Memory/context lens — RUN-GBRAIN-20260820

Provenance note (trial apparatus, not part of the lens result). The memory/context lens worker completed this analysis and returned it to the orchestrator, but both the worker and the orchestrator were terminated by a usage limit before the file was written to disk. This file was recovered verbatim from the worker's returned report. No content was added, and no finding was altered. Formatting may differ in trivial ways from the worker's original file layout.

Lens worker output. Boundary, revision, and IDs as given in the evidence packet. No system execution; no reads outside the frozen checkout except the packet.

0. Scope note and one packet correction

Two packet facts did not survive targeted inspection and should be corrected centrally before publication:

Packet claim Inspected reality Status Anchor
SRC-16: skills/ "(43 skills…)"; boundary table "43 skills" 53 directories, 51 SKILL.md files, 125 files total implemented skills/ tree listing
CMP-16 "Skillpack (43 skills + RESOLVER)" skills/manifest.json declares version: 0.32.3.0 and lists 50 skills, while VERSION is 0.42.25.0 — the manifest is stale relative to the tree implemented skills/manifest.json:1-5; VERSION

Also note: src/schema.sql does not define facts, takes, or query_cache — facts (migration v40), takes (v37), and query_cache (v55/v56) are defined only inside the MIGRATIONS array in src/core/migrate.ts. Any inventory keyed on schema.sql alone will miss the three most memory-relevant tables.

1. Inventory of retained operative parts

Split by content / form / producer / consumer / checks / authority path. RB column: read-back = accumulated or changed through use AND returns to a later invocation or action; retained = state that persists but does not return as input to a later model invocation or action; static = shipped material.

1.1 Facts subsystem (CMP-10 / OBJ-04)

Part Substrate Form Persistence Lineage Producer Consumer Invalidation / regeneration Promotion path RB Status / anchor
facts row (hot memory) Postgres/PGLite facts table natural-language fact + typed fields (kind, notability, visibility, confidence, valid_from/until, claim_metric/value/unit/period) durable, never deleted by consolidation extracted from one conversation turn / page body extractFactsFromTurn (Sonnet by default, facts.extraction_model) recall op, _meta.brain_hot_memory, consolidate, trajectory expired_at, valid_until, superseded_by; forget rewrites the markdown fence → take via consolidate RB implemented — src/core/migrate.ts:2288-2356; src/core/facts/extract.ts:149-254
Facts markdown fence (## Facts block on the entity page) git repo file natural-language table rows durable; system of record same extraction, written to disk first writeFactsToFence parseFactsFence → DB reconcile on extract_facts / rebuild fence is canonical; DB is an index rebuilt from it — RB (survives rebuild; the DB row does not) implemented — src/core/facts/fence-write.ts:1-46
Decayed effective confidence computed, not stored numeric per-call derived from kind + age effectiveConfidence recall, _meta sort, supersession audit, facts_health pure function; recomputed every read — RB (changes ordering of what returns) implemented — src/core/facts/decay.ts:25-56 (halflives: event 7d, commitment/preference 90d, belief/fact 365d)
Facts extraction eligibility predicate in-process module symbolic n/a n/a isFactsBackstopEligible put_page, sync, file_upload, code_import, extract_facts type/slug/length gates; anti-loop dream_generated skip — static implemented — src/core/facts/eligibility.ts:70-98
Facts extraction queue (bounded, in-memory) process memory symbolic ephemeral n/a getFactsQueue().enqueue extraction pipeline cap 100, drop-oldest; drops on shutdown — retained (ephemeral) implemented — src/core/facts/queue.ts:1-40
Facts absorb log ingest_log table symbolic reason codes durable failure events writeFactsAbsorbLog gbrain doctor, admin dashboard none — retained (operator-facing) implemented — src/core/facts/absorb-log.ts:1-40
Phantom-audit / stub-guard JSONL ~/.gbrain/audit/*.jsonl symbolic ISO-week rotated guard fires logStubGuardEvent, phantom pass gbrain doctor none; best-effort writes — retained implemented — src/core/facts/stub-guard-audit.ts:1-31
forget state fence strikethrough + valid_until natural-language + symbolic durable user act forget_fact op / CLI fence parser → DB expired_at legacy pre-v51 rows fall back to DB-only expire and do not survive rebuild — RB (suppresses later return) implemented — src/core/facts/forget.ts:1-33

1.2 Takes / calibration (CMP-11, CMP-19; OBJ-05, OBJ-09, OBJ-10)

Part Substrate Form Persistence Lineage Producer Consumer Invalidation Promotion RB Status / anchor
takes row takes table (+ page fence) NL claim + kind/holder/weight/since_date durable (a) LLM extraction from pages, (b) consolidate bridge from facts consolidate, takes propose --accept, take add think gather (kw + vector), takes CLI/MCP, grading, emotional weight active=false, superseded_by, resolved_* facts → take (kind='fact', holder='self') RB implemented — src/core/migrate.ts:1253-1289; src/core/cycle/phases/consolidate.ts:128-207
take_proposals queue take_proposals table symbolic + NL durable LLM scan of markdown prose propose_takes phase human via gbrain takes propose --accept idempotency key (source_id, page_slug, content_hash, prompt_version) queue → canonical fence only on human accept retained (gated) implemented — src/core/cycle/propose-takes.ts:1-37
take_grade_cache verdict take_grade_cache table symbolic + NL durable judge model over retrieved evidence grade_takes phase calibration scorecard; auto-apply when opted in prompt-version keyed verdict → takes.resolved_* only when cycle.grade_takes.auto_resolve.enabled and confidence ≥ 0.95 retained → RB via profile implemented, evidence stage is a declared stub — src/core/cycle/grade-takes.ts:1-35, 299-305
calibration_profiles row calibration_profiles table NL pattern statements + kebab bias tags + Brier/accuracy/domain_scorecards JSONB durable, append-only per run aggregation over resolved takes calibration_profile phase (Sonnet + voice gate, 2 retries → hand-written template) think <calibration> block; cross-brain mounts superseded by newer generated_at (latest wins; published is not filtered locally) published=true gates cross-brain mount visibility only RB (CLI path only — see §3) implemented — src/core/cycle/calibration-profile.ts:350-381; src/commands/calibration.ts:47-71
think_ab_results table symbolic durable A/B of baseline vs calibration answer think-ab.ts eval reporting — — retained implemented — src/schema.sql:1260-1273

Note the version strings: GRADE_TAKES_PROMPT_VERSION = 'v0.36.1.0-stub' and CALIBRATION_PROFILE_PROMPT_VERSION = 'v0.36.1.0-stub'. The judge prompt and the profile prompt both self-declare stub status at this revision (src/core/cycle/grade-takes.ts:50; src/core/cycle/calibration-profile.ts:41).

1.3 Pages and derived page classes (OBJ-01, OBJ-20)

Split by producer/authority, because they differ materially:

Page class Producer Marker Re-consumed by retrieval? Anti-loop treatment RB Status / anchor
User-authored page human, via git repo none yes eligible for facts extraction RB (acquired material) implemented — src/schema.sql:85-195
Dream-generated page (synthesize phase) Sonnet subagent under allowed_slug_prefixes frontmatter dream_generated: true yes (chunked, indexed) excluded from facts extraction, atom extraction, and transcript discovery RB implemented — src/core/cycle/synthesize.ts:1-27; src/core/facts/eligibility.ts:86-88; src/core/cycle/extract-atoms.ts:182,254,267; src/core/cycle/transcript-discovery.ts:57-72
Synthesis-saved page (think --save) persistSynthesis citations in synthesis_evidence yes remote MCP callers cannot persist (safeSave = remote ? false : …) RB implemented — src/core/operations.ts think handler (safeSave/safeTake)
Atom page extract_atoms phase (Haiku, 1-3 atoms per item) frontmatter.source_hash (16-char sha256) yes idempotency by source_hash; dream_generated excluded RB implemented — src/core/cycle/extract-atoms.ts:1-46
Concept page synthesize_concepts phase; tiers T1 ≥10 / T2 ≥5 / T3 ≥2 / T4 ≥1 atoms; Sonnet narrative for T1/T2, deterministic stub for T3/T4 type: concept yes excludes imported_from-marked atoms RB implemented — src/core/cycle/synthesize-concepts.ts:1-21
Pattern page patterns phase, single Sonnet subagent over recent reflections topic slug from the model yes skipped below min_evidence RB implemented — src/core/cycle/patterns.ts:1-19
Stub / thin page user or import thin body yes enrich_thin develops ≤3 per source per tick; default OFF RB implemented — src/core/cycle/enrich-thin.ts:1-30
Phantom page (unprefixed root slug) accidental writes unprefixed slug until redirected redirect pass migrates fact rows + fence, soft-deletes, bounded 50/cycle retained → removed implemented — src/core/cycle/phantom-redirect.ts:1-37
page_versions snapshot page write path NL + JSONB durable — no consumer found in inspected surface uninspected (consumer)

1.4 Index and graph state (OBJ-02, OBJ-03, OBJ-19)

Part Substrate Form Producer Consumer Invalidation RB Status / anchor
content_chunks + embedding (HNSW) pgvector distributed-parametric + text chunker + embedding provider vector search, think gather embedded_at reset on content change; content_chunks_stale_idx RB implemented — src/schema.sql:279-334
links typed edge links table symbolic auto-link on every page write, regex, no LLM graph traversal (think --anchor), graph signals recomputed on write; pages_links_extracted_at_idx RB implemented — src/schema.sql:424-460
code_edges_chunk / code_edges_symbol tables symbolic resolve_symbol_edges phase (two-pass) code retrieval staleness columns RB implemented — src/schema.sql:370-404
Graph signals (adjacency 1.05×, cross-source 1.10×, session demote 0.95×) computed at query time numeric runPostFusionStages hybrid ranking fail-open; floor-gated RB (past use reshapes present ranking) implemented — src/core/search/graph-signals.ts:1-67
query_cache row query_cache table + embedding symbolic + vector hybridSearch write-back hybridSearch lookup (cosine ≥ 0.92) TTL 3600s; knobs_hash v3 folds mode + embedding column/provider; hit_count/last_hit_at bumped RB — a prior search's output is returned verbatim to a later query implemented — src/core/migrate.ts:2873-2900 + v56 knobs_hash; src/core/search/query-cache.ts:1-30, 76-80
page_aliases / slug_aliases tables symbolic ingest projection of frontmatter aliases: query-side alias match shared normalizer both sides (NFKC + lowercase + whitespace collapse) RB implemented — src/core/migrate.ts:4803-5005; src/core/search/alias-normalize.ts:1-33
timeline_entries table symbolic + NL auto-timeline on write / meeting extraction temporal retrieval, think time-window prompt dedup key widened to (page_id, date, summary, source) RB implemented — src/schema.sql:492-507
pages.emotional_weight (salience) pages column numeric 0..1 recompute_emotional_weight phase, deterministic formula over tags + active takes salience ranking (get_recent_salience, salience sort, query-intent boost) recomputed per cycle; emotional_weight.high_tags config override RB implemented — src/core/cycle/emotional-weight.ts:96-135; src/core/types.ts:495; src/core/engine.ts:1247-1251, 2043
pages.last_retrieved_at pages column timestamp op-layer fire-and-forget write-back on search/query/get_page only (internal callers bypass) LSD stale-page selection (staleBias), not ranking 5-min throttle; default-on with search.track_retrieval escape hatch RB (to a maintenance selection, not to a model's context) implemented — src/core/last-retrieved.ts:1-34, 173-204; src/core/engine.ts:885-886; src/core/postgres-engine.ts:1318-1385
Trajectory points (derived over facts typed claims) computed from facts rows with claim_metric/value + embeddings symbolic + NL findTrajectory engine query + computeTrajectoryStats think <trajectory> block; find_trajectory op; gbrain eval trajectory regression threshold 10% (env-tunable); drift score null under 3 embedded points RB implemented — src/core/trajectory.ts:26-170; src/core/think/index.ts:315-398
Recall cursor state ~/.gbrain/recall-cursors/<source>[.watch].json symbolic recall --since-last-run / --watch tick next recall window selection corrupt / future-shifted → null → 24h default; atomic tmp+rename RB implemented — src/core/recall-cursor-state.ts:1-113

1.5 Execution / control state

Part Substrate Form Producer Consumer RB Status / anchor
subagent_messages table, JSONB content_blocks, schema_version 1/2 NL blocks subagent loop, persisted before tool dispatch crash-replay of the same job retained (replay of one job identity, not cross-task memory) implemented — src/schema.sql:914-936; src/core/minions/handlers/subagent.ts:554, 734, 768, 832-876, 1067, 1108
subagent_tool_executions table, two-phase ledger (pending→complete/failed) symbolic subagent loop crash-replay; also queried by synthesize to learn which slugs children wrote retained + RB for the synthesize orchestrator's write-back step implemented — src/schema.sql:941-967; src/core/cycle/synthesize.ts:6-11
minion_jobs row table symbolic submitters worker claim, timeouts, pg_notify trigger retained implemented — src/schema.sql:776-832, 1276-1289
minion_inbox table (child_done payloads) symbolic child workers parent supervisor retained implemented — src/schema.sql:835-844
dream_verdicts table, keyed (file_path, content_hash) symbolic Haiku "worth processing?" judge synthesize transcript gate RB (a past cheap verdict decides a later run's work) implemented — src/core/migrate.ts v30; src/core/cycle/synthesize.ts:4-6
Audit JSONL family ${GBRAIN_AUDIT_DIR:-~/.gbrain/audit}/<prefix>-YYYY-Www.jsonl symbolic 16 call sites incl. graph-signals, skillopt, shell, supervisor, phantom, batch-retry gbrain doctor, search stats (reads current + previous ISO week) retained (operator/tool-facing, not model-facing) implemented — src/core/audit/audit-writer.ts:1-46
mcp_request_log table, redacted param shape only symbolic MCP dispatch admin SSE, doctor retained implemented — src/mcp/dispatch.ts:75-168
eval_contradictions_cache table, PK (chunk_a_hash, chunk_b_hash, model_id, prompt_version, truncation_policy) symbolic verdict JSONB query-conditioned contradiction judge later probe runs (cost avoidance); dream report RB (verdict reuse), plus report → agent implemented — src/schema.sql:1101-1112; src/core/pglite-engine.ts:3964-4011
eval_contradictions_runs table incl. full report_json symbolic probe runner trend subcommand, doctor retained implemented — src/schema.sql:1118-1135
op_checkpoints, import-checkpoint tables/files symbolic bulk ops resumption retained implemented — src/schema.sql:662-669

1.6 Instruction artifacts (CMP-16, OBJ-14, OBJ-15) — split by mutation authority

Part Substrate Producer Consumer Mutation path RB Status / anchor
Shipped skills/*/SKILL.md (51 files) markdown in agent workspace maintainers host agent platform prompt none at runtime static implemented — skills/ tree
skills/RESOLVER.md router markdown maintainers host agent (dispatch table) none at runtime static implemented — skills/RESOLVER.md:1-40
skills/manifest.json JSON maintainers installer/skillpack tooling none at runtime static (stale vs tree, §0) implemented
skills/_brain-filing-rules.json JSON maintainers synthesize phase's allowed_slug_prefixes (single source of truth) none at runtime static, but load-bearing for enforcement implemented — src/core/cycle/synthesize.ts:14-17
SkillOpt-mutated SKILL.md same file, rewritten in place SkillOpt optimizer after validation gate host agent platform prompt on every subsequent session accepted only when selScore > bestScore + 0.05 on median-of-3 judged rollouts RB — instruction-level self-modification implemented, default OFF (cycle.skillopt.enabled) — src/core/skillopt/validate-gate.ts:55-128; src/core/skillopt/version-store.ts:126-171; src/core/cycle.ts 'skillopt' phase
SkillOpt versions/vNNNN_eE_sS.md + best.md + history.json files under skills/<name>/skillopt/ version store, 5-step ordered write crash-resume; rollback source revertAllPending on crash retained implemented — src/core/skillopt/version-store.ts:1-30, 201-247
SkillOpt proposed.md file optimizer under --no-mutate or bundled-without-flag human review never touches SKILL.md retained (gated) implemented — src/core/skillopt/bundled-skill-gate.ts:53-100
SOUL.md / HEARTBEAT.md / USER.md / ACCESS_POLICY.md markdown templates instantiated in workspace soul-audit skill (agent-run, human-triggered) host agent platform prompt, persistent until regenerated regenerate by re-running soul-audit RB (generated from prior sessions, returns to all later sessions) implemented — templates/SOUL.md.template:1-40; templates/HEARTBEAT.md.template:1-30
Schema pack (gbrain-base, gbrain-base-v2, …) declarative pack file + config maintainers parse, extract, route, cache key, calibration_domains gbrain schema use <name> static/config implemented — src/core/schema-pack/loader.ts, load-active.ts; consumed at src/core/cycle/calibration-profile.ts:327-339
Detected schema candidate (schema detect) computed from pages.source_path prefixes + frontmatter type distribution SQL heuristic, no LLM schema suggest, review-candidates recomputed on demand from disk RB (brain shape reshapes its own type system proposal) implemented — src/core/schema-pack/detect.ts:1-40
Suggested schema change (schema suggest) LLM refinement over runDetect Sonnet-class or heuristic fallback (confidence 0.5) review-candidates → human confidence < 0.6 must not auto-apply retained (gated) implemented — src/core/schema-pack/suggest.ts:1-59
schema-suggest cycle phase wraps runSuggest dream cycle same queue late-cycle retained (gated) implemented — src/core/cycle.ts 'schema-suggest'

1.7 Host-workspace live-context files (not in the packet's registers)

Part Substrate Producer Consumer RB Status / anchor
memory/heartbeat-state.json (awake flag, current location, lastChecks, blockers) workspace JSON heartbeat cron / other producers, outside this checkout generateLiveContext → systemPromptAddition on every assemble() RB (push) implemented — src/core/context-engine.ts:183-197, 413
memory/upcoming-flights.json workspace JSON external producer same RB (push) implemented — src/core/context-engine.ts:199-207, 414
memory/calendar-cache.json (+ lastUpdated staleness) workspace JSON external producer same; >6h emits an explicit staleness warning line RB (push) implemented — src/core/context-engine.ts:218-221, 334-341, 541-543
ops/tasks.md ## Today unchecked items workspace markdown human/agent same; capped at 5 items, 1MB file guard RB (push) implemented — src/core/context-engine.ts:381-407

ingest() on this engine is an explicit no-op (src/core/context-engine.ts:566-569) and ownsCompaction: false — so this engine reads workspace state but writes none of it, and delegates compaction to the OpenClaw runtime (outside the boundary).

1.8 Marked uninspected

  • page_versions consumers — table exists, no reader found in the surface inspected.
  • Full src/core/search/hybrid.ts (1,870 lines): the graph-signals and query-cache seams and the mode/knobs contract were read, not the fusion body. Ranking-stage ordering beyond the four stages named in graph-signals.ts:27-33 is uninspected.
  • src/core/operations.ts (4,751 lines): think, recall, forget_fact, extract_facts headers read. Other ops' read-back behavior is uninspected.
  • Skillpack install/projection mechanics (RTE-17 endpoint side) — the installer path was not read.
  • Everything under test/, evals/, admin/ — excluded by boundary.

2. Write side vs read-back, separated

2.1 Write agency

Write path Agency Trigger Gate
Fact extraction from a turn / page write automatic put_page, sync import, file_upload, code_import (mode queue, fire-and-forget) facts.extraction_enabled kill switch (default ON); eligibility predicate; anti-loop dream_generated
extract_facts MCP op manual (agent-invoked) explicit tool call mode inline, returns truthful counts
Fact fence write automatic, follows extraction — FS page-lock (~/.gbrain/page-locks/<sha256>.lock, PID-liveness + 5-min TTL, 5s timeout); atomic tmp + re-parse + rename
forget_fact manual user/agent fence rewrite, falls back to DB-only for legacy rows
consolidate facts→takes automatic dream cycle phase ≥3 unconsolidated facts per (source_id, entity_slug); oldest ≥24h; cosine cluster ≥0.85; cluster size ≥2; entity page must exist
propose_takes automatic proposal, manual acceptance dream cycle writes only to take_proposals; BAP-10
grade_takes auto-apply manual opt-in cycle.grade_takes.auto_resolve.enabled = true AND confidence ≥ 0.95 config tightening is free; loosening requires --allow-loosen-confidence
calibration_profile automatic dream cycle, after grade ≥5 resolved takes; budget gate ($0.50 default); voice gate
Page write → auto-link, auto-timeline automatic, regex, no LLM every put_page trusted-workspace branch re-enables these for subagent writes
last_retrieved_at bump automatic, fire-and-forget, op-layer only search/query/get_page 5-min throttle; search.track_retrieval off-switch; internal callers bypass entirely
query_cache write automatic after a real hybrid search keyed by (source_id, query_text, knobs_hash)
SkillOpt SKILL.md mutation automatic within an opt-in phase, human-gated for bundled skills cycle.skillopt.enabled (default OFF); benchmark stale AND last_run_at > 7d bundled skills require --allow-mutate-bundled and a held-out set ≥ MIN_HELD_OUT_SIZE, else hard refusal
SOUL.md/HEARTBEAT.md manual operator runs soul-audit none in code (skill-level)
Schema candidate promotion manual schema review-candidates --apply confidence < 0.6 must not auto-apply

2.2 Acquisition and index maintenance vs curation

Acquisition / index maintenance (no epistemic transformation claimed): git sync → pages/chunks import with soft deletes (RTE-15); chunking + embedding (embed phase, embedded_at staleness); auto-link regex edge extraction (RTE-14); alias projection; timeline materialization; resolve_symbol_edges; query_cache population; last_retrieved_at.

Curation (a model or a human decides what is worth keeping, at what strength): notability filter in the extractor prompt (high extract now / medium wait for batch / low skip entirely — src/core/facts/extract.ts:124-131); sync's high-only notability filter (backstop.ts D4); consolidate's cluster-representative selection (highest-confidence fact's text, verbatim, no LLM synthesis at this revision — consolidate.ts:130-133); propose_takes proposal set; grade_takes verdicts; calibration pattern statements; atom/concept tiering.

The consolidate step is worth naming precisely: the take's claim is a copy of the best fact's text, not a synthesis. The header itself flags "v0.31 ships without LLM synthesis to keep the cycle deterministic" (consolidate.ts:9-12). So the facts→takes edge is a selection + re-typing + re-attribution operation (kind='fact', holder='self', weight = mean(confidence)), not a semantic distillation. The label consolidate does not establish semantic preservation — it establishes a copy with a new authority attribution (holder='self') and a new confidence number computed by averaging.

2.3 Named mechanisms

Mechanism Present? How Anchor
Consolidation yes greedy cosine clustering (threshold 0.85, first-member-as-centroid), ≥2 per cluster → one take; contributing facts marked consolidated_at + consolidated_into, never deleted consolidate.ts:128-290
Deduplication yes, several distinct kinds (a) fact insert dedup at cosine 0.95 with supersede support (backstop.ts:1-25); (b) takes semantic upsert on (page_id, claim, since_date) to survive extract_facts's delete-and-reinsert (consolidate.ts:150-175); (c) take_proposals idempotency on (source_id, page_slug, content_hash, prompt_version); (d) atom idempotency on frontmatter.source_hash; (e) timeline (page_id,date,summary,source) unique index multiple
Evolution yes bitemporal writeback: consolidate sorts each cluster chronologically and stamps each older fact's valid_until = next_newer.valid_from; newest keeps NULL consolidate.ts:209-238
Synthesis partial present in think, patterns, synthesize_concepts (T1/T2 only), calibration_profile; absent in consolidate at this revision consolidate.ts:9-12
Invalidation yes expired_at, valid_until, superseded_by, active=false, deleted_at soft delete + 72h purge, query_cache TTL + knobs_hash version bump, eval_contradictions_cache.expires_at, prompt-version cache-key bumps multiple
Decay yes, read-time only effectiveConfidence exponential half-life per kind; nothing is written back decay.ts:44-56
Promotion yes, four ladders facts→takes (automatic, thresholded); proposals→takes (human); verdicts→resolved takes (opt-in + 0.95); atoms→concepts (count tiers); candidate skill→SKILL.md (gated) or →proposed.md (human) multiple
Salience weighting yes emotional_weight = tag boost 0.5 + take density 0.3 + avg weight 0.1 + user-holder ratio 0.1, capped at 1.0; deterministic, no LLM emotional-weight.ts:96-135

2.4 Raw traces vs distilled retained artifacts

Raw trace Distilled artifact Is the raw kept?
Conversation turn text facts row (≤500 chars, sanitized in and out) No — the turn is not stored by this subsystem
Transcript file dream verdict → synthesized page(s) + summary index; atoms Yes, on disk in the corpus dir
facts rows takes row Yes — "NEVER DELETE"; facts stay as audit trail
Resolved takes calibration_profiles narrative Yes
Rollout trajectories SkillOpt selScore, history.json edits Trajectories scored then discarded from the version store; audit JSONL keeps per-task medians
Subagent turns none Yes, subagent_messages (until job row cascade-deletes)

3. Annotation of runtime-owned context routes carrying read-back

Read-back direction is stated from the receiving agent's perspective. Per instruction, post-turn capture / consolidation is counted as write-side maintenance, not a second read-back point.

RTE-07 — _meta.brain_hot_memory injection on every MCP tool response

Field Value Status / anchor
Direction push — the receiving agent does not ask; the dispatcher attaches it to whatever tool result it already requested implemented — src/mcp/dispatch.ts:255-267
Selection signal session-scoped facts first (listFactsBySession); if empty, source-wide facts from the last 24h (listFactsSince) implemented — meta-hook.ts:69-79
Targeting (source_id, session_id); not query- or task-conditioned. Self-exclusion: skipped for recall, extract_facts, forget_fact implemented — meta-hook.ts:47, 49-53
Selection scope + budget top-K default 10, hard cap 25; sorted by decayed effective confidence before truncation implemented — meta-hook.ts:56, 85-88
Delivery + consumption point ToolResult._meta on a successful op; content shape unchanged. Cache key (source_id, session_id, sorted takesHoldersAllowList), TTL 30s, invalidated by bumpHotMemoryCache after extraction implemented — meta-hook.ts:21-29, 107-120
Visibility remote → ['world'] only; local (ctx.remote === false) → all rows implemented — meta-hook.ts:66
Behavioral-faithfulness test none found. Failure degrades silently to no-_meta; the client may ignore unknown _meta entirely absent (test), implemented (degradation) — dispatch.ts:259-267

RTE-08 — recall op → facts

Field Value Status / anchor
Direction pull — the agent issues the tool call implemented — src/core/operations.ts:3545-3655
Selection signal five mutually exclusive branches in priority order: supersessions → entity (via resolveEntitySlug) → session_id → since → unfiltered recent implemented
Targeting agent-supplied entity slug / session / time window; grep substring applied client-side after the DB query implemented — recall handler, if (grep) rows = rows.filter(...)
Scope + budget limit default 50, documented cap 100; activeOnly unless include_expired implemented
Delivery full JSON row projection incl. confidence, notability, superseded_by, consolidated_into; optional pending_consolidation_count piggy-back implemented
Cursor coupling recall --since-last-run reads/writes a per-source file cursor, briefing and watch variants kept separate so quitting a watch doesn't lose the briefing position implemented — recall-cursor-state.ts:1-35
Behavioral-faithfulness test none found absent

RTE-04 — query → hybrid retrieval → ranked chunks

Field Value Status / anchor
Direction pull, with one push-shaped shortcut: a semantic cache hit returns a prior call's results without running retrieval implemented — query-cache.ts:1-30
Selection signal vector + BM25 + RRF + rerank, then four post-fusion stages including graph signals; alias match; recency decay; source boost; intent weighting; emotional_weight salience boost implemented (seams read), fusion body uninspected
Targeting free-text query, source_id-scoped; sourceScopeOpts(ctx) federated-array > scalar precedence implemented — BAP-07
Scope + budget search mode bundle: conservative 4K tokens / limit 10, balanced 12K / 25, tokenmax off / 50 + LLM expansion. Resolution chain: per-call → per-key config → MODE_BUNDLES[search.mode] → balanced implemented — src/core/search/mode.ts; described CLAUDE.md "Search Mode"
Cache correctness knobs_hash (v3) folds mode bundle + active embedding column + provider, so a tokenmax write can't serve a conservative read and a Voyage-1024d row can't serve an OpenAI-1536d read; similarity ≥ 0.92; TTL 3600s implemented — migration v56; mode.ts:KNOBS_HASH_VERSION
Read-back write-side coupling last_retrieved_at bump (5-min throttled, fire-and-forget) — write-side maintenance, feeds LSD selection, not this call's context implemented — last-retrieved.ts:173-204
Behavioral-faithfulness test benchmark numbers exist (P@5 49.1%, R@5 97.9%, +31.4 P@5) but the runs live in the sibling gbrain-evals repo, outside the boundary claimed — packet §4

RTE-05 — think: gather → merge → synthesize

Three separately-wired read-back injections converge here; they must not be collapsed.

Sub-route Direction Selection signal Targeting Scope / budget Delivery point Status
Pages stream pull hybrid search, expansion: false (think supplies its own anchor/graph context) question text gatherLimit 40; excerpt 600 chars each <pages> block implemented — gather.ts:110-116, 184-195
Takes streams pull keyword + vector over active takes, RRF-fused (k=60) on (page_slug, row_num) question text (+ embedding when supplied) takesLimit 30 <takes> block, declared DATA not instructions implemented — gather.ts:118-163; prompt.ts:43-45
Graph stream pull traversePaths(anchor, depth 2, both) requires explicit --anchor first 30 slugs <graph> block implemented — gather.ts:139-153; think/index.ts:282-284
Calibration profile push into the call, opt-in latest calibration_profiles row for holder, ordered generated_at DESC LIMIT 1; published is not filtered on the local path --with-calibration flag; holder default 'garry' pattern statements + bias tags + Brier <calibration> block, placed after retrieval and before the question; system prompt gains anti-bias instructions implemented on the CLI path; absent on the MCP path — think/index.ts:286-313; commands/think.ts:80,119; the think op handler in operations.ts never sets withCalibration
Trajectory push into the call, default-on classifyIntent(question) must be temporal or knowledge_update; entity candidates extracted from question + retrieved slugs; fallback_slugify resolutions rejected per-candidate, concurrency 3, 5s timeout each, limit: 100 points rendered <trajectory> block(s), joined between retrieval/calibration and the question in both prompt shapes; emits TRAJECTORY_INJECTED_N_POINTS warning implemented — think/index.ts:315-398; kill switch think.trajectory_enabled

Behavioral-faithfulness tests on RTE-05: the system prompt carries four hard rules (cite everything, mark weight<0.5 or kind=hunch explicitly, surface both sides of a conflict, name gaps, never instruct) — these are instructions, not tests. The one code-side check is deterministic citation persistence: the model must emit a structured citations array, with a regex fallback in cite-render.ts recovering inline [slug#row] markers when it doesn't. Compliance with the four hard rules is uninspected.

RTE-12 — verdicts → calibration profile → think prompt

Direction: push (write-side aggregation), then push at the think consumption point. The chain's evidence stage is a declared placeholder at this revision: defaultEvidenceRetriever returns a literal [evidence retrieval not yet wired — v0.36.1.0 ship-state] string containing only the take's own claim text (grade-takes.ts:299-305). The module header states most stub-judge verdicts will therefore be unresolvable. Combined with auto_resolve default-OFF, the takes reaching resolved_* in a default deployment arrive through some path other than this phase — which path is uninspected.

PROPOSED-RTE-A — workspace live-context → host agent system prompt (every turn)

Field Value Status / anchor
Direction push — injected on every assemble(); the receiving agent never asks implemented — context-engine.ts:571-600
Selection signal file presence + freshness only; no query conditioning, no LLM, no ranking implemented — context-engine.ts:409-475
Targeting fixed paths under workspaceDir: memory/heartbeat-state.json, memory/upcoming-flights.json, memory/calendar-cache.json, ops/tasks.md implemented
Scope + budget ≤3 upcoming events (4h lookahead), ≤5 open tasks, attendee lists truncated at 3, every string sanitized to ≤100 chars with control chars stripped, tasks file capped at 1MB implemented — context-engine.ts:150-152, 371-378, 403, 383
Delivery point AssembleResult.systemPromptAddition, concatenated with the memory plugin's addition implemented
Behavioral-faithfulness test partial and unusual: the engine refuses to emit a concrete local time when the timezone is unmapped (UNKNOWN_TZ sentinel), rendering an explicit "Local time NOT computed" warning instead of a confidently-wrong value; it also emits a staleness warning when the calendar cache is >6h old. These are honesty guards on the injected content, not tests that the model acted on it implemented — context-engine.ts:171-179, 500-509, 541-543
Note the block ends with > This block is computed on every turn. Trust it over compaction summaries… — an explicit instruction to prefer this channel over the host's own memory implemented — context-engine.ts:546

4. Four separate findings per main read-back path

Statuses are never upgraded. "Deployed wiring" = the material actually reaches a live consumption point in a default configuration. "Activation" = evidence that delivered material changed behavior. "Causal effect" = evidence the change is attributable to the material.

4.1 Facts → _meta.brain_hot_memory (RTE-07)

Finding Status Evidence What the status blocks
Context presence implemented meta-hook.ts:90-108 constructs a payload of up to 10 fact objects with id/text/kind/notability/entity/valid_from/decayed-confidence —
Deployed wiring implemented Both transports pass the hook: src/mcp/server.ts:46 and src/commands/serve-http.ts:1576 —
Activation uninspected Nothing executed. _meta is by design ignorable by clients (dispatch.ts:20-26), so presence in the response is not presence in the model's attended context Blocks: "the downstream agent's answers reflect hot memory." Cannot be asserted.
Causal effect uninspected No experiment; no A/B; the README production counts are claimed and their runs are outside the boundary Blocks: "hot-memory injection improves agent behavior." Cannot be asserted at any strength.

4.2 Facts → recall op (RTE-08)

Finding Status Evidence What the status blocks
Context presence implemented Full row projection returned by the handler —
Deployed wiring implemented recall is a registered scope: 'read' operation, generated into both CLI and MCP surfaces by the contract-first pattern; _meta self-exclusion at meta-hook.ts:47 confirms it is expected to be called by agents —
Activation uninspected No trace of an agent choosing recall over search in any inspected artifact Blocks: "agents actually pull hot memory rather than relying on the push channel."
Causal effect uninspected — Blocks any claim that pull-based recall outperforms or complements push injection.

4.3 Takes + graph → think synthesis (RTE-05, core streams)

Finding Status Evidence What the status blocks
Context presence implemented <pages>/<takes>/<graph> blocks assembled in both prompt shapes (prompt.ts:155-226) —
Deployed wiring implemented think op registered scope: 'write', mutating: true; handler forwards source scope, allow-list, remote flag —
Activation uninspected The prompt asks for citation of every claim and explicit hunch-marking; whether output complies is not inspected. One partial code-side signal exists: cite-render.ts implements a regex fallback for models that omit the structured citations field — the existence of that fallback is evidence the authors expected non-compliance, but it is not evidence of behavior Blocks: "think output is faithfully grounded in the retrieved takes."
Causal effect uninspected — Blocks: "typed/weighted takes improve synthesis over plain page retrieval."

4.4 Calibration profile → think prompt (RTE-12 consumption point)

Finding Status Evidence What the status blocks
Context presence implemented buildCalibrationBlock renders Brier + patterns + bias tags; buildThinkSystemPrompt adds four anti-bias instructions (prompt.ts:88-96, 113-130) —
Deployed wiring split: implemented on the CLI path (gbrain think --with-calibration, commands/think.ts:80,119); absent on the agent-facing MCP path — the think op handler never passes withCalibration, and think's declared params do not include a calibration flag This is the sharpest structural finding in the lens. Blocks: "calibration is part of GBrain's agent-facing memory read-back." It is a human-CLI feature at this revision.
Activation uninspected — Blocks: "surfacing prior + counter-prior changes the answer."
Causal effect uninspected; and the upstream evidence stage is implemented-as-stub grade-takes.ts:299-305 returns a placeholder string as "evidence"; auto_resolve default OFF; both prompt versions self-label -stub Blocks any claim that GBrain's calibration reflects a real forecasting track record. The pipeline shape exists; the evidence input does not.

4.5 Trajectory → think prompt

Finding Status Evidence What the status blocks
Context presence implemented formatTrajectoryBlock output spliced into both prompt shapes at a single defined slot (prompt.ts:132-154) —
Deployed wiring implemented, default ON, but narrowly gated: only for temporal / knowledge_update intent, only for entities that resolve by a means other than fallback_slugify, only when typed-claim facts with claim_metric/claim_value exist Blocks: "trajectory is generally available." It requires the extractor to have emitted metric-shaped claims.
Activation uninspected The code emits a TRAJECTORY_INJECTED_N_POINTS warning string — an instrumentation signal that injection happened, not that it mattered Blocks: "the model's temporal claims are grounded in trajectory."
Causal effect uninspected gbrain eval trajectory exists as a CLI surface; no results in the boundary Blocks any claim of regression-detection or drift-score utility.

4.6 SkillOpt → SKILL.md → host agent (RTE-18 → RTE-17)

Finding Status Evidence What the status blocks
Context presence implemented Accepted candidate text is written verbatim to SKILL.md, the file the host platform loads into its prompt (version-store.ts:159-160) —
Deployed wiring implemented, default OFF and multiply gated cycle.skillopt.enabled default false; bundled skills → proposed.md unless --allow-mutate-bundled; bundled in-place mutation hard-refuses without a held-out set ≥ MIN_HELD_OUT_SIZE (bundled-skill-gate.ts:86-100) Blocks: "GBrain self-modifies its instructions in a default install." It does not.
Activation uninspected The acceptance rule is itself a measurement: median-of-3 judged rollouts per sel-task, accept only if selScore > bestScore + 0.05 (validate-gate.ts:118-121). That is evidence the candidate scored higher on a judged benchmark, not evidence that a later live session behaved differently Blocks: "an optimized skill changes downstream agent behavior in production."
Causal effect uninspected The related functional-area-resolver A/B (+13 to +17pp at 48% size) is a different artifact (evals/functional-area-resolver/, deliberately outside skills/) and its runs are outside the boundary Blocks attributing any behavioral gain to the SkillOpt loop specifically.

4.7 Workspace live context → host system prompt (PROPOSED-RTE-A)

Finding Status Evidence What the status blocks
Context presence implemented formatContextBlock output returned as systemPromptAddition (context-engine.ts:495-549, 598) —
Deployed wiring implemented, conditional on host configuration — requires plugins.slots.contextEngine: "gbrain-context" in openclaw.json; outside OpenClaw the SDK import fails and the fallbacks return undefined memory addition (context-engine.ts:90-94) Blocks: "GBrain injects live context for all host platforms." It is OpenClaw-specific and opt-in.
Activation uninspected The block instructs the model to trust it over compaction summaries; compliance not inspected Blocks: "the time-warp bug class is fixed." The doctrine claims the motivation; the code shows the mechanism.
Causal effect uninspected — Blocks any claim about compaction-loss mitigation.

4.8 Query cache → search results

Finding Status Evidence What the status blocks
Context presence implemented A hit returns the stored results JSONB directly with no keyword search, vector search, expansion, RRF, or dedup (query-cache.ts:4-11) —
Deployed wiring implemented, on in all three shipped modes (cache.enabled: true, threshold 0.92, TTL 3600s) —
Activation implemented-as-mechanism, uninspected-as-behavior The cache is the delivered context on a hit — there is no separate activation question for the retrieval stage. What is uninspected is whether the downstream model's answer differs Blocks: "cache hits are behaviorally equivalent to fresh retrieval." Note the design history: two prior contamination bugs (mode cross-serving, embedding-provider cross-serving) were fixed by widening the key, which is evidence that equivalence was not initially true.
Causal effect claimed only "Cache hits cut all numbers ~50%" is a cost claim in CLAUDE.md, not an inspected measurement Blocks quantitative cost or quality claims.

5. Authority

Referencing the packet's BAP records, with consumer / channel / force / horizon kept explicit rather than substituted by a family label.

BAP Consumer Channel Force Horizon Lens annotation
BAP-01 host platform LLM skillpack markdown in workspace prompt advisory instruction every session where installed 51 SKILL.md files; RESOLVER.md is a routing table, not enforcement. _brain-filing-rules.json is the one skillpack file that becomes binding downstream — synthesize reads it as the subagent allowed_slug_prefixes source, where BAP-05 enforces it in code. Same artifact, two force levels depending on channel.
BAP-02 host platform LLM SOUL.md/HEARTBEAT.md advisory instruction persistent until soul-audit regenerates Templates ship with unresolved <!-- Fill in via soul-audit --> slots (templates/SOUL.md.template:20-23). An un-audited install carries literal placeholder text into the prompt.
BAP-03 subagent loop model buildSystemPrompt(toolDefs, data.system) advisory instruction one subagent job Determinism is a hard requirement here for cache-hit reasons, not epistemic ones (system-prompt.ts:14-21). Caller can bypass the preamble entirely with system_no_tool_preamble.
BAP-05 brain tool executor filterAllowedTools / allowed_slug_prefixes / PROTECTED_JOB_NAMES binding enforcement (code refuses) one subagent job The synthesize header states the allow-list is trusted because PROTECTED_JOB_NAMES prevents MCP from submitting subagent jobs — the binding force depends on a second, separate gate holding.
BAP-06 MCP operation dispatcher scope + localOnly + OperationContext.remote binding enforcement, pre-handler every remote MCP call Fail-closed by type: remote is required and anything not strictly false is untrusted (CLAUDE.md cross-cutting invariants). Materialized in dispatch.ts:205 (remote: opts.remote ?? true).
BAP-07 read-side operations sourceScopeOpts(ctx), takesHoldersAllowList binding enforcement (SQL predicate) every read op Token default is {"takes_holders":["world"]} — non-world takes hidden until explicitly granted (migrate.ts v38). The _meta hot-memory cache key includes the sorted allow-list hash so cached payloads cannot bleed across token tiers (meta-hook.ts:8-9, 52-53).
BAP-08 think synthesis model THINK_SYSTEM_PROMPT_BASE hard rules advisory instruction one think call Four rules: cite everything, mark low-weight/hunch, surface both sides of conflicts, name gaps, never instruct. All advisory. Only the citation rule has a code-side recovery path.
BAP-09 think synthesis model <take> declared DATA + INJECTION_PATTERNS advisory + partial code enforcement one call The sanitizer is reused as the single source of truth across takes, think, and fact extraction — applied to turn text on the way in and to each extracted fact on the way out (extract.ts:155, 207). Emits SANITIZED_N_TAKE_CLAIMS warnings.
BAP-10 human operator propose_takes queue; schema review-candidates; SkillOpt proposed.md binding gate (no write without accept) per proposal Three independent human gates with different shapes: a DB queue, a disk-derived candidate list, and a file. The SkillOpt one is doubly bound — assertBundledMutationHeldOut throws rather than degrading (bundled-skill-gate.ts:92-99).
BAP-12 downstream agent _meta.brain_hot_memory advisory context (client may ignore unknown _meta) per tool call, 30s cache The weakest authority in the system and the most-injected content. Correctly labeled advisory: there is no channel by which GBrain learns whether the payload was read.

Lineage/curation labels do not carry epistemic force. Three cases where a label could be misread:

  1. consolidate promotes facts to takes and stamps holder='self'. That is an attribution assignment, not an inference that the user holds the claim — the underlying fact may have been extracted from a third party's statement in a meeting transcript. The label does not establish that the take is warranted or that its meaning is preserved (the claim text is copied verbatim; only weight is newly computed, as a mean of confidences).
  2. sync imports git markdown into pages/chunks. That is acquisition. Nothing in the import path evaluates the content.
  3. grade_takes writes a "verdict". At this revision the verdict is produced against placeholder evidence. verdict names the row's role in the schema, not its epistemic standing.

published on a calibration profile is likewise an exposure flag for cross-brain mounts, not an acceptance mark: the local think path reads the newest profile regardless of published (commands/calibration.ts:51-71 has no published predicate), while cross-brain.ts:117-120 filters on it. Same field, two meanings; only the mount-side one is a gate.

6. Proposed new records for central registration

Proposed only — the orchestrator owns registration. None of these rename or re-inventory an existing record.

Components

Proposed ID Component Substrate Form Why it needs its own record
PROPOSED-CMP-01 GBrain Context Engine (OpenClaw plugin, createGBrainContextEngine) in-process module + host workspace files mixed (TS + rendered NL block) Carries a per-turn push read-back route into the host system prompt; no existing CMP covers it. Zero LLM calls, ownsCompaction: false.
PROPOSED-CMP-02 Facts markdown-fence writer / parser (fence-write.ts, facts-fence.ts, extract-from-fence.ts) git repo files + FS page-locks natural-language table CMP-10 covers the facts table; the fence is the declared system of record and has different persistence, different concurrency control, and different rebuild survival.
PROPOSED-CMP-03 Trajectory subsystem (trajectory.ts, trajectory-format.ts, findTrajectory) derived over facts typed-claim columns symbolic + NL Feeds a default-ON think injection slot; no existing CMP.
PROPOSED-CMP-04 Shared audit-writer primitive + 16 consumers filesystem JSONL, ISO-week rotated symbolic CMP-20 names the trail; this names the shared primitive and its GBRAIN_AUDIT_DIR override, current+previous-week read window, and best-effort posture.

Operative objects

Proposed ID Object Substrate Producer Consumer
PROPOSED-OBJ-01 Facts fence row (markdown table row on an entity page) git repo file writeFactsToFence fence parser → DB reconcile; survives rebuild
PROPOSED-OBJ-02 Trajectory point / TrajectoryStats (regressions + drift_score) derived findTrajectory + computeTrajectoryStats think <trajectory> block, find_trajectory op
PROPOSED-OBJ-03 emotional_weight salience score pages column recompute_emotional_weight phase salience sort, get_recent_salience, query-intent boost
PROPOSED-OBJ-04 last_retrieved_at retrieval-recency stamp pages column op-layer fire-and-forget write-back LSD staleBias selection
PROPOSED-OBJ-05 Recall cursor record ({schema_version:1, last_run_iso}) ~/.gbrain/recall-cursors/*.json recall --since-last-run / --watch next recall window
PROPOSED-OBJ-06 Take proposal (queued, unaccepted) take_proposals propose_takes human accept gate
PROPOSED-OBJ-07 SkillOpt version snapshot + history.json row + best.md/proposed.md files under skills/<name>/skillopt/ version store 5-step commit crash-resume, rollback, human review
PROPOSED-OBJ-08 Dream verdict (is this transcript worth processing?) dream_verdicts, keyed (file_path, content_hash) Haiku judge synthesize gate on later runs
PROPOSED-OBJ-09 Live-context block (## Live Context (deterministic…)) rendered string formatContextBlock host system prompt, every turn
PROPOSED-OBJ-10 Pattern page pages patterns phase Sonnet subagent retrieval; distinct from OBJ-20 atoms/concepts
PROPOSED-OBJ-11 Detected/suggested schema candidate computed from disk + optional LLM runDetect / runSuggest review-candidates → human

Routes

Proposed ID Route Kind Endpoints
PROPOSED-RTE-A Workspace state files → assemble() → systemPromptAddition context (read-back, push), every turn heartbeat/flights/calendar/tasks → context engine → host LLM
PROPOSED-RTE-B Query → semantic query-cache hit → prior call's results context (read-back, pull, bypasses retrieval) query embedding → query_cache → caller
PROPOSED-RTE-C Facts typed claims → findTrajectory → <trajectory> block context (read-back, push into think) facts → engine → think prompt
PROPOSED-RTE-D Tags + active takes → recompute_emotional_weight → salience ranking state → context tags/takes → pages.emotional_weight → search ordering
PROPOSED-RTE-E Retrieval op → last_retrieved_at → LSD stale-page selection state (write) → maintenance selection ops → pages column → staleBias query
PROPOSED-RTE-F Transcript hash → dream_verdicts → later synthesize gate checking + state transcript → Haiku → cache → gate
PROPOSED-RTE-G Unprefixed phantom page → redirect pass → canonical slug + migrated facts state (repair) pages → resolver → fence + DB migration
PROPOSED-RTE-H Page write → fact fence → DB reconcile on extract_facts/rebuild state (write, markdown-first) write path → fence → facts

Behavioral-authority paths

Proposed ID Consumer Channel Force Horizon
PROPOSED-BAP-A Host platform LLM systemPromptAddition from the context engine, incl. the explicit "trust it over compaction summaries" line advisory instruction + advisory context every turn, while the OpenClaw slot is configured
PROPOSED-BAP-B SkillOpt acceptance runValidationGate strict > bestScore + 0.05 on median-of-3 judged rollouts binding enforcement (candidate rejected in code) one optimization step
PROPOSED-BAP-C SkillOpt bundled-mutation assertBundledMutationHeldOut (throws when held-out < MIN_HELD_OUT_SIZE) binding enforcement (hard refusal, every entry point) per run
PROPOSED-BAP-D grade_takes auto-apply cycle.grade_takes.auto_resolve.enabled + confidence ≥ 0.95 + tighten-only config ratchet binding enforcement, default-closed per verdict
PROPOSED-BAP-E Facts extraction facts.extraction_enabled kill switch + isFactsBackstopEligible + dream_generated anti-loop binding enforcement (code refuses to extract) every write surface
PROPOSED-BAP-F Fence write path FS page-lock (~/.gbrain/page-locks/<sha256>.lock, PID-liveness + 5-min TTL, 5s timeout) + atomic tmp/re-parse/rename binding enforcement (multi-process serialization; DB not written when fence invalid) per page write

7. Files read inside the frozen checkout

All at 9a0bae8, under related-systems/gbrain/. Read-only; nothing executed.

File Extent What was taken
CLAUDE.md full (auto-surfaced) search-mode bundles + knobs_hash v3 contract, cost anchors, trust/source-isolation invariants, skills count claim (29), functional-area-resolver A/B numbers
src/schema.sql index-grep + lines 279-334, 370-404, 424-460, 492-530, 662-669, 776-844, 905-980, 1101-1145, 1260-1344 table inventory; chunks/links/timeline/page_versions/minion/subagent/contradiction schemas; RLS block; confirmed facts/takes/query_cache are absent here
src/core/migrate.ts lines 1240-1330, 2280-2400, 2601-2640, 2865-2930, 4803-5005 (grep) facts DDL v40 (incl. row_num/source_markdown_slug fence round-trip), takes DDL v37 + synthesis_evidence, access_tokens.permissions default v38, query_cache v55 + knobs_hash v56, alias tables
src/core/facts/meta-hook.ts full (133) RTE-07 payload shape, top-K/TTL, cache key incl. allow-list hash, visibility tiering, self-exclusion list
src/core/facts/decay.ts full (63) half-life table, read-time-only decay formula
src/core/facts/extract.ts full (332) extractor system prompt (kinds, notability tiers, typed-claim fields), sanitization in/out, Sonnet default, kill switch, 4-strategy JSON parse
src/core/facts/eligibility.ts full (98) eligible types union, rescue slug prefixes, dream_generated and wiki/agents/ exclusions, 80-char floor
src/core/facts/fence-write.ts header (60) markdown-first invariant, page-lock, atomic write + quarantine
src/core/facts/backstop.ts header (70) five-surface choke point, queue vs inline modes, notability filter policy, provenance source enum
src/core/facts/forget.ts header (45) fence-rewrite forget, strikethrough parse contract, legacy DB-only fallback that does not survive rebuild
src/core/facts/queue.ts header (40) cap 100, per-session in-flight 1, shutdown drop semantics
src/core/facts/absorb-log.ts header (45) failure reason codes into ingest_log
src/core/facts/stub-guard-audit.ts header (35) ISO-week JSONL, two-file read window rationale
src/core/cycle.ts lines 57-230 CyclePhase union, full ALL_PHASES ordering + rationale comments, PHASE_SCOPE taxonomy, which phases are default-OFF
src/core/cycle/phases/consolidate.ts full (297) thresholds (3 facts / 24h / cosine 0.85 / cluster ≥2), verbatim-claim selection, semantic upsert key, chronological valid_until writeback, never-delete rule
src/core/cycle/calibration-profile.ts full (406) pattern/bias-tag prompts, voice gate + template fallback, ≥5-resolved floor, budget gate, published: false on insert, domain_scorecards pack resolution, -stub prompt version
src/core/cycle/propose-takes.ts header (55) queue-only posture, idempotency key, fence dedup, tuned prompt version + reported F1
src/core/cycle/grade-takes.ts header (60) + defaultEvidenceRetriever evidence-retrieval placeholder, auto-resolve default-OFF, 0.95 threshold, tighten-only ratchet, -stub prompt version
src/core/cycle/emotional-weight.ts full (142) deterministic 0..1 formula and its four components; default high-emotion tag set
src/core/cycle/synthesize.ts header (60) dream verdict cache, subagent fan-out, allow-list source of truth, subagent_tool_executions slug harvest, cooldown key
src/core/cycle/patterns.ts header (40) reflection lookback, single-subagent design, trusted-workspace branch
src/core/cycle/extract-atoms.ts header (60) source_hash idempotency, budget, pack gating, known partial-write limitation
src/core/cycle/synthesize-concepts.ts header (50) tier thresholds, T1/T2 Sonnet vs T3/T4 stub narrative
src/core/cycle/enrich-thin.ts header (50) default-OFF, per-tick caps, budget/walltime structure
src/core/cycle/auto-think.ts header (35) default-disabled, cooldown key, not wired into the main dispatcher at this revision
src/core/cycle/phantom-redirect.ts header (50) phantom migration, 50/cycle bound, lock contract, idempotency
src/core/think/gather.ts full (213) four-stream gather, RRF k=60, limits 40/30, depth 2, per-stream failure isolation, block rendering
src/core/think/prompt.ts full (226) system prompt hard rules, output schema, calibration block, trajectory slot placement in both message shapes
src/core/think/index.ts lines 250-428 + greps gather wiring, calibration fetch + NO_CALIBRATION_PROFILE warning, trajectory gating/timeouts/concurrency, prompt assembly
src/core/trajectory.ts full (170) regression threshold + env override, drift score definition and null condition
src/core/last-retrieved.ts full (204) op-layer-only scope, 5-min throttle, config escape hatch, drain machinery
src/core/recall-cursor-state.ts full (121) two cursor variants, atomic write, null-return conditions
src/core/context-engine.ts lines 1-611 (three reads) live-context file set, sanitization, unknown-TZ refusal, staleness warning, ingest no-op, assemble output, "trust it over compaction" line
src/openclaw-context-engine.ts full (66) plugin registration shape, plugins.slots.contextEngine requirement
src/mcp/dispatch.ts full (284) metaHook invocation point and its isolation, param redaction + byte bucketing, remote ?? true default
src/core/operations.ts think op block, recall op block (~3540-3660), forget_fact header remote persist blocking, source-scope threading, absence of a calibration flag, recall branch order and caps
src/commands/calibration.ts getLatestProfile + formatProfileText + op entry (~47-246) no published predicate on local reads
src/core/calibration/cross-brain.ts grep, lines ~6-120 published=true mount-side filter
src/core/search/query-cache.ts lines 1-120 cache semantics, threshold/TTL defaults, row-id derivation with knobs_hash
src/core/search/graph-signals.ts lines 1-90 three signal magnitudes, floor gate, fail-open, audit-writer use
src/core/search/alias-normalize.ts lines 1-56 shared read/write normalizer
src/core/skillopt/version-store.ts full (247) 5-step commit ordering, crash-resume, writeProposed
src/core/skillopt/validate-gate.ts full (176) median-of-3, epsilon 0.05, must-abort handling
src/core/skillopt/bundled-skill-gate.ts full (100) bundled detection, mutate policy, hard-refusal on missing held-out
src/core/schema-pack/detect.ts header (40) SQL-only heuristic, privacy posture
src/core/schema-pack/suggest.ts header (60) LLM refinement, heuristic fallback confidence 0.5, <0.6 no-auto-apply rule
src/core/schema-pack/review.ts header (50) disk-derived candidates vs redacted audit
src/core/audit/audit-writer.ts header (50) ISO-week filenames, GBRAIN_AUDIT_DIR, best-effort, two-week read window, remote-deploy blind spot
src/core/minions/system-prompt.ts header (60) determinism-for-cache requirement, override paths
src/core/minions/handlers/subagent.ts grep only (replay call sites) replay ordering, persist-before-dispatch, stable-id shim
skills/ tree listing, RESOLVER.md lines 1-40, manifest.json head + counts 53 dirs / 51 SKILL.md / 125 files; manifest version 0.32.3.0 listing 50 skills
templates/SOUL.md.template, templates/HEARTBEAT.md.template heads (40 / 30) placeholder slots, cadence beats referencing skill files and gbrain commands
VERSION full 0.42.25.0

Grep-only sweeps (no file substantively read): emotional_weight consumers across src/; last_retrieved_at readers; dream_generated call sites; withCalibration call sites; metaHook wiring; createAuditWriter consumers; eval_contradictions_cache engine methods.

8. Limitations, each paired with the conclusion it prevents

# Limitation Exact conclusion it prevents
1 Nothing was executed (cold trial, read-only checkout). Prevents any observed or causally supported status anywhere in this lens. Every activation and causal-effect finding in §4 is uninspected for this reason alone.
2 src/core/search/hybrid.ts (1,870 lines) was not line-read; only the graph-signals, query-cache, and mode seams. Prevents "the ranking pipeline's full stage ordering and its interaction with emotional_weight, recency decay, autocut, and two-pass is as described." Only the four post-fusion stages named in graph-signals.ts:27-33 are grounded.
3 src/core/operations.ts (4,751 lines) was read only at think, recall, forget_fact, extract_facts. Prevents "the inventory of agent-facing read-back surfaces is complete." Other ops among the ~47 may carry read-back not seen — get_recent_salience, find_anomalies, get_recent_transcripts, find_trajectory, whoknows are all named in doctrine but their handlers are uninspected.
4 The host agent platform (OpenClaw, Claude Code, Codex, …) is outside the checkout. Prevents "the injected _meta, skillpack, SOUL/HEARTBEAT, or live-context block reaches the model's attended context." Everything ends at GBrain's output boundary; the consumption side is unobservable here. This is the packet's own declared boundary character and it bites hardest on the push routes (RTE-07, RTE-17, PROPOSED-RTE-A).
5 gbrain-evals and evals/functional-area-resolver/ runs are outside the boundary. Prevents upgrading any benchmark number (P@5 49.1%, R@5 97.9%, +31.4 P@5, propose_takes 0.952/0.922 F1, resolver +13–17pp) above claimed.
6 defaultEvidenceRetriever is a declared placeholder and both calibration-wave prompt versions self-label -stub; whether any production path injects a real retriever was not traced. Prevents "GBrain's calibration profile reflects a graded forecasting track record." Also prevents concluding the reverse — that no real retriever exists — since injection is a supported opt (opts.evidenceRetriever) and callers were not exhaustively traced.
7 Skillpack install / projection (RTE-17 producer side) was not read. Prevents "the 51 shipped SKILL.md files are what a given install actually loads." The manifest's staleness (50 skills, version 0.32.3.0 vs VERSION 0.42.25.0) makes this gap load-bearing rather than cosmetic.
8 page_versions has a schema definition but no reader was found in the inspected surface. Prevents both "page version history is read back" and "page version history is dead state." The consumer is uninspected, not absent.
9 Config defaults were read from code comments and header docs, not from a running config. Several read-back routes are gated by config keys (think.trajectory_enabled, search.track_retrieval, cycle.skillopt.enabled, cycle.enrich_thin.enabled, cycle.grade_takes.auto_resolve.enabled, facts.extraction_enabled, search.mode). Prevents any claim about what a particular deployment has switched on. Every "default ON/OFF" in this report is a code-declared default, not an observed setting.
10 src/core/minions/handlers/subagent.ts was grep-read, not line-read (1,275 lines). Prevents "crash replay is the only consumer of subagent_messages" — one additional consumer of subagent_tool_executions (the synthesize orchestrator) surfaced from a different file, which suggests there may be others not surfaced.
11 Only two of four templates/*.md.template files were read (USER.md.template, ACCESS_POLICY.md.template not read). Prevents completing the OBJ-15 split — USER.md and ACCESS_POLICY.md remain uninspected as to producer, content shape, and whether either carries binding rather than advisory force.