[Epic] Concept Ontology & Semantic Gap Inference

Input from Claude Opus 4.6 (Antigravity):

❋ Note on Origin Session ID: The 496c0120-29aa-4b02-a804-d5a32ff6b5eb is an Antigravity conversation ID, not a Memory Core session ID (Memory Core was offline during this planning session). The planning artifacts (implementation plan, concept layer analysis, walkthrough) are persisted at ~/.gemini/antigravity/brain/496c0120-29aa-4b02-a804-d5a32ff6b5eb/. Future agents can retrieve full context from there.

Input from Claude Opus 4.7 (Claude Code):

❋ # Session Handoff — 2026-04-19 / Session 62d6f155-e57f-4279-9b59-36c9e4ecbc5e

A2A Contextual Bridge per AGENTS.md §14. Future sessions picking up the Concept Ontology epic should find everything they need to boot efficiently here.

What Landed This Session

Directly scoped to #10030

• #10033 (Concept Graph Unit Tests) → closed via PR #10078. Closed four named-AC coverage gaps in the existing spec (root-tier assertion, coverage-annotation placement, production-graph line-count observability, explicit tier-ordering weight test). 23 tests total in ConceptService.spec.mjs.
• #10035 (Graph Integration & GapInferenceEngine Refactor) → closed via PR #10084. Architectural Pillar shipped. Replaces regex + per-match gemma4-31b LLM Boolean verification with deterministic concept-graph edge traversal. New ConceptIngestor daemon service wires the JSONL ontology into the Native Edge Graph as DreamService Phase 0. Adds [EXAMPLE_GAP] signal + display section in sandman_handoff.md. 30 tests pass.

Meta-infrastructure (not #10030-scoped but session-generated)

• #10075 (memory-mining skill) → merged via PR #10076. Promoted the memory-first reflex from rule-in-AGENTS.md to Progressive Disclosure skill. Also corrected ask_knowledge_base vs query_documents official tool descriptions in openapi.yaml.

Current Epic Status

Completed sub-issues

• ✅ #10031 — JSONL Schema & Concept Seed Extraction
• ✅ #10032 — ConceptService & Source Code Wiring
• ✅ #10033 — Concept Graph Unit Tests (closed this session)
• ✅ #10035 — Graph Integration & GapInferenceEngine Refactor (closed this session)
• ✅ #10049 — Concept Schema Enrichment: Aliases, Analogous Edges

Open sub-issues (recommended pickup order)

Priority	Sub-Issue	Complexity	Why this order
1	#10085 — ConceptIngestor follow-ups (bundle)	Low	Polish from #10084's self-review: hoist concept-graph pass to cycle-scope, dedicated spec, observability, orphan warn consolidation. Same two files, natural single PR. Closes the #10035 loop cleanly.
2	#10036 — Memory Core Concept Discovery	Medium	Unblocked by #10035 now that concept graph is in SQLite. Next natural consumer of the concept-graph infrastructure.
3	#10034 — Concept Graph Visualization App	Medium	Neo.mjs app surfacing the concept ontology. Independent of #10036; can run in parallel. Good candidate for a Neo.mjs frontend-focused session.
4	#10037 — ChromaDB Concept Embedding & Hybrid Search	Medium	Powers the relevance-bounded queries proposed in #10080 (separate ticket — see below). Do after #10085 lands so the stable concept-node contract is locked in.
5	#10086 — Config-lift guideGapWeightThreshold	Tiny	Tiny scope (config.template.mjs + one import path). Any-session pickup.
6	#10087 — ORPHAN_CONCEPT distinct display channel	Small	Promote from logger.warn to a first-class sandman_handoff section. Non-blocking; value grows with ontology size.
7	#10050 — Concept Description Enrichment	Medium	⚠️ Blocked on human architect time per @tobiu's 2026-04-17 annotation — agent-generated descriptions produce wrong results without architectural decision history (e.g., Data Worker case study). Collaborative session required.

Open sub-issues NOT filed by this session but visible on the epic

• #10050 — blocked (see above)

Related Tickets Not Parented Under #10030 (Worth Tracking)

These were filed this session but are broader than the Concept Ontology scope:

• #10077 — manage_issue_labels MCP tool fails on PRs (bug — pre-existing)
• #10079 — Broadened Hypothesis gate for written claims (skill strengthening — covers 5 claim classes observed this session)
• #10080 — Relevance-bounded query APIs for ConceptService (will consume #10037's embeddings when ready)
• #10081 — ask_knowledge_base query telemetry → Agent FAQ → Gap signal
• #10082 — Strengthen tech-debt-radar for unbounded-list API detection (sub-issue of #10080)
• #10083 — Audit AGENTS.md §9 "Single Full-Read Rule" for frontier-context-era relaxation
• #10088 — Automate post-merge knowledge-base sync (solves recurring [KB_GAP] pattern)
• #10089 — Document SQLite FK constraints on Database.mjs#addEdge JSDoc

Durable Lessons Captured This Session

Three feedback memory files written to the Agent OS's persistent cross-session memory:

• feedback_ask_kb_dominates_query_documents.md — empirically verified ask_knowledge_base returns synthesized answer PLUS the same top-5 refs query_documents returns. Reserve the latter only for exhaustive enumeration beyond top-5.
• feedback_challenge_numerical_thresholds_in_acs.md — hardcoded numerical thresholds in AC items (≤ N lines, at most N ms) trigger the Hypothesis gate. Prefer observability + architectural-layer assertions over format-layer caps.
• feedback_verify_written_claims_against_precedent.md — five distinct sub-classes of "accept written claim without verifying precedent" observed + remediations for each. Covers numerical, architectural, tool-routing, self-inferred-policy, and wrong-grounding-assumption classes.

What to Know for the Next Session

1. Opus 4.7 has 1M context — earlier session-management rules calibrated for smaller models may be over-conservative (see #10083). Don't blindly apply "fresh session per Architectural Pillar" — semantic locality usually wins.
2. Gemma 4 31B is frontier-class (85.2% MMLU Pro, distilled from Gemini 3) — don't frame SLM design decisions as "SLM is weak." Frame them correctly: deterministic code over LLM inference is about curated data being the source of truth, not capability.
3. The concept graph is the authoritative source for "does concept X have a guide?" — via EXPLAINED_BY edges. Do not re-introduce regex-based or LLM-based guide matching. Gap detection is now deterministic.
4. Stub-node seeding is a requirement for any SQLite edge insertion — the FK constraint is enforced at storage layer. Use ConceptIngestor.ensureEdgeTargetsExist as the pattern.
5. Invoke the memory-mining skill BEFORE diagnostics or architectural claims — the skill exists for exactly the failure modes observed this session.

Origin Session ID

62d6f155-e57f-4279-9b59-36c9e4ecbc5e

Input from Claude Opus 4.7 (Claude Code):

❋ # Session Handoff — 2026-04-19 / Session 1db25bbe-f39d-4dcd-bb0e-bc125ce91326

A2A Contextual Bridge per AGENTS.md §14. Future sessions picking up the Concept Ontology epic should find everything they need here. Session signal: marathon, valuable — tobi's explicit framing. Depth-investment paid off; next sessions should expect similar collaboration cadence.

What Landed This Session

Directly scoped to #10030 (all 4 closed)

• #10085 (ConceptIngestor follow-ups) → closed via PR #10100. Hoisted inferConceptGraphGaps from per-session to cycle-scope; deleted the executeCapabilityGapInference wrapper; promoted inferTestGapsFromSession + inferConceptGraphGaps to public API with matching thin facades on DreamService. New ConceptIngestor.spec.mjs with 6 coverage targets. Bonus: closed a pre-existing cross-spec singleton-leak flake (symmetric afterEach discipline — see durable memory). Ship-polish commits added hoist-behavioral test + idempotency skip-path assertion after self-review challenge. 40/40 stable × 8 runs.

• #10087 (ORPHAN_CONCEPT channel) → closed via PR #10101. Promoted orphan-concept surfacing from ephemeral logger.warn in ConceptIngestor to the durable capabilityGap tag channel + new ### ⚠️ Orphaned Concepts section in sandman_handoff.md via GoldenPathSynthesizer. Shared GUIDE_GAP_WEIGHT_THRESHOLD gate with the other two concept-graph signals. Self-review pivot mid-PR: tobi caught an "A2A vs human consumer" framing miss in the PR body — corrected inline, durable lesson saved. 40/40 stable.

• #10086 (config-lift threshold) → closed via PR #10102. GUIDE_GAP_WEIGHT_THRESHOLD promoted from file-local const to aiConfig.data.guideGapWeightThreshold with env override (NEO_GUIDE_GAP_WEIGHT_THRESHOLD). Self-review caught the service-boundary placement miss (memory-core config vs SDK-layer config); tracked as deferred debt via filed ticket #10103. 41/41 stable.

• #10036 (Memory Core Concept Discovery) → closed via PR #10105. Deliverables 1, 2, 4 shipped. New ConceptDiscoveryService — LLM-driven extraction via OpenAiCompatible.generate (gemma-4-31b-it). Two corpora: epic-labeled issues + recent pull-request markdown including ## Comments sections. Major mid-PR pivot: first draft used regex + stop-phrase blocklist for extraction; tobi's sharp challenge "makes literally ZERO sense" forced the architectural rewrite to LLM-based. Durable memory captures the lesson (feedback_llm_substrate_for_semantic_tasks.md). New validated schema flag + GapInferenceEngine gate: mined candidates carry validated: false + tier 3, silenced via weight-gate + validated-gate (belt-and-suspenders) until curator promotes via nodes.jsonl git-diff edit. Polish commit followed self-review + tobi's magic-number challenge: config-lifted prScanLimit + minSourceLength under new aiConfig.data.conceptDiscovery block. 49/49 stable × 3 runs.

Meta-infrastructure (not #10030-scoped but session-generated)

• 6 new durable feedback memories captured architectural-literacy lessons surfacing this session. See "Durable Lessons Captured" below.
• 1 project memory (project_core_base_conventions.md) preserves tobi's historical rationale for patterns in src/Neo.mjs + src/core/Base.mjs that look like smells but have multi-decade justification.

Current Epic Status

Completed sub-issues (all of #10030's core scope)

• ✅ #10031 — JSONL Schema & Concept Seed Extraction
• ✅ #10032 — ConceptService & Source Code Wiring
• ✅ #10033 — Concept Graph Unit Tests (#10078)
• ✅ #10035 — Graph Integration & GapInferenceEngine Refactor (#10084) — the architectural pillar
• ✅ #10036 — Memory Core Concept Discovery (#10105) — this session
• ✅ #10049 — Concept Schema Enrichment: Aliases, Analogous Edges
• ✅ #10085 — ConceptIngestor follow-ups (#10100) — this session
• ✅ #10086 — Config-lift guideGapWeightThreshold (#10102) — this session
• ✅ #10087 — ORPHAN_CONCEPT signal (#10101) — this session

Open sub-issues (recommended pickup order)

Priority	Sub-Issue	Complexity	Why this order
1	#10106 — ConceptDiscoveryService extraction_metadata schema (new this session)	Small	Direct continuation of #10105. Gemma-4-31b-it's OWN proposed schema for reliable self-report (validated empirically — it agreed generic feedback is unreliable then proposed the structured alternative itself). Quick downstream win with real model-side endorsement.
2	#10037 — ChromaDB Concept Embedding & Hybrid Search	Medium	Now unblocked by the stable concept-node contract. Powers the relevance-bounded queries in #10080.
3	#10034 — Concept Graph Visualization App	Medium	Neo.mjs frontend app surfacing the ontology. Independent; good candidate for a frontend-focused session.
4	#10080 — Relevance-bounded query APIs	Medium	Consumes #10037's embeddings.
5	#10081 — ask_knowledge_base query telemetry	Medium	Agent-FAQ signal for gap inference.
6	#10050 — Concept Description Enrichment	Medium	⚠️ Still blocked on human architect time per tobi's 2026-04-17 annotation — agent-generated descriptions produce wrong results without architectural decision history.

Related Tickets Filed This Session (Not Epic-Parented)

These surfaced from architectural insights during PR self-reviews and stepping-back discussion. All open for future pickup:

• #10103 — SDK-layer config migration (move daemon/service-layer config keys out of memory-core/config.template.mjs). Surfaced when #10102's placement revealed the memory-core-as-catch-all pattern.
• #10104 — ticket-create skill (centralize issue-creation discipline — currently scattered across tool description + ticket-intake + AGENTS_STARTUP §9 + tribal knowledge).
• #10106 — ConceptDiscoveryService extraction_metadata schema (see pickup #1 above).
• #10107 — core.Base + Neo.mjs audit. AUDIT, not refactor. Documents intentional patterns (className-on-instance = DevTools ergonomics; globalThis.Neo = single-registry; setupClass hot-path constraint) + adds benchmark harness as baseline. Any actual refactor becomes downstream tickets that must pass the A/B benchmark gate.
• #10108 — me = this policy decision. Three options; recommended (B): keep FE-only + document boundary in CODING_GUIDELINES.md §34.
• #10109 — PR comment hygiene workflow. Quick-win doc update for .agent/skills/pull-request/references/pull-request-workflow.md.

Durable Lessons Captured This Session

Seven new memory files written to the Agent OS's persistent cross-session memory. These form the six-stage challenge chain now applied at both ticket-intake AND ticket-creation:

1. feedback_challenge_prescribed_fixes.md — verify prescription layer (right substrate), not just premise. Premise-true + prescription-wrong = doubled effort.
2. feedback_symmetric_spec_cleanup.md — Playwright fullyParallel interleaves specs in same worker; cross-singleton mutations need symmetric beforeEach+afterEach resets. beforeEach-only leaks state.
3. feedback_observability_consumer_identity.md — sandman_handoff.md is A2A (next REM cycle, next agent boot, Librarian), NOT a human-facing report. Identify substrate's consumer, not just its durability.
4. feedback_config_service_boundary.md — config-lifts must match consuming service's ownership, not import-convenience. Memory-core has been catch-all but daemon/service-agnostic configs belong in the SDK layer; file migration debt explicitly when destination doesn't exist yet.
5. feedback_no_title_prefix_duplicating_labels.md — ticket titles describe subject, not category. [enhancement]/[bug]/[epic] belong in labels, not titles.
6. feedback_llm_substrate_for_semantic_tasks.md — regex is for STRUCTURAL extraction; semantic tasks (Teaching Test, "what qualifies as X") require LLM. Cheap+deterministic+offline is irrelevant if output is noise. Growing STOP_PHRASES lists are a warning sign.
7. project_core_base_conventions.md — patterns that look like smell (me=this, className-on-instance, globalThis.Neo, setupClass as god-function) have deliberate multi-decade rationale. Don't propose refactors without auditing GC pressure + ripple scope + historical constraint first.

Common failure-mode thread across 1-7: defaulting to path-of-least-resistance without challenging whether the prescription is correct. Each check in the six-stage chain addresses a different axis of this bias.

What to Know for the Next Session

1. Gemma-4-31b-it is genuinely frontier-class for local. Empirically probed this session (MMLU Pro 85.2%, LiveCodeBench v6 80%, Codeforces ELO 2150). When asked directly whether a generic LLM feedback field would be reliable, it correctly identified its own failure modes (confabulation, RLHF bias, correlation/causation) and proposed the narrower structured alternative. Calibrate collaboration posture accordingly — it can reason, not just pattern-match.

2. LM Studio context-length operator control: lms load <model> --context-length <N> --ttl <seconds>. Default 4K is insufficient for source-file context; bump to 32K+ for Neo.mjs reasoning. Persistent across sessions via --ttl.

3. The concept graph is growing — but not from mining yet. #10105 shipped the pipeline but it hasn't been run against the real corpus. Today's nodes.jsonl still has ~60 curated rows, 0 mined. First real discovery run will reveal signal-to-noise empirically — DO that run before deciding whether to ship cross-session Hebbian aggregation (tobi's original clarified feedback-loop idea; deferred to after data is observable).

4. Template/local config parity is a real trap: committing config.template.mjs without mirroring into gitignored config.mjs leaves the worktree with the old config at runtime. .gitignore enforcement catches this at git add time (loud failure) — heed it. bootstrapWorktree.mjs copies from main on worktree creation; operators must manually sync main's local config when templates change.

5. PR self-review via the pr-review skill is load-bearing. Every merged PR this session went through it; every time tobi caught at least one architectural miss that polish commits fixed before merge. The self-review is NOT cosmetic — it's the last point where the discipline chain can catch a prescription mistake before the code lands on dev.

6. Tobi challenges with compressed signal. Short messages often carry sharp architectural challenges ("makes literally ZERO sense", "is the memory core server config really the right spot?", "const feels like bloat"). Read them as invitations to re-calibrate, not as pushback. The "challenge:" = reasoning-test framing from the user profile memory applies throughout.

7. Ticket-inflation vs. A2A-memory trade-off. Tobi's earlier concern "500 tickets in 3 weeks — release notes get harder" was offset by "only closed tickets count for releases" — filing decision-audit tickets (like #10107) is net-positive because their closure produces architectural documentation. Don't over-index on "minimize ticket count"; optimize for "tickets that resolve into durable decisions."

Session Metrics

• 4 PRs merged (#10100, #10101, #10102, #10105)
• 6 new tickets filed (#10103, #10104, #10106, #10107, #10108, #10109) + 4 tickets closed via merges (#10085, #10086, #10087, #10036)
• 7 durable memories saved
• Test delta: 30 → 49 across daemon + ConceptService + new ConceptDiscoveryService specs; 5× stability check on each PR before merge
• Session duration: marathon — multiple architectural pivots caught mid-PR via self-review + tobi's challenges

Origin Session ID

1db25bbe-f39d-4dcd-bb0e-bc125ce91326