You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
artifact receipt gate in docs/steering-activation-path.md
Still not landed:
complete artifact receipt with exact repo, file path, resolved revision, local path, size, and SHA-256 digest for every model/tokenizer/SAE file
storage receipt after artifact resolution
policy admission / agent-registry grant records
GPT-2 Small model loading
blocks.6.hook_resid_pre SAE loading
real activation injection
real baseline vs steered completion
local smoke record with status: applied
Scope
This issue is GPT-2 Small only.
Gemma sourcesets are explicitly out of scope for closure because Gemma model access depends on operator acceptance of Google/Gemma terms. Gemma remains registered but blocked until terms/access and artifact readiness are verified separately.
Implement the path needed for:
agent-machine steer serve --sourceset gpt2-small.res-jb
POST /steer
The served endpoint must remain compatible with Noetica's existing local steering client shape:
Resolve the gpt2-small.res-jb sourceset from the SteeringSourceset registry.
Fetch or locate GPT-2 Small model artifacts from openai-community/gpt2 with digest verification.
Fetch or locate the residual-stream SAE artifact from jbloom/GPT2-Small-SAEs-Reformatted / SAELens release gpt2-small-res-jb / SAE id blocks.6.hook_resid_pre with digest verification.
Emit a storage receipt after successful artifact resolution.
Load GPT-2 Small.
Load the blocks.6.hook_resid_pre SAE.
Run a baseline completion.
Run a steered completion by injecting the requested SAE feature vector at the specified layer and strength.
Return a Noetica-compatible SteeringResult with status: applied, baseline, steered, diff_summary, feature_id, layer, and strength.
Emit or stage evidence sufficient for a local smoke record.
Acceptance criteria
agent-machine steer serve --sourceset gpt2-small.res-jb starts the real local steering endpoint when required artifacts are present and verified.
agent-machine steer serve-stub ... remains available and returns not_configured or noop only.
If artifacts, storage receipts, grants, policy admission, or runtime dependencies are missing, the real serve path fails closed with a clear diagnostic and does not silently downgrade to fake applied.
A local smoke record is committed showing:
sourceset: gpt2-small.res-jb
feature index: 10200
layer: 6-res-jb
strength: 5
prompt
baseline text
steered text
status: applied
evidence hash / receipt reference
Noetica can set NEURONPEDIA_BASE_URL=http://localhost:8080, call /api/steer, and receive the applied response without code changes.
Native validation / CI passes.
Non-goals
Do not implement Gemma activation in this issue.
Do not bypass policy/grant/storage/evidence gates.
Do not store credentials or tokens in the repository.
Goal
Implement a controlled local activation path for GPT-2 Small residual-stream SAE steering, using the already-registered
gpt2-small.res-jbsourceset.Current status
PR #38 landed the fail-closed preflight / real-path entrypoint tranche.
PR #39 landed the steering artifact receipt contract and pending fixture.
Issue #34 is not complete. It must remain open until a real
status: appliedlocal smoke record exists.Landed by PR #38:
requirements-steering.txtoptional runtime dependency listagent-machine steer preflight --sourceset gpt2-small.res-jbagent-machine steer serve --sourceset gpt2-small.res-jb --host 127.0.0.1 --port 8080not_configuredbehavior when dependencies/artifacts/receipts/admission are absentdocs/steering-activation-path.mdLanded by PR #39:
contracts/steering-artifact-receipt.schema.jsongpt2-small.res-jbartifact receipt fixtureSteeringArtifactReceiptdocs/steering-artifact-receipts.mddocs/steering-activation-path.mdStill not landed:
blocks.6.hook_resid_preSAE loadingstatus: appliedScope
This issue is GPT-2 Small only.
Gemma sourcesets are explicitly out of scope for closure because Gemma model access depends on operator acceptance of Google/Gemma terms. Gemma remains registered but blocked until terms/access and artifact readiness are verified separately.
Implement the path needed for:
The served endpoint must remain compatible with Noetica's existing local steering client shape:
Required implementation work
gpt2-small.res-jbsourceset from theSteeringSourcesetregistry.openai-community/gpt2with digest verification.jbloom/GPT2-Small-SAEs-Reformatted/ SAELens releasegpt2-small-res-jb/ SAE idblocks.6.hook_resid_prewith digest verification.blocks.6.hook_resid_preSAE.SteeringResultwithstatus: applied,baseline,steered,diff_summary,feature_id,layer, andstrength.Acceptance criteria
agent-machine steer serve --sourceset gpt2-small.res-jbstarts the real local steering endpoint when required artifacts are present and verified.agent-machine steer serve-stub ...remains available and returnsnot_configuredornooponly.applied.gpt2-small.res-jb102006-res-jb5status: appliedNEURONPEDIA_BASE_URL=http://localhost:8080, call/api/steer, and receive the applied response without code changes.Non-goals