Define cskg.vnext with FSMS memory and Identity-Prime boundary controls

[mdheller]

Purpose

Define cskg.vnext as a registry-backed graph profile for TriTRPC vNext, aligned with fail-closed boundary watchers, FSMS memory operators, Identity-Is-Prime linkage controls, and the Gödel-style proof/evidence substrate.

Source materials

• Common Sense Knowledge Graph specification.docx
• CSKG_into_TriTRPC_VNext_alignment_blueprint_2026-03-31.docx
• fsms_memory_pack.zip
• identity_is_prime_reference_v2_patched.zip
• Existing issue Align vNext with fail-closed boundary watchers and Minimal Universe worldlets #66: fail-closed boundary watchers / Minimal Universe worldlets

Core decision

Do not put raw KGTK strings, relation labels, source lists, or sentences directly onto hot frames.

Use TriTRPC vNext as the transport/control/admission plane:

• CTRL243 for profile/lane/evidence/fallback/routefmt;
• KIND243 for unary/stream/beacon/error;
• Handle243 for routes, schema/context manifests, policy handles, dictionary handles;
• Braid243 + State243 for transport-semantic posture;
• Beacon-A/B/C for capability, intent/reservation, and commit/receipt.

Keep canonical CSKG truth in payload records and manifests.

Native profile

Add draft profile:

cskg.vnext

Default:

• profile=Path-A
• lane=classical
• routefmt=handle
• evidence=sampled|exact|verified depending on admission stage
• no hot-path raw KGTK relation strings

Payload records

EdgeRecord {
  id,
  node1,
  relation,
  node2,
  node1_labels[],
  node2_labels[],
  relation_labels[],
  relation_dimensions[],
  sources[],
  sentences[]
}

AuxRecord {
  edge_id,
  key,
  value,
  media_type,
  content_hash?
}

QuerySpec {
  seed_ids,
  relation_filter,
  dimension_filter,
  source_filter,
  language_scope,
  limit,
  cursor
}

ContextManifest {
  profile_version,
  dataset_version,
  vocabulary_versions,
  policy_epoch,
  ordering_rule
}

Canonicalization rule: parse KGTK boundary syntax -> trim boundary whitespace -> unescape multi-value cells -> deduplicate -> sort arrays by raw UTF-8 bytes -> encode deterministically.

FSMS memory alignment

Map memory objects:

• WorkingMemoryState: request/session context; TTL-bounded.
• EpisodeBundle: append-only trace of query/retrieval/graph traversal/action/validation.
• SemanticMemoryRelease: immutable promoted CSKG/latent/topic/hypergraph release.
• ProceduralMemoryBundle: policies/playbooks/operators used to act.
• PromotionRule: promotes from working/episodic to semantic memory only when evidence, review, and policy gates pass.
• ForgettingPolicy: TTL, compaction, purge, archive, and legal-hold behavior.

Identity-Is-Prime alignment

Identity controls are not optional.

• Identity-bearing or quasi-identifier CSKG rows require DoNotLearn and DoNotLink by default.
• SameAs/entity-resolution edges are high-risk and must route through Identity-Prime policy/proof artifacts before promotion.
• Prime-topic identity basis vectors must be referenced by handles or private-vault refs, not hot-path labels.
• Cross-context merges require explicit capability, provenance, proof artifact, and receipt.

Gödelization alignment

Every admitted graph/memory object needs:

• canonical quoted representation;
• stable code/hash/handle;
• substitution/rewrite receipt when transformed;
• proof/evidence relation when admitted;
• promotion/admission receipt when moved into semantic memory.

CSKG edges are candidate claims until receipted. SameAs merges are proof obligations, not ordinary labels.

Watcher gates

Add cskg.vnext watcher sequence:

1. SchemaWatcher: validate CSKG required columns / EdgeRecord fields.
2. CanonicalizationWatcher: enforce array canonicalization and decode->encode stability.
3. VocabularyWatcher: relation and dimension values must be in registered vocabularies.
4. SourceWatcher: source token and provenance admissibility.
5. PrivacyWatcher: identity, quasi-identifier, sentence text, source leakage.
6. IdentityPrimeWatcher: SameAs / entity-merge / cross-context linkage controls.
7. MemoryWatcher: working/episodic/semantic/procedural write authorization.
8. LearningWatcher: block training/vectorization when labels are missing or sensitive.
9. ClaimWatcher: CSKG edge is a claim/evidence item, not automatically truth.
10. ReceiptWatcher: emit BoundaryReceipt before admission.

Worked fixture targets

Add positive and negative fixtures:

• PutEdge ConceptNet-style IsA edge succeeds.
• PutEdge VisualGenome sentence succeeds only as public/common-sense edge, not identity memory.
• PutEdge with missing relation rejects.
• PutEdge with unknown relation rejects or quarantines.
• PutEdge SameAs linking two human identifiers quarantines pending Identity-Prime proof.
• ScanEdges with stale context handle rejects.
• BulkPutEdges with malformed pipe-list normalization rejects.
• Promotion to SemanticMemoryRelease without PromotionRule receipt rejects.
• Vectorization/training on unclassified sentence text rejects.

Acceptance criteria

• Draft spec/drafts/cskg_vnext_profile.md.
• Draft spec/drafts/annex_x_boundary_admission_profile.md cross-linking cskg.vnext.
• Add EdgeRecord/AuxRecord/QuerySpec/ContextManifest schemas.
• Add BoundaryReceipt examples for CSKG ingest, query, and semantic-memory promotion.
• Add FSMS memory object mapping to vNext beacons.
• Add Identity-Prime high-risk edge handling notes.
• Add positive/negative golden fixtures.
• Ensure no hot-path raw CSKG strings are required when handles/context manifests suffice.

Non-goals

• Do not replace stable v1.
• Do not force CSKG into ternary-native Path-B payloads prematurely.
• Do not encode CSKG relation;dimension as Braid243; Braid243 describes operation posture, not graph semantics.
• Do not treat SameAs/entity resolution as ordinary graph edges for memory promotion.
• Do not permit fail-open memory promotion or learning from unclassified graph text.