-    model: 'fixture-model',
-    maximumCharge: { externallyEnforcedMaximumUsd: 0.0001 },
-    execute: async () => String(surface),
-    receipt: () => ({
-      model: 'fixture-model',
-      inputTokens: 1,
-      outputTokens: 1,
-      actualCostUsd: 0.0001,
-    }),
-  })
-  if (!paid.succeeded) throw paid.error
-  return paid.value
-}
-
-function fixtureProfile() {
-  return {
-    name: 'candidate',
-    prompt: {
-      systemPrompt: 'BASELINE',
-      instructions: ['Inspect the repository, implement the fix, and run tests.'],
-    },
-    model: { default: 'provider/model', reasoningEffort: 'high' as const },
-    harness: 'codex' as const,
-    resources: { failOnError: true as const },
-  }
-}
-
-function alignedBundle(
-  bundle: Omit<ReturnType<typeof createCandidateExecutionFixture>['bundle'], 'digest'>,
-  profile: { prompt?: { systemPrompt?: string } },
-) {
-  return {
-    ...bundle,
-    profile: {
-      ...bundle.profile,
-      prompt: {
-        ...bundle.profile.prompt,
-        systemPrompt: profile.prompt?.systemPrompt,
-      },
-    },
-  }
-}
-
-function alignedSealedBundle(
-  bundle: Omit<ReturnType<typeof createCandidateExecutionFixture>['bundle'], 'digest'>,
-  profile: { prompt?: { systemPrompt?: string } },
-) {
-  const aligned = alignedBundle(bundle, profile)
-  const { profileDiffIds: _profileDiffIds, ...lineage } = aligned.lineage
-  return buildAgentCandidateBundle({
-    profile: { kind: 'candidate-profile', profile: aligned.profile },
-    code: aligned.code,
-    execution: aligned.execution,
-    memory: aligned.memory,
-    lineage,
-  })
-}
-
 afterEach(() => {
+  cleanupCandidateExperimentFixtures()
   cleanupCandidateFixtures()
-  vi.restoreAllMocks()
 })
 
 describe('agent improvement lifecycle', () => {
-  it('refuses memory persistence before human approval', async () => {
-    const writeBack = vi.fn()
-    await expect(
-      proposeAgentImprovement({
-        runId: 'analysis-run-memory-writeback',
-        profile: fixtureProfile(),
-        analysis: { registry: registry([finding]), inputs: {}, findingsStore: null, log: () => {} },
-        improvement: {
-          surface: 'memory',
-          memory: {
-            document: '# Durable memory\n',
-            writeBack,
-          },
-          generator: proposer,
-          scenarios,
-          judge,
-          agent,
-          budget: { generations: 1, populationSize: 1, holdoutFraction: 0.5 },
-        },
-      }),
-    ).rejects.toThrow('cannot write memory before human approval')
-    expect(writeBack).not.toHaveBeenCalled()
-  })
-
-  it('disposes a retained code winner when candidate construction fails', async () => {
-    const { execFileSync } = await import('node:child_process')
-    const { mkdtempSync, readFileSync, rmSync, writeFileSync } = await import('node:fs')
-    const { tmpdir } = await import('node:os')
-    const { join } = await import('node:path')
-    const repoRoot = mkdtempSync(join(tmpdir(), 'improvement-cycle-code-cleanup-'))
-    const git = (...args: string[]) =>
-      execFileSync('git', args, { cwd: repoRoot, encoding: 'utf8', stdio: 'pipe' })
-    try {
-      git('init', '-q', '-b', 'main')
-      git('config', 'user.email', 'improvement-cycle@test.local')
-      git('config', 'user.name', 'improvement-cycle-test')
-      writeFileSync(join(repoRoot, 'module.txt'), 'BASELINE\n')
-      git('add', 'module.txt')
-      git('commit', '-qm', 'baseline')
-
-      const buildCandidate = vi.fn(async () => {
-        throw new Error('candidate construction failed')
-      })
-      await expect(
-        proposeAgentImprovement({
-          runId: 'analysis-run-code-build-failure',
-          profile: fixtureProfile(),
-          analysis: {
-            registry: registry([finding]),
-            inputs: {},
-            findingsStore: null,
-            log: () => {},
-          },
-          improvement: {
-            surface: 'code',
-            scenarios,
-            judge: {
-              name: 'code-marker',
-              dimensions: [{ key: 'marker', description: 'Code contains the promoted marker.' }],
-              score: ({ artifact }) => {
-                const composite = artifact.includes('PROMOTED') ? 1 : 0
-                return { dimensions: { marker: composite }, composite, notes: '' }
-              },
-            },
-            agent: async (surface, _scenario, context) => {
-              const paid = await context.cost.runPaidCall({
-                channel: 'agent',
-                actor: 'fixture',
-                model: 'fixture-model',
-                maximumCharge: { externallyEnforcedMaximumUsd: 0.0001 },
-                execute: async () => {
-                  if (typeof surface === 'string') throw new Error('expected code surface')
-                  return readFileSync(join(surface.worktreeRef, 'module.txt'), 'utf8')
-                },
-                receipt: () => ({
-                  model: 'fixture-model',
-                  inputTokens: 1,
-                  outputTokens: 1,
-                  actualCostUsd: 0.0001,
-                }),
-              })
-              if (!paid.succeeded) throw paid.error
-              return paid.value
-            },
-            code: {
-              repoRoot,
-              generator: {
-                kind: 'fixture',
-                async generate({ worktreePath }) {
-                  writeFileSync(join(worktreePath, 'module.txt'), 'PROMOTED\n')
-                  return { applied: true, summary: 'write promoted marker' }
-                },
-              },
-            },
-            promotionGate: {
-              name: 'fixture-ship',
-              decide: async () => ({
-                decision: 'ship',
-                reasons: ['candidate passes the fixture comparison'],
-                contributingGates: [],
-              }),
-            },
-            budget: { generations: 1, populationSize: 1, holdoutFraction: 0.25 },
-          },
-          buildCandidate,
-        }),
-      ).rejects.toThrow('candidate construction failed')
-
-      expect(buildCandidate).toHaveBeenCalledOnce()
-      expect(git('worktree', 'list', '--porcelain').match(/^worktree /gm)).toHaveLength(1)
-    } finally {
-      rmSync(repoRoot, { recursive: true, force: true })
-    }
-  })
-
-  it('binds typed candidate hooks to their equivalent measured profile commands', () => {
-    expect(() =>
-      assertCandidateProfileBinding(
-        {
-          hooks: {
-            beforeTool: [
-              {
-                command: "node 'path with space.js'",
-                timeoutMs: 1_000,
-                env: { MODE: 'check' },
-              },
-            ],
-          },
-        },
-        {
-          hooks: {
-            beforeTool: [
-              {
-                executable: 'node',
-                args: [{ kind: 'public', value: 'path with space.js' }],
-                timeoutMs: 1_000,
-                env: { MODE: { kind: 'public', value: 'check' } },
-              },
-            ],
-          },
-        },
-      ),
-    ).not.toThrow()
-  })
-
-  it('analyzes, measures, approves, executes, grades, and links one exact receipt', async () => {
-    const fixture = createCandidateExecutionFixture()
-    const { digest: _digest, ...bundleInput } = fixture.bundle
-    const profile = fixtureProfile()
-    const proposed = await proposeAgentImprovement({
-      runId: 'analysis-run-1',
-      profile,
-      analysis: { registry: registry([finding]), inputs: {}, findingsStore: null, log: () => {} },
-      improvement: {
-        surface: 'prompt',
-        generator: proposer,
-        scenarios,
-        judge,
-        agent,
-        budget: { generations: 1, populationSize: 2, reps: 3, holdoutFraction: 0.5 },
+  it('measures the exact paired matrix before review and activation', async () => {
+    const rig = createCandidateExperimentFixture()
+    const executed: string[] = []
+    const result = await runAgentCandidateExperiment({
+      experiment: rig.experiment,
+      runId: 'paired-run-1',
+      maxConcurrency: 2,
+      placeCell: (input) => {
+        executed.push(`${input.arm}:${input.benchmarkCell.repetition}`)
+        return rig.placeCell(input)
       },
-      buildCandidate: ({ improvement }) => alignedSealedBundle(bundleInput, improvement.profile),
-      now: () => new Date('2026-07-10T01:00:00.000Z'),
     })
 
-    expect(proposed.proposal.evaluation.decision.outcome).toBe('ship')
-    expect(proposed.proposal.evaluation.overall.delta).toBeGreaterThan(0)
-    expect(proposed.proposal.changedSurfaces).toEqual(['prompt'])
-    expect(proposed.proposal.findings).toEqual([finding])
-    expect(proposed.proposal.candidateBundle?.profile.prompt?.systemPrompt).toBe('PROMOTED')
-    expect(proposed.proposal.candidateBundle?.digest).toMatch(/^sha256:[a-f0-9]{64}$/)
-    expect(verifyAgentImprovementProposal(proposed.proposal)).toEqual(proposed.proposal)
-    expect(() =>
-      verifyAgentImprovementProposal({ ...proposed.proposal, runId: 'tampered-run' }),
-    ).toThrow(/proposal digest does not match/)
-    expect(
-      createAgentImprovementProposal({
-        runId: 'analysis-run-1',
-        baselineProfile: profile,
-        findings: proposed.analysis.analystResult.findings,
-        evaluation: proposed.proposal.evaluation,
-        candidateBundle: proposed.proposal.candidateBundle,
-        now: () => new Date('2026-07-10T01:00:00.000Z'),
-      }),
-    ).toEqual(proposed.proposal)
-    expect(() =>
-      createAgentImprovementProposal({
-        runId: 'analysis-run-unmeasured',
-        baselineProfile: profile,
-        findings: proposed.analysis.analystResult.findings,
-        evaluation: proposed.proposal.evaluation,
-        candidateBundle: alignedBundle(bundleInput, {
-          ...proposed.improvement.profile,
-          prompt: { systemPrompt: 'UNMEASURED' },
-        }),
-      }),
-    ).toThrow(/does not bind the exact candidate bundle/)
-    expect(() =>
-      createAgentImprovementProposal({
-        runId: 'analysis-run-baseline-drift',
-        baselineProfile: { ...profile, name: 'different-baseline' },
-        findings: proposed.analysis.analystResult.findings,
-        evaluation: proposed.proposal.evaluation,
-        candidateBundle: proposed.proposal.candidateBundle,
-      }),
-    ).toThrow(/does not bind the included baseline profile/)
+    expect(executed.sort()).toEqual([
+      'baseline:0',
+      'baseline:1',
+      'baseline:2',
+      'candidate:0',
+      'candidate:1',
+      'candidate:2',
+    ])
+    expect(result.measurements).toHaveLength(3)
+    expect(result.evaluation).toMatchObject({
+      experiment: { digest: rig.experiment.digest },
+      overall: { baseline: 0, candidate: 1, delta: 1, n: 3 },
+      decision: { outcome: 'ship' },
+      evaluation: { executionCostUsd: 0, searchCostUsd: 0, totalCostUsd: 0 },
+    })
 
-    const firstMeasuredCell = proposed.improvement.raw.raw.baselineOnHoldout.cells[0]
-    if (!firstMeasuredCell) throw new Error('expected a measured heldout cell')
-    const primaryObjective = Object.keys(firstMeasuredCell.judgeScores)[0]
-    const primaryScore = primaryObjective
-      ? firstMeasuredCell.judgeScores[primaryObjective]
-      : undefined
-    const primaryDimension = primaryScore ? Object.keys(primaryScore.dimensions)[0] : undefined
-    if (!primaryObjective || !primaryScore || !primaryDimension) {
-      throw new Error('expected a measured objective and dimension')
-    }
-    const addSecondJudge = (cell: typeof firstMeasuredCell) => ({
-      ...cell,
-      judgeScores: {
-        ...cell.judgeScores,
-        secondary: {
-          composite: cell.judgeScores[primaryObjective]!.composite,
-          dimensions: { ...cell.judgeScores[primaryObjective]!.dimensions },
-        },
-      },
+    const proposal = createAgentImprovementProposal({
+      runId: 'paired-run-1',
+      findings: [finding],
+      evaluation: result.evaluation,
+      now: () => new Date('2026-07-10T01:00:00.000Z'),
     })
-    const multiJudgeResult = {
-      ...proposed.improvement.raw,
-      raw: {
-        ...proposed.improvement.raw.raw,
-        baselineOnHoldout: {
-          ...proposed.improvement.raw.raw.baselineOnHoldout,
-          cells: proposed.improvement.raw.raw.baselineOnHoldout.cells.map(addSecondJudge),
-        },
-        winnerOnHoldout: {
-          ...proposed.improvement.raw.raw.winnerOnHoldout,
-          cells: proposed.improvement.raw.raw.winnerOnHoldout.cells.map(addSecondJudge),
-        },
-      },
+    expect(proposal.changedSurfaces).toEqual(['prompt'])
+    expect(verifyAgentImprovementProposal(proposal)).toEqual(proposal)
+    const { digest: _proposalDigest, ...proposalWithoutDigest } = proposal
+    const judgeDerivedProposalWithoutDigest = {
+      ...proposalWithoutDigest,
+      findings: proposal.findings.map((item) => ({ ...item, derived_from_judge: true })),
     }
-    const multiJudgeComparison = createAgentImprovementMeasuredComparison({
-      result: multiJudgeResult,
-      measuredSurface: 'prompt',
-      baselineProfile: profile,
-      candidateBundle: proposed.proposal.candidateBundle,
-      metadata: {
-        executionTask: {
-          itemId: 'scenario-1',
-          instructionDigest: `sha256:${'c'.repeat(64)}`,
-        },
-      },
-    })
-    expect(multiJudgeComparison.metadata).toEqual({
-      executionTask: {
-        itemId: 'scenario-1',
-        instructionDigest: `sha256:${'c'.repeat(64)}`,
-      },
+    const judgeDerivedProposal = {
+      ...judgeDerivedProposalWithoutDigest,
+      digest: canonicalCandidateDigest(judgeDerivedProposalWithoutDigest),
+    }
+    expect(() => verifyAgentImprovementProposal(judgeDerivedProposal)).toThrow(/judge-derived/)
+
+    const review = reviewAgentImprovementProposal(proposal, {
+      decision: 'approve',
+      reviewedBy: 'operator@example.com',
+      reason: 'All three paired tasks improved.',
+      now: () => new Date('2026-07-10T01:01:00.000Z'),
     })
-    expect(multiJudgeComparison.objectives).toEqual(
-      expect.arrayContaining([
-        expect.objectContaining({ kind: 'objective', name: primaryObjective }),
-        expect.objectContaining({ kind: 'objective', name: 'secondary' }),
-        expect.objectContaining({
-          kind: 'dimension',
-          objective: primaryObjective,
-          name: primaryDimension,
-        }),
-        expect.objectContaining({
-          kind: 'dimension',
-          objective: 'secondary',
-          name: primaryDimension,
-        }),
-      ]),
-    )
+    expect(verifyAgentImprovementReview(review)).toEqual(review)
 
-    const constantScoreResult = {
-      ...proposed.improvement.raw,
-      baseline: {
-        ...proposed.improvement.raw.baseline,
-        compositeMean: 0.5,
-        perScenario: Object.fromEntries(scenarios.map(({ id }) => [id, 0.5])),
-      },
-      winner: {
-        ...proposed.improvement.raw.winner,
-        compositeMean: 0.8,
-        perScenario: Object.fromEntries(scenarios.map(({ id }) => [id, 0.8])),
-      },
-      lift: 0.8 - 0.5,
-      provenance: {
-        ...proposed.improvement.raw.provenance,
-        baselineHoldoutComposite: 0.5,
-        winnerHoldoutComposite: 0.8,
-        heldOutLift: 0.8 - 0.5,
-      },
-      raw: {
-        ...proposed.improvement.raw.raw,
-        baselineOnHoldout: {
-          ...proposed.improvement.raw.raw.baselineOnHoldout,
-          cells: proposed.improvement.raw.raw.baselineOnHoldout.cells.map((cell) => ({
-            ...cell,
-            judgeScores: Object.fromEntries(
-              Object.entries(cell.judgeScores).map(([name, score]) => [
-                name,
-                {
-                  ...score,
-                  composite: 0.5,
-                  dimensions: Object.fromEntries(
-                    Object.keys(score.dimensions).map((dimension) => [dimension, 0.5]),
-                  ),
-                },
-              ]),
-            ),
-          })),
-        },
-        winnerOnHoldout: {
-          ...proposed.improvement.raw.raw.winnerOnHoldout,
-          cells: proposed.improvement.raw.raw.winnerOnHoldout.cells.map((cell) => ({
-            ...cell,
-            judgeScores: Object.fromEntries(
-              Object.entries(cell.judgeScores).map(([name, score]) => [
-                name,
-                {
-                  ...score,
-                  composite: 0.8,
-                  dimensions: Object.fromEntries(
-                    Object.keys(score.dimensions).map((dimension) => [dimension, 0.8]),
-                  ),
-                },
-              ]),
-            ),
-          })),
+    const activation = createAgentImprovementActivation(proposal, review, {
+      targets: [
+        {
+          surface: 'prompt',
+          identity: 'tenant/default/profile',
+          expectedBaseDigest: promptSurfaceDigest(rig),
         },
-      },
-    }
-    const constantScoreComparison = createAgentImprovementMeasuredComparison({
-      result: constantScoreResult,
-      measuredSurface: 'prompt',
-      baselineProfile: profile,
-      candidateBundle: proposed.proposal.candidateBundle,
-    })
-    expect(constantScoreComparison.overall.confidenceInterval).toMatchObject({
-      lower: constantScoreComparison.overall.delta,
-      upper: constantScoreComparison.overall.delta,
+      ],
+      fundingOwner: 'tenant/default',
+      authorizedBy: 'operator@example.com',
+      now: () => new Date('2026-07-10T01:02:00.000Z'),
     })
+    expect(verifyAgentImprovementActivation({ proposal, review, activation })).toEqual(activation)
+  })
 
-    const compoundProfile = {
-      ...proposed.improvement.profile,
-      tools: { inspect_repository: true },
-    }
-    const compoundInput = alignedBundle(bundleInput, compoundProfile)
-    const compoundBundle = sealAgentCandidateBundle({
-      ...compoundInput,
-      profile: { ...compoundInput.profile, tools: compoundProfile.tools },
+  it('runs isolated memory as a real candidate surface', async () => {
+    const rig = createCandidateExperimentFixture({ candidateSurface: 'memory' })
+    const result = await runAgentCandidateExperiment({
+      experiment: rig.experiment,
+      runId: 'memory-run-1',
+      placeCell: rig.placeCell,
     })
-    const compoundComparison = createAgentImprovementMeasuredComparison({
-      result: {
-        ...proposed.improvement.raw,
-        winner: {
-          ...proposed.improvement.raw.winner,
-          surface: JSON.stringify(compoundProfile),
-        },
-      },
-      measuredSurface: 'agent-profile',
-      baselineProfile: profile,
-      candidateBundle: compoundBundle,
+    const proposal = createAgentImprovementProposal({
+      runId: 'memory-run-1',
+      findings: [finding],
+      evaluation: result.evaluation,
     })
-    const compoundProposal = createAgentImprovementProposal({
-      runId: 'analysis-run-compound-profile',
-      baselineProfile: profile,
-      findings: proposed.analysis.analystResult.findings,
-      evaluation: compoundComparison,
-      candidateBundle: compoundBundle,
-    })
-    expect(compoundProposal.changedSurfaces).toEqual(['prompt', 'tools'])
-    const { digest: _compoundDigest, ...compoundWithoutDigest } = compoundProposal
-    const omittedSurfaceProposal = canonicalCandidateDocument<typeof compoundProposal>({
-      ...compoundWithoutDigest,
-      changedSurfaces: ['prompt'],
-    }).value
-    expect(() => verifyAgentImprovementProposal(omittedSurfaceProposal)).toThrow(
-      /changed surfaces do not match/,
-    )
 
-    const codeFixture = createCandidateExecutionFixture(true)
-    const { digest: _codeDigest, ...codeBundleInput } = codeFixture.bundle
-    if (!codeFixture.task.repository) throw new Error('expected repository fixture')
-    const patch = Buffer.from('diff --git a/source.ts b/source.ts\n', 'utf8')
-    const candidateTree = 'f'.repeat(codeFixture.task.repository.baseTree.length)
-    const codeSurface: CodeSurface = {
-      kind: 'code',
-      worktreeRef: '/tmp/measured-code-winner',
-      baseRef: 'main',
-      baseCommit: codeFixture.task.repository.baseCommit,
-      baseTree: codeFixture.task.repository.baseTree,
-      candidateCommit: 'e'.repeat(codeFixture.task.repository.baseCommit.length),
-      candidateTree,
-      patch: {
-        format: 'git-diff-binary',
-        sha256: embeddedCandidateArtifact(patch).sha256,
-        byteLength: patch.byteLength,
-      },
+    expect(proposal.changedSurfaces).toEqual(['memory'])
+    expect(result.measurements).toHaveLength(3)
+    for (const measurement of result.measurements) {
+      expect(measurement.baseline.receipt.memory.mode).toBe('disabled')
+      expect(measurement.candidate.receipt.memory.mode).toBe('isolated')
     }
-    const codeBundle = sealAgentCandidateBundle({
-      ...alignedBundle(codeBundleInput, profile),
-      code: {
-        kind: 'git-patch',
-        repository: { kind: 'github', owner: 'owner', repo: 'repo' },
-        baseCommit: codeSurface.baseCommit,
-        baseTree: codeSurface.baseTree,
-        candidateTree: codeSurface.candidateTree,
-        patch: { format: 'git-diff-binary', artifact: embeddedCandidateArtifact(patch) },
-      },
+  })
+
+  it('rejects receipt substitution and stale activation targets', async () => {
+    const rig = createCandidateExperimentFixture()
+    const result = await runAgentCandidateExperiment({
+      experiment: rig.experiment,
+      runId: 'binding-run-1',
+      placeCell: rig.placeCell,
     })
-    const codeResult = {
-      ...proposed.improvement.raw,
-      winner: { ...proposed.improvement.raw.winner, surface: codeSurface },
+    const evidence = result.measurements[0]?.candidate
+    if (!evidence) throw new Error('expected candidate evidence')
+    const benchmarkCell = {
+      suiteDigest: rig.experiment.benchmark.suite.digest,
+      taskIndex: 0,
+      repetition: 0,
     }
-    const codeComparison = createAgentImprovementMeasuredComparison({
-      result: codeResult,
-      measuredSurface: 'code',
-      baselineProfile: profile,
-      candidateBundle: codeBundle,
-    })
-    const codeProposal = createAgentImprovementProposal({
-      runId: 'analysis-run-code',
-      baselineProfile: profile,
-      findings: proposed.analysis.analystResult.findings,
-      evaluation: codeComparison,
-      candidateBundle: codeBundle,
-      now: () => new Date('2026-07-10T01:30:00.000Z'),
-    })
-    expect(codeProposal.changedSurfaces).toEqual(['code'])
     expect(
-      reviewAgentImprovementProposal(codeProposal, {
-        decision: 'approve',
-        reviewedBy: 'operator@example.com',
-        reason: 'The measured code winner matches the sealed patch.',
-      }).decision,
-    ).toBe('approve')
-    expect(() =>
-      createAgentImprovementMeasuredComparison({
-        result: {
-          ...codeResult,
-          winner: {
-            ...codeResult.winner,
-            surface: { ...codeSurface, candidateTree: '0'.repeat(candidateTree.length) },
-          },
-        },
-        measuredSurface: 'code',
-        baselineProfile: profile,
-        candidateBundle: codeBundle,
+      verifyCandidateExecutionEvidence(evidence, {
+        experiment: rig.experiment,
+        arm: 'candidate',
+        benchmarkCell,
+        seed: 101,
       }),
-    ).toThrow(/does not match the measured code winner/)
+    ).toEqual(evidence)
     expect(() =>
-      createAgentImprovementMeasuredComparison({
-        result: {
-          ...proposed.improvement.raw,
-          winner: { ...proposed.improvement.raw.winner, surface: 'measured memory' },
-        },
-        measuredSurface: 'memory',
-        baselineProfile: profile,
-        candidateBundle: proposed.proposal.candidateBundle,
-      }),
-    ).toThrow(/memory improvement proposals require a content-addressed bundle binding/)
-    expect(() =>
-      createAgentImprovementMeasuredComparison({
-        result: {
-          ...proposed.improvement.raw,
-          winner: {
-            ...proposed.improvement.raw.winner,
-            surface: '# Measured skill document',
-          },
-        },
-        measuredSurface: 'skills',
-        baselineProfile: profile,
-        candidateBundle: proposed.proposal.candidateBundle,
+      verifyCandidateExecutionEvidence(evidence, {
+        experiment: rig.experiment,
+        arm: 'baseline',
+        benchmarkCell,
+        seed: 101,
       }),
-    ).toThrow(/skill-document improvement proposals require a content-addressed bundle binding/)
+    ).toThrow(/substituted|bind/)
 
-    const review = reviewAgentImprovementProposal(proposed.proposal, {
+    const proposal = createAgentImprovementProposal({
+      runId: 'binding-run-1',
+      findings: [],
+      evaluation: result.evaluation,
+    })
+    const review = reviewAgentImprovementProposal(proposal, {
       decision: 'approve',
       reviewedBy: 'operator@example.com',
-      reason: 'Measured winner passed the held-back scenarios.',
-      now: () => new Date('2026-07-10T02:00:00.000Z'),
+      reason: 'Exact measured candidate approved.',
     })
-    expect(verifyAgentImprovementReview(review)).toEqual(review)
     expect(() =>
-      verifyAgentImprovementReview({ ...review, reason: 'Tampered after review.' }),
-    ).toThrow(/review digest does not match/)
-
-    const traceStore = new InMemoryTraceStore()
-    let request: AgentCandidateExecutorRequest | undefined
-    const executor: AgentCandidateExecutorPort = {
-      execute: async (input) => {
-        request = input
-        await traceStore.appendRun({
-          runId: input.trace.runId,
-          scenarioId: 'approved-candidate',
-          startedAt: 100,
-          endedAt: 200,
-          status: 'completed',
-          tags: { ...input.trace.tags },
-        })
-        return { executionId: input.executionId, termination: { kind: 'exit', exitCode: 0 } }
-      },
-      stop: async () => ({ stopped: true }),
-      capture: async () => ({
-        taskOutcome: unchangedTaskOutcomeCapture(fixture),
+      createAgentImprovementActivation(proposal, review, {
+        targets: [
+          {
+            surface: 'prompt',
+            identity: 'tenant/default/profile',
+            expectedBaseDigest: canonicalCandidateDigest('stale-profile'),
+          },
+        ],
+        fundingOwner: 'tenant/default',
+        authorizedBy: 'operator@example.com',
       }),
-    }
-    const outputs = createCandidateOutputFixture()
-    const executed = await executeApprovedAgentCandidate({
-      proposal: proposed.proposal,
-      review,
-      authorizeReview: async (candidateReview) => candidateReview.digest === review.digest,
-      task: fixture.task,
-      ports: fixture.ports,
-      execution: {
-        executor,
-        traceStore,
-        claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
-        ...outputs,
-      },
-    })
+    ).toThrow(/activation targets/)
+  })
 
-    expect(request).toBeDefined()
-    expect(executed.finalization.succeeded).toBe(true)
-    if (!executed.finalization.succeeded) throw new Error('expected successful finalization')
-    expect(executed.evidence).toMatchObject({
-      proposalDigest: proposed.proposal.digest,
-      reviewDigest: review.digest,
-      executionId: fixture.task.executionId,
-      succeeded: true,
-      receipt: {
-        bundleDigest: proposed.proposal.candidateBundle?.digest,
-        executionPlanDigest: expect.stringMatching(/^sha256:/),
-        materializationReceiptDigest: expect.stringMatching(/^sha256:/),
-        digest: expect.stringMatching(/^sha256:/),
-      },
-      materializationReceipt: {
-        digest: executed.finalization.receipt.value.materializationReceiptDigest,
-        profilePlan: { material: { files: expect.any(Array) } },
-      },
+  it('does not create a proposal from an inconclusive comparison', async () => {
+    const rig = createCandidateExperimentFixture({ scoreFor: () => 1 })
+    const result = await runAgentCandidateExperiment({
+      experiment: rig.experiment,
+      runId: 'no-lift-run',
+      placeCell: rig.placeCell,
     })
-    expect(
-      verifyCandidateExecutionEvidence([executed.evidence], {
-        proposal: proposed.proposal,
-        review,
-        expectedCount: 1,
-      }),
-    ).toEqual([executed.evidence])
-    expect(() =>
-      verifyCandidateExecutionEvidence([executed.evidence, executed.evidence], {
-        proposal: proposed.proposal,
-        review,
-        expectedCount: 2,
-      }),
-    ).toThrow(/reuses an execution id/)
-    expect(() =>
-      verifyCandidateExecutionEvidence([{ ...executed.evidence, succeeded: false }], {
-        proposal: proposed.proposal,
-        review,
-        expectedCount: 1,
-      }),
-    ).toThrow()
+
+    expect(result.evaluation.decision.outcome).not.toBe('ship')
     expect(() =>
-      verifyCandidateExecutionEvidence([forgeProfileActivation(executed.evidence)], {
-        proposal: proposed.proposal,
-        review,
-        expectedCount: 1,
+      createAgentImprovementProposal({
+        runId: 'no-lift-run',
+        findings: [],
+        evaluation: result.evaluation,
       }),
-    ).toThrow(/profile activation file path, mode, and content must match the canonical plan/)
+    ).toThrow(/passing experiment/)
   })
 
-  it('records rejection and refuses to execute it', async () => {
-    const fixture = createCandidateExecutionFixture()
-    const { digest: _digest, ...bundleInput } = fixture.bundle
-    const proposed = await proposeAgentImprovement({
-      runId: 'analysis-run-2',
-      profile: fixtureProfile(),
-      analysis: { registry: registry([finding]), inputs: {}, findingsStore: null, log: () => {} },
-      improvement: {
-        surface: 'prompt',
-        generator: proposer,
-        scenarios,
-        judge,
-        agent,
-        budget: { generations: 1, populationSize: 2, reps: 3, holdoutFraction: 0.5 },
-      },
-      buildCandidate: ({ improvement }) => alignedBundle(bundleInput, improvement.profile),
+  it('does not activate a rejected proposal', async () => {
+    const rig = createCandidateExperimentFixture()
+    const result = await runAgentCandidateExperiment({
+      experiment: rig.experiment,
+      runId: 'rejected-run',
+      placeCell: rig.placeCell,
+    })
+    const proposal = createAgentImprovementProposal({
+      runId: 'rejected-run',
+      findings: [],
+      evaluation: result.evaluation,
     })
-    const review = reviewAgentImprovementProposal(proposed.proposal, {
+    const review = reviewAgentImprovementProposal(proposal, {
       decision: 'reject',
       reviewedBy: 'operator@example.com',
-      reason: 'The change is not appropriate for this deployment.',
-      feedback: 'Keep the baseline behavior.',
-    })
-
-    await expect(
-      executeApprovedAgentCandidate({
-        proposal: proposed.proposal,
-        review,
-        authorizeReview: async () => true,
-        task: fixture.task,
-        ports: fixture.ports,
-        execution: {
-          executor: {
-            execute: async () => {
-              throw new Error('must not execute')
-            },
-            stop: async () => ({ stopped: true }),
-            capture: async () => ({}),
-          },
-          traceStore: new InMemoryTraceStore(),
-          claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
-          ...createCandidateOutputFixture(),
-        },
-      }),
-    ).rejects.toThrow('not an approval')
-  })
-
-  it('rejects a sealed build candidate whose digest was tampered', async () => {
-    const fixture = createCandidateExecutionFixture()
-    const { digest: _digest, ...bundleInput } = fixture.bundle
-
-    await expect(
-      proposeAgentImprovement({
-        runId: 'analysis-run-tampered-bundle',
-        profile: fixtureProfile(),
-        analysis: { registry: registry([finding]), inputs: {}, findingsStore: null, log: () => {} },
-        improvement: {
-          surface: 'prompt',
-          generator: proposer,
-          scenarios,
-          judge,
-          agent,
-          budget: { generations: 1, populationSize: 2, reps: 3, holdoutFraction: 0.5 },
-        },
-        buildCandidate: ({ improvement }) => ({
-          ...alignedSealedBundle(bundleInput, improvement.profile),
-          digest: candidateSha('0'),
-        }),
-      }),
-    ).rejects.toThrow('built candidate bundle digest is invalid')
-  })
-
-  it('refuses a structurally valid approval that its authority does not recognize', async () => {
-    const fixture = createCandidateExecutionFixture()
-    const { digest: _digest, ...bundleInput } = fixture.bundle
-    const proposed = await proposeAgentImprovement({
-      runId: 'analysis-run-unauthorized',
-      profile: fixtureProfile(),
-      analysis: { registry: registry([finding]), inputs: {}, findingsStore: null, log: () => {} },
-      improvement: {
-        surface: 'prompt',
-        generator: proposer,
-        scenarios,
-        judge,
-        agent,
-        budget: { generations: 1, populationSize: 2, reps: 3, holdoutFraction: 0.5 },
-      },
-      buildCandidate: ({ improvement }) => alignedBundle(bundleInput, improvement.profile),
-    })
-    const review = reviewAgentImprovementProposal(proposed.proposal, {
-      decision: 'approve',
-      reviewedBy: 'forged@example.com',
-      reason: 'Self-authored approval.',
+      reason: 'Not suitable for this deployment.',
     })
 
-    await expect(
-      executeApprovedAgentCandidate({
-        proposal: proposed.proposal,
-        review,
-        authorizeReview: async () => false,
-        task: fixture.task,
-        ports: fixture.ports,
-        execution: {
-          executor: {
-            execute: async () => {
-              throw new Error('must not execute')
-            },
-            stop: async () => ({ stopped: true }),
-            capture: async () => ({}),
+    expect(() =>
+      createAgentImprovementActivation(proposal, review, {
+        targets: [
+          {
+            surface: 'prompt',
+            identity: 'tenant/default/profile',
+            expectedBaseDigest: promptSurfaceDigest(rig),
           },
-          traceStore: new InMemoryTraceStore(),
-          claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
-          ...createCandidateOutputFixture(),
-        },
+        ],
+        fundingOwner: 'tenant/default',
+        authorizedBy: 'operator@example.com',
       }),
-    ).rejects.toThrow('not authorized')
+    ).toThrow(/requires an approval/)
   })
 
-  it('rejects judge-derived findings before they can steer a proposal', async () => {
+  it('blocks optimizer write-back before measurement and approval', async () => {
     await expect(
       proposeAgentImprovement({
-        runId: 'analysis-run-3',
-        profile: fixtureProfile(),
-        analysis: {
-          registry: registry([{ ...finding, derived_from_judge: true }]),
-          inputs: {},
-          findingsStore: null,
-          log: () => {},
+        runId: 'unsafe-memory-write',
+        profile: { name: 'fixture' },
+        analysis: {} as never,
+        improvement: { memory: { writeBack: async () => undefined } } as never,
+        buildExperiment: async () => {
+          throw new Error('must not build an experiment')
         },
-        improvement: {
-          surface: 'prompt',
-          generator: proposer,
-          scenarios,
-          judge,
-          agent,
-          budget: { generations: 1, populationSize: 2, reps: 3, holdoutFraction: 0.5 },
+        placeCell: () => {
+          throw new Error('must not place a cell')
         },
       }),
-    ).rejects.toThrow(/judge/i)
+    ).rejects.toThrow(/cannot write memory before approval/)
   })
 })
 
-function forgeProfileActivation(evidence: CandidateExecutionEvidence): CandidateExecutionEvidence {
-  const { digest: _activationDigest, ...activationWithoutDigest } = evidence.profileActivation
-  const first = evidence.profileActivation.files[0]
-  if (!first) throw new Error('expected a materialized profile file')
-  const profileActivation = canonicalCandidateDocument<
-    CandidateExecutionEvidence['profileActivation']
-  >({
-    ...activationWithoutDigest,
-    files: [
-      { ...first, content: `${first.content}\nforged` },
-      ...evidence.profileActivation.files.slice(1),
-    ],
-  }).value
-  const { digest: _evidenceDigest, ...evidenceWithoutDigest } = evidence
-  return canonicalCandidateDocument<CandidateExecutionEvidence>({
-    ...evidenceWithoutDigest,
-    profileActivation,
-  }).value
+function promptSurfaceDigest(rig: CandidateExperimentFixture): `sha256:${string}` {
+  const profile = agentCandidateProfileAsAgentProfile(rig.experiment.baseline.profile)
+  return canonicalCandidateDigest({
+    prompt: profile.prompt ?? null,
+    instructions: profile.resources?.instructions ?? null,
+  })
 }
diff --git a/tests/knowledge-improvement-job.test.ts b/tests/knowledge-improvement-job.test.ts
index a6f97805..232aa829 100644
--- a/tests/knowledge-improvement-job.test.ts
+++ b/tests/knowledge-improvement-job.test.ts
@@ -1,12 +1,7 @@
 import { mkdir, mkdtemp, readdir, readFile, rm, stat, writeFile } from 'node:fs/promises'
 import { tmpdir } from 'node:os'
 import { dirname, join, relative } from 'node:path'
-import type {
-  AgentCandidateBundle,
-  AgentCandidateKnowledge,
-  AgentImprovementMeasuredComparison,
-  AgentProfile,
-} from '@tangle-network/agent-interface'
+import type { AgentImprovementActivation } from '@tangle-network/agent-interface'
 import {
   addSourceText,
   defineReadinessSpec,
@@ -15,12 +10,13 @@ import {
   knowledgeImprovementCandidateRef,
   withKnowledgeImprovementCandidate,
 } from '@tangle-network/agent-knowledge'
-import { describe, expect, it } from 'vitest'
-import { canonicalCandidateDigest } from '../src/candidate-execution/digest'
-import { agentCandidateProfileAsAgentProfile } from '../src/candidate-execution/profile'
+import { afterEach, describe, expect, it } from 'vitest'
+import { createAgentCandidateWorkspacePort } from '../src/candidate-execution/workspace-archive'
 import {
+  createAgentImprovementActivation,
   createAgentImprovementProposal,
   reviewAgentImprovementProposal,
+  runAgentCandidateExperiment,
 } from '../src/intelligence/improvement-cycle'
 import { runKnowledgeImprovementJob } from '../src/knowledge'
 import type { SuperviseOptions } from '../src/runtime/supervise/supervise'
@@ -28,10 +24,19 @@ import type { SupervisorProfile } from '../src/runtime/supervise/supervisor-agen
 import type { SupervisedResult } from '../src/runtime/supervise/types'
 import {
   candidateBundle,
-  candidateSha,
+  cleanupCandidateFixtures,
   createCandidateOutputFixture,
   redigestCandidateBundle,
 } from './helpers/candidate-execution-fixture'
+import {
+  cleanupCandidateExperimentFixtures,
+  createCandidateExperimentFixture,
+} from './helpers/candidate-experiment-fixture'
+
+afterEach(() => {
+  cleanupCandidateExperimentFixtures()
+  cleanupCandidateFixtures()
+})
 
 async function withKb(fn: (root: string) => Promise<void>): Promise<void> {
   const root = await mkdtemp(join(tmpdir(), 'agent-runtime-knowledge-job-'))
@@ -129,91 +134,6 @@ function isMissingFile(error: unknown): boolean {
   return error instanceof Error && 'code' in error && error.code === 'ENOENT'
 }
 
-function measuredKnowledgeComparison(
-  baselineProfile: AgentProfile,
-  bundle: AgentCandidateBundle,
-): AgentImprovementMeasuredComparison {
-  const confidenceInterval = {
-    level: 0.95,
-    lower: 0,
-    upper: 0,
-    method: 'paired-bootstrap' as const,
-    statistic: 'mean' as const,
-    resamples: 2_000,
-  }
-  return {
-    kind: 'agent-improvement-measured-comparison',
-    benchmark: { name: 'development', version: '1', splitDigest: candidateSha('f') },
-    baselineProfileDigest: canonicalCandidateDigest(baselineProfile),
-    candidateBundleDigest: bundle.digest,
-    overall: {
-      name: 'composite',
-      baseline: 0,
-      candidate: 1,
-      delta: 1,
-      confidenceInterval: { ...confidenceInterval, lower: 1, upper: 1 },
-      n: 1,
-      direction: 'higher-is-better',
-      unit: 'score',
-    },
-    objectives: [
-      {
-        availability: 'measured',
-        kind: 'objective',
-        name: 'knowledge-readiness',
-        baseline: 0,
-        candidate: 1,
-        delta: 1,
-        confidenceInterval: { ...confidenceInterval, lower: 1, upper: 1 },
-        n: 1,
-        direction: 'higher-is-better',
-        unit: 'score',
-      },
-      {
-        availability: 'unavailable',
-        kind: 'cost',
-        name: 'cost',
-        direction: 'lower-is-better',
-        unit: 'usd',
-        reason: 'This deterministic fixture does not measure cost.',
-      },
-      {
-        availability: 'unavailable',
-        kind: 'latency',
-        name: 'latency',
-        direction: 'lower-is-better',
-        unit: 'milliseconds',
-        reason: 'This deterministic fixture does not measure latency.',
-      },
-    ],
-    candidate: { label: 'frozen knowledge candidate' },
-    decision: {
-      outcome: 'ship',
-      reasons: ['The paired held-out comparison approved this exact knowledge candidate.'],
-      contributingChecks: [{ name: 'heldout', passed: true }],
-    },
-    power: {
-      sufficient: true,
-      n: 1,
-      minimumDetectableDelta: 0,
-      confidenceLevel: 0.95,
-      scaleAssumed: true,
-      sharedScorerChannel: false,
-      reason: 'Fixture measurement is deterministic.',
-    },
-    provenance: {
-      kind: 'agent-eval-loop',
-      schema: '1.0.0',
-      runId: 'knowledge-heldout',
-      recordDigest: candidateSha('1'),
-      baselineContentHash: candidateSha('2'),
-      candidateContentHash: candidateSha('3'),
-    },
-    diff: 'knowledge candidate bytes changed',
-    evaluation: { generationsExplored: 1, durationMs: 1, totalCostUsd: 0 },
-  }
-}
-
 const KNOWLEDGE_IMPROVEMENT_JOB_TEST_TIMEOUT_MS = 15_000
 
 describe('runKnowledgeImprovementJob', () => {
@@ -395,46 +315,58 @@ describe('runKnowledgeImprovementJob', () => {
         )
         const liveBeforeApproval = await liveKnowledgeBytes(root)
         const baseBundle = candidateBundle()
-        const baselineProfile = agentCandidateProfileAsAgentProfile(baseBundle.profile)
         const bundle = redigestCandidateBundle(baseBundle, { knowledge })
+        const rig = createCandidateExperimentFixture({
+          baseline: baseBundle,
+          candidate: bundle,
+          configureFixture: (fixture) => {
+            fixture.ports.artifacts = artifacts
+            const materialize = fixture.ports.workspaces.materialize
+            const archivedWorkspaces = createAgentCandidateWorkspacePort()
+            fixture.ports.workspaces.materialize = async (input) =>
+              input.role === 'knowledge'
+                ? archivedWorkspaces.materialize(input)
+                : materialize(input)
+          },
+        })
+        const measured = await runAgentCandidateExperiment({
+          experiment: rig.experiment,
+          runId: 'runtime-job-approved',
+          placeCell: rig.placeCell,
+        })
         const proposal = createAgentImprovementProposal({
           runId: 'runtime-job-approved',
-          baselineProfile,
           findings: [],
-          evaluation: measuredKnowledgeComparison(baselineProfile, bundle),
-          candidateBundle: bundle,
+          evaluation: measured.evaluation,
           now: () => new Date('2026-07-13T01:00:00.000Z'),
         })
-        const mismatchedKnowledge: AgentCandidateKnowledge = {
-          ...knowledge,
-          candidate: { ...knowledge.candidate, candidateHash: candidateSha('0') },
-        }
-        const mismatchedBundle = redigestCandidateBundle(bundle, {
-          knowledge: mismatchedKnowledge,
-        })
-        expect(() =>
-          createAgentImprovementProposal({
-            runId: 'runtime-job-mismatched',
-            baselineProfile,
-            findings: [],
-            evaluation: measuredKnowledgeComparison(baselineProfile, bundle),
-            candidateBundle: mismatchedBundle,
-          }),
-        ).toThrow(/measured comparison does not bind the exact candidate bundle/)
         const review = reviewAgentImprovementProposal(proposal, {
           decision: 'approve',
           reviewedBy: 'operator@example.com',
           reason: 'Approve the exact frozen knowledge candidate.',
           now: () => new Date('2026-07-13T01:01:00.000Z'),
         })
+        const activation = createAgentImprovementActivation(proposal, review, {
+          targets: [
+            {
+              surface: 'knowledge',
+              identity: root,
+              expectedBaseDigest: knowledge.candidate.baseHash,
+            },
+          ],
+          fundingOwner: 'tenant/default',
+          authorizedBy: 'operator@example.com',
+          now: () => new Date('2026-07-13T01:02:00.000Z'),
+        })
 
         expect(await liveKnowledgeBytes(root)).toEqual(liveBeforeApproval)
         const approvedUpdate = async () => {
           throw new Error('candidate-ready promotion must not rerun the updater')
         }
         const runApproved = (
-          approvedProposal: AgentImprovementProposal,
-          approvedReview: AgentImprovementReview,
+          approvedProposal: typeof proposal,
+          approvedReview: typeof review,
+          approvedActivation: AgentImprovementActivation,
         ) =>
           runKnowledgeImprovementJob({
             root,
@@ -448,11 +380,12 @@ describe('runKnowledgeImprovementJob', () => {
             approval: {
               proposal: approvedProposal,
               review: approvedReview,
-              authorizeReview: async (candidateReview) =>
-                candidateReview.digest === approvedReview.digest,
+              activation: approvedActivation,
+              authorizeActivation: async (candidateActivation) =>
+                candidateActivation.digest === approvedActivation.digest,
             },
           })
-        const promoteApproved = () => runApproved(proposal, review)
+        const promoteApproved = () => runApproved(proposal, review, activation)
 
         await expect(
           withKnowledgeImprovementCandidate(
@@ -465,23 +398,6 @@ describe('runKnowledgeImprovementJob', () => {
         ).rejects.toThrow(/snapshot changed during use/)
         expect(await liveKnowledgeBytes(root)).toEqual(liveBeforeApproval)
 
-        const mismatchedProposal = createAgentImprovementProposal({
-          runId: 'runtime-job-mismatched',
-          baselineProfile,
-          findings: [],
-          evaluation: measuredKnowledgeComparison(baselineProfile, mismatchedBundle),
-          candidateBundle: mismatchedBundle,
-        })
-        const mismatchedReview = reviewAgentImprovementProposal(mismatchedProposal, {
-          decision: 'approve',
-          reviewedBy: 'operator@example.com',
-          reason: 'Attempt to approve a candidate identity that was never measured.',
-        })
-        await expect(runApproved(mismatchedProposal, mismatchedReview)).rejects.toThrow(
-          /does not match the measured candidate/,
-        )
-        expect(await liveKnowledgeBytes(root)).toEqual(liveBeforeApproval)
-
         const promoted = await promoteApproved()
 
         expect(promoted.promoted).toBe(true)
diff --git a/tests/mcp/worktree-harness.test.ts b/tests/mcp/worktree-harness.test.ts
index df2337a6..5f31b7b7 100644
--- a/tests/mcp/worktree-harness.test.ts
+++ b/tests/mcp/worktree-harness.test.ts
@@ -1,6 +1,7 @@
 import { execFileSync, spawnSync } from 'node:child_process'
 import { createHash } from 'node:crypto'
 import {
+  chmodSync,
   existsSync,
   mkdirSync,
   mkdtempSync,
@@ -77,20 +78,17 @@ describe('runWorktreeHarness profile materialization', () => {
     }
   })
 
-  it('delivers mounted inputs, captures the worker edit, and excludes tracked and untracked inputs', async () => {
-    const repoRoot = initializeRepository({
-      'src/value.ts': 'export const value = 1\n',
-      'profile-tracked.txt': 'repository version\n',
-    })
+  it('delivers mounted inputs, captures the worker edit, and excludes profile inputs', async () => {
     const runId = 'profile-real-path'
     const newInputPath = '.agent-profile/[literal]*:context.txt'
-    const trackedInputPath = 'profile-tracked.txt'
     const newInputMarker = 'PROFILE_NEW_INPUT_8f08e9f8'
-    const trackedInputMarker = 'PROFILE_TRACKED_INPUT_c521cd4d'
     const systemMarker = 'SYSTEM_PROMPT_2ef237df'
     const promptInstructionMarker = 'PROMPT_INSTRUCTION_6dcc8c4e'
     const resourceInstructionMarker = 'RESOURCE_INSTRUCTION_75b98d68'
     const taskMarker = 'TASK_PROMPT_d5ade1ac'
+    const repoRoot = initializeRepository({
+      'src/value.ts': 'export const value = 1\n',
+    })
     const profile: AgentProfile = {
       model: { default: 'gpt-5.4', reasoningEffort: 'xhigh' },
       prompt: {
@@ -103,14 +101,6 @@ describe('runWorktreeHarness profile materialization', () => {
             path: newInputPath,
             resource: { kind: 'inline', name: 'new-context', content: newInputMarker },
           },
-          {
-            path: trackedInputPath,
-            resource: {
-              kind: 'inline',
-              name: 'tracked-context',
-              content: trackedInputMarker,
-            },
-          },
         ],
         instructions: {
           kind: 'inline',
@@ -131,7 +121,6 @@ describe('runWorktreeHarness profile materialization', () => {
         codexReproducible: true,
         runHarness: async (options: RunLocalHarnessOptions) => {
           expect(readFileSync(join(options.cwd, newInputPath), 'utf8')).toBe(newInputMarker)
-          expect(readFileSync(join(options.cwd, trackedInputPath), 'utf8')).toBe(trackedInputMarker)
           const prompt = options.invocation?.args[1] ?? ''
           expect(options.codexReproducible).toBe(true)
           expect(options.invocation?.args).toContain('project_doc_max_bytes=0')
@@ -146,10 +135,6 @@ describe('runWorktreeHarness profile materialization', () => {
 
           writeFileSync(join(options.cwd, 'src/value.ts'), 'export const value = 2\n')
           writeFileSync(join(options.cwd, newInputPath), 'worker changed new profile input\n')
-          writeFileSync(
-            join(options.cwd, trackedInputPath),
-            'worker changed tracked profile input\n',
-          )
           return successfulHarnessResult()
         },
       })
@@ -159,13 +144,11 @@ describe('runWorktreeHarness profile materialization', () => {
         expect(execution.out.patch).toContain('diff --git a/src/value.ts b/src/value.ts')
         expect(execution.out.patch).toContain('+export const value = 2')
         expect(execution.out.patch).not.toContain(newInputPath)
-        expect(execution.out.patch).not.toContain(trackedInputPath)
         expect(execution.out.patch).not.toContain(newInputMarker)
-        expect(execution.out.patch).not.toContain(trackedInputMarker)
         expect(execution.out.stats).toEqual({ filesChanged: 1, insertions: 1, deletions: 1 })
         expect(execution.out.profileMaterialization).toMatchObject({
           workspacePlanDigest: expect.stringMatching(/^sha256:[a-f0-9]{64}$/u),
-          writtenPaths: [newInputPath, trackedInputPath],
+          writtenPaths: [newInputPath],
           unsupported: [],
           environmentNames: [],
           flags: [],
@@ -188,6 +171,51 @@ describe('runWorktreeHarness profile materialization', () => {
     }
   })
 
+  it('rejects profile inputs that shadow repository files', async () => {
+    const repoRoot = initializeRepository({
+      'profile-tracked.txt': 'repository version\n',
+      'src/value.ts': 'export const value = 1\n',
+    })
+    const runId = 'profile-existing-file'
+    const runHarness = vi.fn()
+    chmodSync(join(repoRoot, 'profile-tracked.txt'), 0o755)
+    git(repoRoot, ['add', 'profile-tracked.txt'])
+    git(repoRoot, ['commit', '-q', '-m', 'make profile fixture executable'])
+    const previousUmask = process.umask(0o022)
+    try {
+      await expect(
+        runWorktreeHarness({
+          repoRoot,
+          profile: {
+            resources: {
+              files: [
+                {
+                  path: 'profile-tracked.txt',
+                  executable: true,
+                  resource: {
+                    kind: 'inline',
+                    name: 'tracked-context',
+                    content: 'repository version\n',
+                  },
+                },
+              ],
+            },
+          },
+          harness: 'codex',
+          taskPrompt: 'task',
+          runId,
+          runHarness,
+        }),
+      ).rejects.toThrow(/Refusing to replace existing workspace file: profile-tracked\.txt/u)
+      expect(runHarness).not.toHaveBeenCalled()
+      expect(existsSync(join(repoRoot, '.agent-worktrees', runId))).toBe(false)
+      expect(git(repoRoot, ['branch', '--list', `delegate/${runId}`])).toBe('')
+    } finally {
+      process.umask(previousUmask)
+      rmSync(repoRoot, { recursive: true, force: true })
+    }
+  })
+
   it('materializes every Claude-supported structural axis and delivers instructions once', async () => {
     const repoRoot = initializeRepository({ 'src/value.ts': 'export const value = 1\n' })
     const runId = 'supported-structural-axes'
@@ -198,8 +226,8 @@ describe('runWorktreeHarness profile materialization', () => {
       nativeAgent: '.claude/agents/native-reviewer.md',
       command: '.claude/commands/audit.md',
       subagent: '.claude/agents/helper.md',
-      mcp: '.mcp.json',
-      settings: '.claude/settings.json',
+      mcp: '.tangle/claude-mcp.json',
+      settings: '.tangle/claude-settings.json',
     } as const
     try {
       const run = await runWorktreeHarness({
@@ -276,8 +304,16 @@ describe('runWorktreeHarness profile materialization', () => {
           const settings = JSON.parse(
             readFileSync(join(options.cwd, paths.settings), 'utf8'),
           ) as Record<string, unknown>
-          expect(settings).toHaveProperty('enabledMcpjsonServers')
           expect(settings).toHaveProperty('hooks')
+          expect(options.invocation?.args).toEqual(
+            expect.arrayContaining([
+              '--mcp-config',
+              paths.mcp,
+              '--strict-mcp-config',
+              '--settings',
+              paths.settings,
+            ]),
+          )
           const prompt = options.invocation?.args[1] ?? ''
           expect(count(prompt, resourceInstructionMarker)).toBe(1)
           expect(count(prompt, 'DIRECT_SYSTEM_a3563f03')).toBe(1)
diff --git a/tests/sandbox-approved-candidate.test.ts b/tests/sandbox-approved-candidate.test.ts
index 955aa2af..dba08b7e 100644
--- a/tests/sandbox-approved-candidate.test.ts
+++ b/tests/sandbox-approved-candidate.test.ts
@@ -2,18 +2,26 @@ import { chmodSync, mkdirSync, writeFileSync } from 'node:fs'
 import { dirname, join } from 'node:path'
 
 import { InMemoryTraceStore } from '@tangle-network/agent-eval'
+import {
+  type CandidateExperimentExecutionInput,
+  sealCandidateBenchmarkSuite,
+  sealCandidateExperiment,
+} from '@tangle-network/agent-eval/contract'
 import type {
   AgentCandidateBundle,
-  AgentImprovementMeasuredComparison,
-  AgentProfile,
+  AgentCandidateExperiment,
 } from '@tangle-network/agent-interface'
-import type { Process, ProcessStatus, SandboxInstance } from '@tangle-network/sandbox'
+import type {
+  CreateSandboxOptions,
+  Process,
+  ProcessStatus,
+  SandboxInstance,
+} from '@tangle-network/sandbox'
 import { afterEach, describe, expect, it, vi } from 'vitest'
 
 import { InMemoryAgentCandidateExecutionClaimStore } from '../src/candidate-execution/claim'
 import {
   canonicalCandidateBytes,
-  canonicalCandidateDigest,
   embeddedCandidateArtifact,
 } from '../src/candidate-execution/digest'
 import {
@@ -22,12 +30,8 @@ import {
 } from '../src/candidate-execution/knowledge'
 import type { AgentCandidateExecutorRequest } from '../src/candidate-execution/types'
 import {
-  createAgentImprovementProposal,
-  reviewAgentImprovementProposal,
-} from '../src/intelligence/improvement-cycle'
-import {
-  createSandboxApprovedCandidateExecutor,
-  sandboxApprovedCandidateExecutionSupport,
+  createSandboxCandidateExperimentExecutor,
+  sandboxCandidateExperimentExecutionSupport,
 } from '../src/intelligence/sandbox-approved-candidate'
 import {
   candidateSha,
@@ -35,13 +39,15 @@ import {
   createCandidateOutputExecutionFixture,
   createCandidateOutputFixture,
   redigestCandidateBundle,
+  replaceCandidateFixtureTask,
 } from './helpers/candidate-execution-fixture'
 
 afterEach(cleanupCandidateFixtures)
 
-describe('Sandbox approved candidate executor', () => {
-  it('runs one exact bounded-output task in a fresh strict-egress sandbox', async () => {
+describe('sandbox candidate experiment executor', () => {
+  it('runs one exact bounded-output cell with profile and knowledge bytes', async () => {
     const fixture = createCandidateOutputExecutionFixture('application/json', 1_024)
+    const baseline = fixture.bundle
     const knowledgeBytes = Buffer.from('# Exact runbook\nUse the verified endpoint.\n', 'utf8')
     const knowledgeMaterial = {
       kind: 'agent-candidate-workspace-manifest' as const,
@@ -55,15 +61,8 @@ describe('Sandbox approved candidate executor', () => {
       ],
     }
     const knowledgeManifest = embeddedCandidateArtifact(canonicalCandidateBytes(knowledgeMaterial))
-    const knowledgeSnapshot = {
-      kind: 'agent-candidate-workspace-snapshot' as const,
-      digest: knowledgeManifest.sha256,
-      material: knowledgeMaterial,
-      manifest: knowledgeManifest,
-      archive: embeddedCandidateArtifact(Buffer.from('knowledge-archive', 'utf8')),
-    }
     const retrievalConfigBytes = Buffer.from('{"topK":4}\n', 'utf8')
-    fixture.bundle = redigestCandidateBundle(fixture.bundle, {
+    const candidate = redigestCandidateBundle(baseline, {
       knowledge: {
         candidate: {
           kind: 'knowledge-improvement-candidate',
@@ -75,56 +74,28 @@ describe('Sandbox approved candidate executor', () => {
           evidenceHash: candidateSha('4'),
           promotionPlanHash: candidateSha('5'),
         },
-        snapshot: knowledgeSnapshot,
+        snapshot: {
+          kind: 'agent-candidate-workspace-snapshot',
+          digest: knowledgeManifest.sha256,
+          material: knowledgeMaterial,
+          manifest: knowledgeManifest,
+          archive: embeddedCandidateArtifact(Buffer.from('knowledge-archive', 'utf8')),
+        },
         retrievalConfig: embeddedCandidateArtifact(retrievalConfigBytes),
         evaluation: embeddedCandidateArtifact(Buffer.from('{"score":1}\n', 'utf8')),
       },
     })
+    const experiment = candidateExperiment(fixture, candidate)
     const materialize = fixture.ports.workspaces.materialize
     fixture.ports.workspaces.materialize = async (input) => {
-      if (input.role !== 'knowledge') return materialize(input)
+      if (input.role !== 'knowledge') return await materialize(input)
       const path = join(input.destination, 'runbook.md')
       mkdirSync(dirname(path), { recursive: true })
       writeFileSync(path, knowledgeBytes)
       chmodSync(path, 0o644)
     }
-    const baselineProfile = { name: 'baseline' }
-    const proposal = createAgentImprovementProposal({
-      runId: 'sandbox-proposal-1',
-      baselineProfile,
-      findings: [],
-      evaluation: measuredProfileResult(baselineProfile, fixture.bundle),
-      candidateBundle: fixture.bundle,
-      now: () => new Date('2026-07-13T12:00:00.000Z'),
-    })
-    const review = reviewAgentImprovementProposal(proposal, {
-      decision: 'approve',
-      reviewedBy: 'operator@example.com',
-      reason: 'Held-back comparison passed.',
-      now: () => new Date('2026-07-13T12:01:00.000Z'),
-    })
-    const status: ProcessStatus = {
-      pid: 42,
-      command: 'codex',
-      cwd: '/workspace/task',
-      running: false,
-      exitCode: 0,
-      startedAt: new Date('2026-07-13T12:02:00.000Z'),
-      exitedAt: new Date('2026-07-13T12:02:01.000Z'),
-    }
-    const process = {
-      wait: vi.fn(async () => 0),
-      status: vi.fn(async () => status),
-      stdout: async function* () {
-        yield '{"accepted":true}\n'
-      },
-      kill: vi.fn(async () => undefined),
-    } as unknown as Process
+    const process = completedProcess('{"accepted":true}\n')
     const writes: Array<{ path: string; content: string; mode: number }> = []
-    const spawnExact = vi.fn(async () => {
-      spawned = true
-      return process
-    })
     let spawned = false
     let deleted = false
     const sandbox = {
@@ -136,8 +107,11 @@ describe('Sandbox approved candidate executor', () => {
         }),
       },
       process: {
-        spawnExact,
-        list: vi.fn(async () => (spawned ? [status] : [])),
+        spawnExact: vi.fn(async () => {
+          spawned = true
+          return process
+        }),
+        list: vi.fn(async () => (spawned ? [stoppedStatus(0)] : [])),
         get: vi.fn(async () => process),
       },
       delete: vi.fn(async () => {
@@ -146,7 +120,7 @@ describe('Sandbox approved candidate executor', () => {
     } as unknown as SandboxInstance
     const create = vi.fn(async () => sandbox)
     const outputs = createCandidateOutputFixture()
-    const adapter = createSandboxApprovedCandidateExecutor({
+    const adapter = createSandboxCandidateExperimentExecutor({
       client: {
         create,
         get: vi.fn(async () => (deleted ? null : sandbox)),
@@ -156,10 +130,10 @@ describe('Sandbox approved candidate executor', () => {
       ...outputs,
       traceStore: new InMemoryTraceStore(),
       claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
-      authorizeReview: async (candidateReview) => candidateReview.digest === review.digest,
     })
 
-    const evidence = await adapter.execute({ proposal, review, task: fixture.task })
+    const input = candidateCell(experiment, fixture)
+    const evidence = await adapter.execute(input)
 
     expect(create).toHaveBeenCalledWith(
       expect.objectContaining({
@@ -174,7 +148,7 @@ describe('Sandbox approved candidate executor', () => {
         },
         metadata: expect.objectContaining({
           kind: 'agent-candidate-execution',
-          executionId: evidence.executionId,
+          executionId: input.executionId,
           executionPlanDigest: evidence.receipt.executionPlanDigest,
           outputMediaType: 'application/json',
           outputMaxBytes: 1_024,
@@ -182,14 +156,15 @@ describe('Sandbox approved candidate executor', () => {
       }),
       expect.objectContaining({ signal: expect.any(AbortSignal), timeoutMs: 120_000 }),
     )
+    const spawnExact = sandbox.process.spawnExact as ReturnType<typeof vi.fn>
     expect(spawnExact).toHaveBeenCalledOnce()
     const [executable, args, launch] = spawnExact.mock.calls[0]!
     expect(executable).toBe('codex')
     expect(args).toEqual(expect.any(Array))
     expect(launch).toMatchObject({
       cwd: '/workspace/task',
-      stdin: fixture.task.instruction,
-      timeoutMs: fixture.task.limits.timeoutMs,
+      stdin: fixture.task.task.instruction,
+      timeoutMs: fixture.task.task.limits.timeoutMs,
     })
     expect(launch.env).not.toHaveProperty('PATH')
     expect(launch.env).toHaveProperty('MODEL_GATEWAY_TOKEN', 'protected')
@@ -198,23 +173,14 @@ describe('Sandbox approved candidate executor', () => {
       [CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV]:
         '/workspace/task/.tangle/knowledge-retrieval-config.json',
     })
-    expect(writes.length).toBeGreaterThan(0)
     expect(deleted).toBe(true)
     expect(evidence).toMatchObject({
-      proposalDigest: proposal.digest,
-      reviewDigest: review.digest,
-      succeeded: true,
       materializationReceipt: {
-        profilePlan: { material: { files: expect.any(Array) } },
-      },
-      profileActivation: {
-        profilePlan: { digest: evidence.materializationReceipt.profilePlan.digest },
-        files: expect.any(Array),
+        profileActivation: { files: expect.any(Array) },
       },
       receipt: {
-        executorCapture: expect.objectContaining({
-          sha256: expect.stringMatching(/^sha256:/),
-        }),
+        bundleDigest: candidate.digest,
+        executorCapture: expect.objectContaining({ sha256: expect.stringMatching(/^sha256:/) }),
         taskOutcome: {
           material: {
             outcome: {
@@ -225,39 +191,28 @@ describe('Sandbox approved candidate executor', () => {
         },
       },
     })
-    for (const file of evidence.profileActivation.files) {
+    for (const file of evidence.materializationReceipt.profileActivation.files) {
       const written = writes.find((entry) => entry.path === `/workspace/task/${file.path}`)
       expect(written?.mode).toBe(file.mode)
       expect(Buffer.from(written?.content ?? '', 'base64').toString('utf8')).toBe(file.content)
     }
-    const writtenKnowledge = writes.find(
-      (entry) => entry.path === '/workspace/task/.tangle/knowledge/runbook.md',
-    )
-    expect(Buffer.from(writtenKnowledge?.content ?? '', 'base64')).toEqual(knowledgeBytes)
-    const writtenRetrievalConfig = writes.find(
-      (entry) => entry.path === '/workspace/task/.tangle/knowledge-retrieval-config.json',
-    )
-    expect(Buffer.from(writtenRetrievalConfig?.content ?? '', 'base64')).toEqual(
-      retrievalConfigBytes,
-    )
-    await expect(
-      adapter.executor.stop(
-        {
-          executionId: evidence.executionId,
-          executionPlanDigest: evidence.receipt.executionPlanDigest,
-        },
-        {
-          traceStore: new InMemoryTraceStore(),
-          reason: 'completed',
-          signal: new AbortController().signal,
-          deadlineAtMs: Date.now() + 1_000,
-        },
+    expect(
+      Buffer.from(
+        writes.find((entry) => entry.path.endsWith('/knowledge/runbook.md'))?.content ?? '',
+        'base64',
+      ),
+    ).toEqual(knowledgeBytes)
+    expect(
+      Buffer.from(
+        writes.find((entry) => entry.path.endsWith('knowledge-retrieval-config.json'))?.content ??
+          '',
+        'base64',
       ),
-    ).resolves.toEqual({ stopped: true })
+    ).toEqual(retrievalConfigBytes)
   })
 
-  it('declares and enforces its bounded capability instead of claiming universality', async () => {
-    expect(sandboxApprovedCandidateExecutionSupport).toEqual({
+  it('declares and enforces its bounded capability', async () => {
+    expect(sandboxCandidateExperimentExecutionSupport).toEqual({
       outcomes: ['output'],
       outputMediaTypes: ['text/*', 'application/json', '*+json'],
       code: ['disabled'],
@@ -277,7 +232,7 @@ describe('Sandbox approved candidate executor', () => {
         egress: ['blocked', 'strict'],
       },
     })
-    const adapter = createSandboxApprovedCandidateExecutor({
+    const adapter = createSandboxCandidateExperimentExecutor({
       client: {
         create: vi.fn(async () => {
           throw new Error('unsupported requests must fail before sandbox creation')
@@ -290,7 +245,6 @@ describe('Sandbox approved candidate executor', () => {
       outputArtifacts: {} as never,
       traceStore: new InMemoryTraceStore(),
       claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
-      authorizeReview: async () => true,
     })
     const request = minimalExecutorRequest()
 
@@ -314,8 +268,74 @@ describe('Sandbox approved candidate executor', () => {
     ).rejects.toThrow(/only UTF-8 text and JSON/)
   })
 
+  it('blocks all sandbox egress when the signed task disables model calls', async () => {
+    const fixture = createCandidateOutputExecutionFixture('application/json', 32)
+    replaceCandidateFixtureTask(fixture, {
+      limits: { ...fixture.task.task.limits, maxModelCalls: 0 },
+    })
+    const experiment = candidateExperiment(fixture)
+    const process = completedProcess('{"accepted":true}\n')
+    const { adapter, create } = sandboxAdapter(fixture, process, () => stoppedStatus(0))
+
+    await adapter.execute(candidateCell(experiment, fixture))
+
+    expect(create).toHaveBeenCalledWith(
+      expect.objectContaining({ egressPolicy: { mode: 'blocked' } }),
+      expect.any(Object),
+    )
+  })
+
+  it('recovers bounded output from sandbox metadata in a fresh worker', async () => {
+    const executionPlanDigest = candidateSha('a')
+    const process = completedProcess('{"recovered":true}\n')
+    let deleted = false
+    const sandbox = {
+      id: 'sandbox-recovery',
+      metadata: {
+        kind: 'agent-candidate-execution',
+        executionId: 'recovery-execution',
+        executionPlanDigest,
+        outputMaxBytes: 64,
+      },
+      process: {
+        list: vi.fn(async () => [stoppedStatus(0)]),
+        get: vi.fn(async () => process),
+      },
+      delete: vi.fn(async () => {
+        deleted = true
+      }),
+    } as unknown as SandboxInstance
+    const adapter = createSandboxCandidateExperimentExecutor({
+      client: {
+        create: vi.fn(async () => sandbox),
+        get: vi.fn(async () => (deleted ? null : sandbox)),
+        list: vi.fn(async () => (deleted ? [] : [sandbox])),
+      },
+      ports: {} as never,
+      grader: {} as never,
+      outputArtifacts: {} as never,
+      traceStore: new InMemoryTraceStore(),
+      claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
+      resultTimeoutMs: 100,
+    })
+    const request = { executionId: 'recovery-execution', executionPlanDigest }
+    const signal = new AbortController().signal
+
+    const capture = await adapter.executor.capture(request, {
+      traceStore: new InMemoryTraceStore(),
+      signal,
+    })
+    await adapter.executor.dispose?.(request, { signal })
+
+    expect(Buffer.from(capture.taskOutcome?.bytes ?? []).toString('utf8')).toBe(
+      '{"recovered":true}\n',
+    )
+    expect(deleted).toBe(true)
+  })
+
   it('kills oversized output and deletes the failed sandbox', async () => {
-    const { fixture, proposal, review } = approvedExecution(4)
+    const fixture = createCandidateOutputExecutionFixture('application/json', 4)
+    const experiment = candidateExperiment(fixture)
     let running = true
     let resolveWait!: (exitCode: number) => void
     const wait = new Promise<number>((resolve) => {
@@ -334,18 +354,21 @@ describe('Sandbox approved candidate executor', () => {
       },
       kill,
     } as unknown as Process
-    const { adapter, deleted } = sandboxAdapter(fixture, proposal, review, process, status)
+    const { adapter, deleted } = sandboxAdapter(fixture, process, status)
 
-    await expect(adapter.execute({ proposal, review, task: fixture.task })).rejects.toThrow(
+    await expect(adapter.execute(candidateCell(experiment, fixture))).rejects.toThrow(
       /output exceeds/,
     )
     expect(kill).toHaveBeenCalledWith('SIGKILL', { tree: true })
     expect(deleted()).toBe(true)
   })
 
-  it('kills and deletes a sandbox whose process exceeds the task timeout', async () => {
-    const { fixture, proposal, review } = approvedExecution(32)
-    fixture.task.limits = { ...fixture.task.limits, timeoutMs: 25 }
+  it('records timeout evidence and deletes the sandbox', async () => {
+    const fixture = createCandidateOutputExecutionFixture('application/json', 32)
+    replaceCandidateFixtureTask(fixture, {
+      limits: { ...fixture.task.task.limits, timeoutMs: 100 },
+    })
+    const experiment = candidateExperiment(fixture)
     let running = true
     let resolveWait!: (exitCode: number) => void
     const wait = new Promise<number>((resolve) => {
@@ -359,35 +382,101 @@ describe('Sandbox approved candidate executor', () => {
     const process = {
       wait: vi.fn(async () => wait),
       status: vi.fn(async () => status()),
-      stdout: async function* () {},
+      stdout: async function* () {
+        yield 'partial'
+      },
       kill,
     } as unknown as Process
-    const { adapter, deleted } = sandboxAdapter(fixture, proposal, review, process, status)
+    const { adapter, deleted } = sandboxAdapter(fixture, process, status)
 
-    await expect(adapter.execute({ proposal, review, task: fixture.task })).rejects.toThrow(
-      /approved candidate execution failed/,
-    )
+    const evidence = await adapter.execute(candidateCell(experiment, fixture))
+    expect(evidence.receipt.termination).toEqual({ kind: 'timeout', timeoutMs: 100 })
     expect(kill).toHaveBeenCalledWith('SIGKILL', { tree: true })
     expect(deleted()).toBe(true)
   })
-})
 
-function approvedExecution(maxBytes: number) {
-  const fixture = createCandidateOutputExecutionFixture('application/json', maxBytes)
-  const baselineProfile = { name: 'baseline' }
-  const proposal = createAgentImprovementProposal({
-    runId: `sandbox-proposal-${maxBytes}`,
-    baselineProfile,
-    findings: [],
-    evaluation: measuredProfileResult(baselineProfile, fixture.bundle),
-    candidateBundle: fixture.bundle,
+  it('fails before durable success when sandbox deletion is not proven', async () => {
+    const fixture = createCandidateOutputExecutionFixture('application/json', 32)
+    const experiment = candidateExperiment(fixture)
+    const process = completedProcess('{"accepted":true}\n')
+    const status = () => stoppedStatus(0)
+    const { adapter, sandbox } = sandboxAdapter(fixture, process, status, true)
+
+    await expect(adapter.execute(candidateCell(experiment, fixture))).rejects.toThrow(
+      /resource disposal|sandbox deletion failed/,
+    )
+    expect(sandbox.delete).toHaveBeenCalledOnce()
   })
-  const review = reviewAgentImprovementProposal(proposal, {
-    decision: 'approve',
-    reviewedBy: 'operator@example.com',
-    reason: 'Held-back comparison passed.',
+})
+
+function candidateExperiment(
+  fixture: ReturnType<typeof createCandidateOutputExecutionFixture>,
+  candidate: AgentCandidateBundle = redigestCandidateBundle(fixture.bundle, {
+    profile: {
+      ...fixture.bundle.profile,
+      prompt: { ...fixture.bundle.profile.prompt, systemPrompt: 'Candidate prompt.' },
+    },
+  }),
+): AgentCandidateExperiment {
+  return sealCandidateExperiment({
+    kind: 'agent-candidate-experiment',
+    digestAlgorithm: 'rfc8785-sha256',
+    baseline: fixture.bundle,
+    candidate,
+    candidateLineage: { source: 'human' },
+    benchmark: sealCandidateBenchmarkSuite({
+      tasks: [fixture.task.task],
+      reps: 1,
+      seeds: [101],
+    }),
+    policy: {
+      confidenceLevel: 0.95,
+      resamples: 500,
+      bootstrapSeed: 1_337,
+      deltaThreshold: 0,
+      minProductiveRuns: 3,
+      budgetUsd: 1,
+      criticalDimensions: [],
+      regressionTolerance: 0.05,
+    },
   })
-  return { fixture, proposal, review }
+}
+
+function candidateCell(
+  experiment: AgentCandidateExperiment,
+  fixture: ReturnType<typeof createCandidateOutputExecutionFixture>,
+) {
+  const input: CandidateExperimentExecutionInput & {
+    executionId: string
+    executionRoots: typeof fixture.task.executionRoots
+    stagingRoots: typeof fixture.task.stagingRoots
+  } = {
+    experiment,
+    arm: 'candidate',
+    bundle: experiment.candidate,
+    task: experiment.benchmark.tasks[0]!,
+    benchmarkCell: {
+      suiteDigest: experiment.benchmark.suite.digest,
+      taskIndex: 0,
+      repetition: 0,
+    },
+    seed: 101,
+    executionId: 'sandbox-execution-1',
+    executionRoots: fixture.task.executionRoots,
+    stagingRoots: fixture.task.stagingRoots,
+  }
+  return input
+}
+
+function completedProcess(output: string): Process {
+  return {
+    wait: vi.fn(async () => 0),
+    status: vi.fn(async () => stoppedStatus(0)),
+    stdout: async function* () {
+      yield output
+    },
+    kill: vi.fn(async () => undefined),
+  } as unknown as Process
 }
 
 function stoppedStatus(exitCode: number): ProcessStatus {
@@ -404,15 +493,14 @@ function stoppedStatus(exitCode: number): ProcessStatus {
 
 function sandboxAdapter(
   fixture: ReturnType<typeof createCandidateOutputExecutionFixture>,
-  proposal: ReturnType<typeof createAgentImprovementProposal>,
-  review: ReturnType<typeof reviewAgentImprovementProposal>,
   process: Process,
   currentStatus: () => ProcessStatus,
+  failDelete = false,
 ) {
   let spawned = false
   let wasDeleted = false
   const sandbox = {
-    id: `sandbox-${proposal.runId}`,
+    id: 'sandbox-candidate-test',
     metadata: undefined,
     fs: { write: vi.fn(async () => undefined) },
     process: {
@@ -424,13 +512,18 @@ function sandboxAdapter(
       get: vi.fn(async () => process),
     },
     delete: vi.fn(async () => {
+      if (failDelete) throw new Error('sandbox deletion failed')
       wasDeleted = true
     }),
   } as unknown as SandboxInstance
   const outputs = createCandidateOutputFixture()
-  const adapter = createSandboxApprovedCandidateExecutor({
+  const create = vi.fn(async (options: CreateSandboxOptions) => {
+    ;(sandbox as unknown as { metadata: Record<string, unknown> }).metadata = options.metadata ?? {}
+    return sandbox
+  })
+  const adapter = createSandboxCandidateExperimentExecutor({
     client: {
-      create: vi.fn(async () => sandbox),
+      create,
       get: vi.fn(async () => (wasDeleted ? null : sandbox)),
       list: vi.fn(async () => (wasDeleted ? [] : [sandbox])),
     },
@@ -438,112 +531,16 @@ function sandboxAdapter(
     ...outputs,
     traceStore: new InMemoryTraceStore(),
     claimStore: new InMemoryAgentCandidateExecutionClaimStore(),
-    authorizeReview: async (candidateReview) => candidateReview.digest === review.digest,
   })
-  return { adapter, deleted: () => wasDeleted }
-}
-
-function measuredProfileResult(
-  baselineProfile: AgentProfile,
-  bundle: AgentCandidateBundle,
-): AgentImprovementMeasuredComparison {
-  return {
-    kind: 'agent-improvement-measured-comparison',
-    benchmark: { name: 'development', version: '1', splitDigest: candidateSha('f') },
-    baselineProfileDigest: canonicalCandidateDigest(baselineProfile),
-    candidateBundleDigest: bundle.digest,
-    overall: {
-      name: 'composite',
-      baseline: 0,
-      candidate: 1,
-      delta: 1,
-      confidenceInterval: {
-        level: 0.95,
-        lower: 1,
-        upper: 1,
-        method: 'paired-bootstrap',
-        statistic: 'mean',
-        resamples: 2_000,
-      },
-      n: 3,
-      direction: 'higher-is-better',
-      unit: 'score',
-    },
-    objectives: [
-      {
-        kind: 'objective',
-        name: 'profile-quality',
-        availability: 'measured',
-        baseline: 0,
-        candidate: 1,
-        delta: 1,
-        confidenceInterval: {
-          level: 0.95,
-          lower: 1,
-          upper: 1,
-          method: 'paired-bootstrap',
-          statistic: 'mean',
-          resamples: 2_000,
-        },
-        n: 3,
-        direction: 'higher-is-better',
-        unit: 'score',
-      },
-      measuredResourceObjective('cost', 'usd'),
-      measuredResourceObjective('latency', 'milliseconds'),
-    ],
-    candidate: { label: 'measured profile' },
-    decision: {
-      outcome: 'ship',
-      reasons: ['paired heldout interval cleared the promotion threshold'],
-      contributingChecks: [{ name: 'heldout', passed: true }],
-    },
-    power: {
-      sufficient: true,
-      n: 3,
-      minimumDetectableDelta: 0.5,
-      confidenceLevel: 0.95,
-      scaleAssumed: true,
-      sharedScorerChannel: true,
-      reason: 'paired sample can detect the measured effect',
-    },
-    provenance: {
-      kind: 'agent-eval-loop',
-      schema: 'tangle.loop-provenance.v2',
-      runId: 'heldout-1',
-      recordDigest: candidateSha('1'),
-      baselineContentHash: candidateSha('2'),
-      candidateContentHash: candidateSha('3'),
-    },
-    diff: 'measured profile replacement',
-    evaluation: { generationsExplored: 1, durationMs: 1, totalCostUsd: 0 },
-  }
-}
-
-function measuredResourceObjective(
-  kind: 'cost' | 'latency',
-  unit: 'usd' | 'milliseconds',
-): AgentImprovementMeasuredComparison['objectives'][number] {
-  return {
-    kind,
-    name: kind,
-    availability: 'unavailable',
-    reason: `${kind} was not captured by this fixture`,
-    direction: 'lower-is-better' as const,
-    unit,
-  }
+  return { adapter, sandbox, create, deleted: () => wasDeleted }
 }
 
 function minimalExecutorRequest(): AgentCandidateExecutorRequest {
   return {
-    executionPlan: {
-      value: {
-        material: {
-          task: { outcome: { kind: 'output', mediaType: 'text/plain', maxBytes: 10 } },
-          codeKind: 'disabled',
-        },
-      },
+    benchmark: {
+      task: { outcome: { kind: 'output', mediaType: 'text/plain', maxBytes: 10 } },
     },
+    executionPlan: { value: { material: { codeKind: 'disabled' } } },
     inputs: {},
     memory: { mode: 'disabled' },
   } as unknown as AgentCandidateExecutorRequest
@@ -558,27 +555,27 @@ function withRequest(
     mediaType?: string
   },
 ): AgentCandidateExecutorRequest {
-  const material = request.executionPlan.value.material
+  const outcome = request.benchmark.task.outcome
   return {
     ...request,
+    benchmark: {
+      ...request.benchmark,
+      task: {
+        ...request.benchmark.task,
+        outcome:
+          change.taskOutcome ??
+          (outcome.kind === 'output'
+            ? { ...outcome, mediaType: change.mediaType ?? outcome.mediaType }
+            : outcome),
+      },
+    },
     executionPlan: {
       ...request.executionPlan,
       value: {
         ...request.executionPlan.value,
         material: {
-          ...material,
-          task: {
-            ...material.task,
-            outcome:
-              change.taskOutcome ??
-              (material.task.outcome.kind === 'output'
-                ? {
-                    ...material.task.outcome,
-                    mediaType: change.mediaType ?? material.task.outcome.mediaType,
-                  }
-                : material.task.outcome),
-          },
-          codeKind: change.codeKind ?? material.codeKind,
+          ...request.executionPlan.value.material,
+          codeKind: change.codeKind ?? request.executionPlan.value.material.codeKind,
         },
       },
     },