GBrain agentic-system analysis

Type: types/agentic-system-analysis-result.md

Run identity

Run state: kb/reports/state/agentic-system-analysis/AAS-2026-09-23-gbrain-01/run-state.md

Generated review: kb/agentic-systems/reviews/gbrain-garrytan.md

Memory analysis report: kb/reports/state/agentic-system-analysis/AAS-2026-09-23-gbrain-01/memory-report.md

Memory analysis report SHA-256: 0051622a58b2fc0961f072803d7f78cd3b758f6ff3e78af198b4bf3e6b6e14a1

The caller asked for a refresh of kb/agentic-systems/reviews/gbrain.md. Destination inspection rejected that path as not a generated review of the same source (it is a legacy review), so this run publishes under the qualified slug above. The legacy review's prose was not read.

Boundary and evidence

Intended use: a current, code-grounded account of GBrain as a whole system, including how it delivers its agent-facing parts (plugins, generated skill trees, scaffolded and stub skills, shared brain skills, bootstrap writes, upgrade handling) to host harnesses and projects.

Target class: memory/knowledge/context-engineering system that also ships an agent operating layer: a skill set with a resolver, harness hooks, a durable background job queue with an LLM sub-agent, a nightly maintenance cycle, and a skill optimizer. Boundary kind: complete artifact, partial loop. GBrain supplies memory, retrieval, context delivery, skills and background maintenance. The host harness (Claude Code, Codex, OpenClaw, Hermes, opencode and others) owns the main agent loop, the host model's decisions, skill activation and tool approval.

Functional inclusions: the gbrain CLI and operation contract; stdio and HTTP MCP servers and their dispatcher; persistence coordinator and both engines (PGLite, Postgres); retrieval, think synthesis, and push-context paths; fact, take and page write/curation routes; the dream cycle, autopilot and Minions queue; SkillOpt; the plugin generator, committed plugin trees, release publication of the codex-plugin branch; skillpack scaffold, harness bridge, stub skills and skill serving; shared brain skills and native router; bootstrap and in-agent install writers; self-upgrade, migrations and post-upgrade sweep.

Exclusions and the conclusions they prevent:

• Host harness internals (skill loading, hook execution, context assembly, permission prompts): prevent any claim that a delivered skill, hook output or instruction block was activated or changed host behavior.
• Model providers (Anthropic, Voyage, OpenAI and others): prevent claims about model internals, provider-side versioning of undated aliases, and output quality.
• The closed-source Memorable CLI and its remote extraction API: prevents any claim about procedure extraction, revision scoring or recall quality beyond GBrain's own capture hook and documentation.
• Claude and Codex marketplace services: prevent claims about how and when installed plugins refresh.
• The separate gbrain-evals repository and the author's production deployment: prevent verification of reported benchmark and scale figures.
• Deployed operation: no system was executed; all findings are implementation or doctrine, not observed runs.

Frozen revision: 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29 (v0.54.1.0, commit date 2026-09-23). Analysis cutoff: 2026-09-23. Evidence tier: code-grounded.

Source register

Source ID	Kind	Identity/location	Revision or capture	Evidence layer	Inspected scope	Citation anchors	Access gaps and conclusion prevented
SRC-1	Git	https://github.com/garrytan/gbrain; access root /home/zby/llm/commonplace/related-systems/gbrain	6040075c6cb95be5881cc2e1b76ef7d71f4e5d29	Implementation: TypeScript under src/, scripts/, .github/workflows/, plugin manifests. Doctrine/design: README.md, AGENTS.md, CLAUDE.md, docs/, skills/. No observed-run or causal evidence.	Operation contract and MCP dispatch; persistence, facts, search, think, context and hook code; cycle, Minions, SkillOpt; skillpack, shared-skills, bootstrap, agent-install, upgrade; plugin generation and release workflow; memory-boundary, shared-skills, skillopt and memorable docs	Full commit-relative paths on the canonical records below, all at this revision	Worktree and current HEAD are never evidence. Large parts of src/core (company-brain, connectors, code-intel, google, schema-pack internals, admin SPA) were not read; they are not load-bearing here except where named. No dynamic execution, so no observed operation.

Shared records

Components

CMP-1 — gbrain CLI, src/cli.ts. Implementation conclusion status: wired. Builds CLI commands from operation cliHints plus CLI-only commands, and runs every operation with remote: false, the trusted-local lane. First tries to delegate to a running gbrain serve over a local socket. Writes still require a verified local writer registration (see RTE-22). Source: SRC-1, src/cli.ts:63-85,697-740,1569-1665.

// Local CLI invocation — the user owns the machine; do not apply remote-caller // confinement (e.g., cwd-locked file_upload). remote: false, --- src/cli.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CMP-2 — MCP servers and shared dispatcher: stdio src/mcp/server.ts, OAuth HTTP src/commands/serve-http.ts, legacy bearer HTTP src/mcp/http-transport.ts, all through src/mcp/dispatch.ts. Implementation conclusion status: wired. Every MCP call runs with remote: true. Surfaces (verbs, starter, full) filter the visible operation set; HTTP checks OAuth scopes at list and call time; localOnly operations dispatch only on stdio; --source-guard blocks writes without an explicit source binding. Source: SRC-1, src/mcp/dispatch.ts:499-720, src/mcp/surface.ts:88-156, src/commands/serve-http.ts:2474-2604.

if (op.localOnly && opts.transport !== 'stdio') { return unknownToolEnvelope(name, opts.allowedOps); } --- src/mcp/dispatch.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CMP-3 — Operation contract, src/core/operations.ts and src/core/ops/contract.ts. Implementation conclusion status: wired. One registry of 100+ operations with scope, requiredScopes, localOnly, mutating, verb and a handler; CLI, MCP and --tools-json are generated from it. OperationContext.remote is required. Source: SRC-1, src/core/operations.ts:1-4,126-221, src/core/ops/contract.ts:277-530.

• True when the caller is remote/untrusted (MCP over stdio/HTTP, or any agent-facing entry point).
• False for local CLI invocations by the owner of the machine. --- src/core/ops/contract.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CMP-4 — Brain engine and persistence coordinator: PGLite (embedded Postgres via WASM) or Postgres with pgvector, behind BrainEngine; src/core/persistence/* admits page, fact and skill mutations through a write-request queue, write-through to Markdown files, and a trigger guard for legacy writers. Implementation conclusion status: wired. Symbolic code over an RDBMS plus Markdown files. Source: SRC-1, src/core/persistence/memory-mutations.ts:47-148, src/core/persistence/page-prepare.ts:97-245, src/core/persistence/authority.ts:25-37, src/core/persistence/writer-guard-schema.ts:1.

CMP-5 — AI gateway and distributed-parametric model components, src/core/ai/gateway.ts, src/core/ai/defaults.ts, src/core/model-config.ts. Roles: chat/synthesis (default anthropic:claude-sonnet-4-6; think resolves tier deep with fallback alias opus = anthropic:claude-opus-4-7), query expansion (anthropic:claude-haiku-4-5-20251001), embeddings (voyage:voyage-4 at 1024 dimensions for new installs; legacy zeroentropyai:zembed-1), reranking (voyage:rerank-2.5), fact extraction, LLM judges, SkillOpt optimizer/target/judge tiers. Existence and call wiring: wired. Parameter changes during operation: absent within the searched boundary (ABS-6); GBrain issues inference calls only. Identity pinning: configurable defaults, none hard-pinned; the Haiku default names a dated snapshot, while the Sonnet, Opus, Voyage defaults are provider model names whose resolution is provider-controlled (uninspected). Provider internals: uninspected. Source: SRC-1, src/core/ai/gateway.ts:128-129, src/core/ai/defaults.ts:44-60, src/core/model-config.ts:60-70, src/core/think/index.ts:488-493.

const DEFAULT_EXPANSION_MODEL = 'anthropic:claude-haiku-4-5-20251001'; const DEFAULT_CHAT_MODEL = 'anthropic:claude-sonnet-4-6'; --- src/core/ai/gateway.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

const modelUsed = await resolveModel(engine, { cliFlag: opts.model, configKey: 'models.think', tier: 'deep', fallback: 'opus', // think is the high-stakes synthesis op; opus is the right default }); --- src/core/think/index.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CMP-6 — Dream cycle, autopilot and Minions queue: src/core/cycle.ts and src/core/cycle/* run an ordered phase list (lint, backlinks, sync, synthesize, extract, extract_facts, extract_atoms, patterns, synthesize_concepts, consolidate, propose_takes, grade_takes, calibration_profile, drift, enrich_thin, skillopt, embed, orphans, schema-suggest, purge and others); gbrain dream runs one cycle; gbrain autopilot submits a cycle job per interval; src/core/minions/* provides a durable job queue whose subagent handler is crash-resumable. Implementation conclusion status: wired. Many LLM phases are off by default or gated by config. Source: SRC-1, src/core/cycle.ts:61-202, src/commands/dream.ts:1-24, src/commands/autopilot.ts:1-17, src/core/minions/handlers/subagent.ts:1-24.

• crash-resumable: subagent_messages + subagent_tool_executions together
• are the single source of truth about where the conversation is. On
• resume after a worker kill, we load all committed rows, trust any tool
• execution marked 'complete' or 'failed', and re-run 'pending' ones only --- src/core/minions/handlers/subagent.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CMP-7 — SkillOpt optimizer, src/core/skillopt/*. Implementation conclusion status: wired. Treats a SKILL.md body as the optimized artifact: target-model rollouts on benchmark tasks, two optimizer reflect calls (failures, successes), a judge-scored validation gate, versioned commit. See RTE-11. Source: SRC-1, src/core/skillopt/orchestrator.ts:40-80,540-745.

CMP-8 — Plugin generator and published plugin lanes: scripts/generate-plugin-tree.ts builds the committed plugin/ and plugin-variants/ trees; scripts/check-plugin-tree.sh gates drift; .claude-plugin/, .codex-plugin/, .agents/plugins/marketplace.json, openclaw.plugin.json and the shell launcher .agents/gbrain-launcher declare the lanes; .github/workflows/release.yml force-publishes a history-less codex-plugin branch. Implementation conclusion status: wired. Source: SRC-1, scripts/generate-plugin-tree.ts:17-30,96-318, scripts/check-plugin-tree.sh:36-60, .github/workflows/release.yml:285-354.

CMP-9 — Skillpack scaffold and harness bridge, src/core/skillpack/*, src/commands/skillpack/*. Implementation conclusion status: wired. Copies bundled skills into a workspace or into a harness's native skills directory, optionally as stubs, records install-time hashes for harness installs, and offers a read-only reference lens. Source: SRC-1, src/core/skillpack/scaffold.ts:1-26,97-151, src/core/skillpack/harness-bridge.ts:1-31,283-313,604-730, src/core/skillpack/bridge-state.ts:1-64.

CMP-10 — Shared brain skills: brain-resident skill packs with immutable revisions, publication policy and membership (src/core/shared-skills/*), the harness-side adapter that syncs authorized revisions into a local cache and renders a router skill, and src/core/harness/native-router.ts, which installs that router into the harness skills directory. Implementation conclusion status: wired. Source: SRC-1, src/core/shared-skills/adapter.ts:48-200, src/core/harness/native-router.ts:27-78, src/core/shared-skills/publication.ts:28-32.

CMP-11 — Bootstrap and in-agent install writers, src/core/bootstrap/*, src/core/agent-install/*, templates/bootstrap/*. Implementation conclusion status: wired. Write identity/instruction templates, harness hooks, MCP registrations and a marker-delimited instruction block, each with ownership receipts. Source: SRC-1, src/core/bootstrap/render.ts:1-39, src/core/bootstrap/hooks.ts:1-25, src/core/bootstrap/host-specs.ts:184-239, src/core/bootstrap/instructions-block.ts:1-108, src/core/agent-install/setup.ts:120-298.

CMP-12 — Upgrade machinery: maybeEmitUpdateMarker in src/cli.ts, src/commands/check-update.ts, src/commands/upgrade.ts (binary swap, post-upgrade, migrations, skill reference sweep), src/commands/apply-migrations.ts, src/core/upgrade-checkpoint.ts, and host-agent migration manuals in skills/migrations/. Implementation conclusion status: wired. Source: SRC-1, src/cli.ts:351-412, src/commands/upgrade.ts:198-219,336-468,800-915, src/commands/migrations/index.ts:1-11.

CMP-13 — Host harness and host model (excluded participant). Claude Code, Codex, OpenClaw, Hermes, opencode and others load GBrain skills, call MCP tools, run hooks and own the main loop. Conclusion status for activation of anything GBrain delivers: uninspected. Named so that routes can state where GBrain's responsibility ends.

Operative objects

OBJ-1 — Pages. Markdown files in brain repositories plus DB page rows, chunks and embeddings; frontmatter typed by a schema pack. put_page stamps provenance fields (source_kind, ingested_via, ingested_at). Natural-language content with symbolic metadata; storage: files and RDBMS (with vector columns). Consumers: search, recall, think, context push, dream phases. Source: SRC-1, src/core/persistence/page-prepare.ts:176-236, gbrain.yml.

OBJ-2 — Facts. facts rows plus a page's ## Facts fence; each carries provenance (source), valid_from, optional valid_until (TTL), visibility (world default, or private), embedding, expired_at, superseded_by, and withdrawal state. Natural-language claims with symbolic metadata. Source: SRC-1, src/core/verbs.ts:106-162, src/core/persistence/memory-prepare.ts:28-90, src/core/facts/withdrawal.ts:11-86.

const visibility = typeof p.visibility === 'string' ? p.visibility : 'world'; --- src/core/verbs.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

Amendment (memory lens): the fence is canonical and the cycle reconciles DB rows from it (RTE-9). source_session is not stored in the fence, so a rebuilt row loses its session, and session-keyed hot-memory selection falls back to 24-hour recency (CLM-10). Source: SRC-1, src/core/ops/facts.ts:46, src/core/cycle/extract-facts.ts:1-25.

OBJ-3 — Takes, take proposals, grade cache and calibration profiles. Takes are held positions with conviction; take_proposals hold LLM-proposed takes pending human acceptance; take_grade_cache holds LLM verdicts; calibration_profiles hold Brier/accuracy scorecards plus voice-gated narrative statements and bias tags. Natural language plus numeric scores in RDBMS. Source: SRC-1, src/core/cycle/propose-takes.ts:1-37, src/core/cycle/grade-takes.ts:1-33, src/core/cycle/calibration-profile.ts:1-24.

OBJ-4 — Link and timeline edges. Typed graph edges and timeline entries extracted deterministically from page references on trusted writes, or added through add_link. Symbolic, RDBMS. Source: SRC-1, src/core/persistence/page-prepare.ts:214, docs/guides/memory-boundaries.md.

OBJ-5 — Bundled skill set (static, shipped). skills/*/SKILL.md with frontmatter triggers:, skills/RESOLVER.md, skills/conventions/, skills/manifest.json. Natural-language instructions; files in the repository and the installed package. Not memory: changed only by releases or by local editing of copies. Source: SRC-1, skills/RESOLVER.md, skills/_AGENT_README.md:121-131.

OBJ-6 — Generated plugin trees and manifests: plugin/skills/, plugin-variants/gbrain-coding, plugin-variants/gbrain-daily, each with a generated README version stamp and its own manifests and launcher; lane membership recorded with a reason per addition or exclusion in skills/plugin-lanes.json. Files; natural-language skills plus JSON manifests. Source: SRC-1, scripts/generate-plugin-tree.ts:204-318, skills/plugin-lanes.json, .claude-plugin/plugin.json.

OBJ-7 — Installed skill copies and their receipts. Workspace scaffold copies (no receipt); harness-bridge copies in ~/.claude/skills, ~/.agents/skills or opencode skills, full or stub, with per-file SHA-256 install receipts in ~/.gbrain/skillpack-bridge-state.json (schema gbrain-skillpack-bridge-v1); third-party pack receipts in ~/.gbrain/skillpack-state.json (pinned commit or tarball digest, no per-file hashes). Source: SRC-1, src/core/skillpack/bridge-state.ts:1-64, src/core/skillpack/state.ts:1-57.

• reference lens can run a true three-way comparison: installed-hash vs
• current-file vs current-source separates a LOCAL EDIT (file no longer
• matches what we wrote) from UPSTREAM DRIFT (file untouched, source moved). --- src/core/skillpack/bridge-state.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

OBJ-8 — Shared brain skill packs and the harness-side cache. Brain-resident skillpack.json and skills/ in a source's content root, published as immutable revisions with sealed file hashes and a policy epoch; on the harness side, revisions/ cache files, router/SKILL.md, active.json, receipt.json with owned_files and pending_files hashes, and a native router skill named gbrain-shared-<hash24>. Natural-language skills plus symbolic manifests. Source: SRC-1, src/core/shared-skills/adapter.ts:48-130, src/core/harness/native-router.ts:27-78, docs/guides/shared-brain-skills.md.

OBJ-9 — Host configuration written by bootstrap and in-agent install: workspace identity templates (CLAUDE.md, AGENTS.md, SOUL.md), agent.json, a marker-delimited instruction block in ~/.claude/CLAUDE.md or $CODEX_HOME/AGENTS.md, hook entries keyed by _gbrain markers in Claude settings, MCP registrations, and receipts (bootstrap/receipt.json, harness.json, in-agent owned_files). Natural-language instructions plus JSON/TOML configuration. Source: SRC-1, src/core/bootstrap/instructions-block.ts:32-33, src/core/bootstrap/host-specs.ts:184-239, src/core/bootstrap/format.ts:119-200, src/core/agent-install/setup.ts:120-153.

export const AMBIENT_WRITEBACK_BLOCK_BEGIN = ''; export const AMBIENT_WRITEBACK_BLOCK_END = ''; --- src/core/bootstrap/instructions-block.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

OBJ-10 — SkillOpt artifacts under skills/<name>/skillopt/: history.json (pending/committed rows with the applied edits, selection score and delta), versions/vNNNN_eE_sS.md, best.md, proposed.md, and rejected.json (bounded LRU of rejected edits with reasons, keyed to the skill text they were proposed against); plus the author-supplied benchmark (skillopt-benchmark.jsonl) and optional held-out set. Natural-language skill text; symbolic judges. Source: SRC-1, src/core/skillopt/version-store.ts:1-60, src/core/skillopt/rejected-buffer.ts:1-40, src/core/skillopt/types.ts:36-83,240-262.

export type Judge = | { kind: 'rule'; checks: RuleCheck[] } | { kind: 'llm'; rubric: string; model?: string } | { kind: 'qrels'; expected_slugs: string[]; k: number }; --- src/core/skillopt/types.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

OBJ-11 — Superseded. This record combined derived session state with raw transcript files, which differ in lineage and consumers. It is split into OBJ-16 (session context state) and OBJ-17 (raw transcript corpus). OBJ-11 denotes no current object; see Reconciliation.

OBJ-12 — Derived pages: dream synthesize pages from transcripts, patterns pages, enrich_thin enrichments, and think --save synthesis pages with a ## Gaps section. Natural language; files and RDBMS. Source: SRC-1, src/core/cycle/synthesize.ts:1-48, src/core/cycle/patterns.ts:1-21, src/core/think/index.ts:966-1000.

OBJ-13 — Upgrade state: update-check cache, ~/.gbrain/migrations/completed.jsonl, ~/.gbrain/upgrade-checkpoint.json bound to a database-URL hash, pending-host-work.jsonl, and the natural-language migration manuals in skills/migrations/vX.md addressed to the host agent. Source: SRC-1, src/commands/migrations/index.ts:1-11, src/core/upgrade-checkpoint.ts:1-45.

OBJ-14 — Bootstrap agent-workspace memory files: MEMORY.md (standing rules, open commitments, active context), memory/YYYY-MM-DD.md daily logs, memory/reference/, USER.md, SOUL.md, HEARTBEAT.md, in a workspace Git repository that hooks push to a private remote. The host agent writes them by editing files, as the workspace AGENTS.md instructs; Claude Code loads MEMORY.md through an @MEMORY.md import, and the SessionStart hook pushes a digest of three sections (RTE-5). Natural language; files and repo. The templates are shipped doctrine; the file contents change through use. Implementation conclusion status for loading and digest: wired; for agent writing: claimed (doctrine the host agent follows). Source: SRC-1, templates/bootstrap/AGENTS.md.template:28-44,100-124, templates/bootstrap/CLAUDE.md.template:8-11, templates/bootstrap/memory-README.md.template:1-21, src/commands/hook.ts:118,486-490,603-631.

• Operational state (open commitments, corrections, active context) → MEMORY.md --- templates/bootstrap/AGENTS.md.template @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

OBJ-15 — Stop-hook live buffer, ~/.gbrain/transcripts/live/<session>.txt: one JSON line per turn with a 400-character excerpt of the last assistant message, collected after 7 days. Raw trace. Read back only as the "Last session activity" line at the next SessionStart, from the newest file of any session. Source: SRC-1, src/commands/hook.ts:633-663,1431-1450.

...(exchange ? { exchange: exchange.slice(0, 400) } : {}), --- src/commands/hook.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

OBJ-16 — Session context state: session_context_state rows keyed (source_id, client_id, session_id) holding banked standing entities, the delta wake cursor and a checkpoint link manifest of verified page slugs. Symbolic; RDBMS. Idle cursors are collected after 7 days. Consumers: post-compaction SessionStart re-entry and delta (RTE-5, RTE-26). Source: SRC-1, src/core/context/session-state.ts:1-55, src/core/ops/facts.ts:682-715.

OBJ-17 — Raw transcript corpus, in dream.synthesize.session_corpus_dir or else <config>/transcripts/corpus: session-end transcripts, compaction segments, writeback turn files, a per-session ledger, harvest receipts and claim sidecars, secret-scanned before write and retained 30 days. Raw trace. Inputs to RTE-6, RTE-24 and the synthesize branch of RTE-8. Source: SRC-1, src/commands/hook.ts:1320-1430,1630-1810, src/core/sweep.ts:26-35,582-586.

OBJ-18 — Fact withdrawal ledger, fact_withdrawals, rows keyed (source_id, visibility, fact fingerprint) written by forget. A later remember or extraction of the same normalized claim is refused. Symbolic; RDBMS. Force: enforcement over future writes. Source: SRC-1, src/core/facts/withdrawal.ts:11-40, src/core/facts/single-prepare.ts:19-24.

const inserted = await tx.executeRaw(INSERT INTO fact_withdrawals(source_id,visibility,fact_hash) ---src/core/facts/withdrawal.ts@6040075c6cb95be5881cc2e1b76ef7d71f4e5d29`

OBJ-19 — Access and feedback metadata: pages.last_retrieved_at, updated by recall and used to break entity-card ties, and context_volunteer_events, a log of delivered push items read only by operator statistics and doctor, pruned after 90 days. Symbolic access metadata, not content. Source: SRC-1, src/core/ops/facts.ts:395-420, src/core/verbs/entity-card.ts:10-14, docs/guides/push-context.md.

Routes

RTE-1 — Explicit fact write (remember verb over MCP, gbrain remember locally). Implementation conclusion status: wired. The caller supplies a fact, required provenance, kind, optional TTL and visibility (default world). The coordinator resolves the entity page (or memory/unattributed), queues a write request, rejects a claim previously withdrawn, embeds the fact when a provider exists, and classifies it: exact fingerprint match is duplicate; cosine ≥ 0.95 against a same-entity, same-visibility, unexpired candidate is superseded (same kind) or duplicate; otherwise inserted. It appends to the page's ## Facts fence, re-renders the page and inserts or expires facts rows. Return: status and ID. Revision admission: the caller proposes, the similarity rule decides supersession; no truth check applies. Rollback: withdrawal (RTE-7); an expired row keeps its history. Source: SRC-1, src/core/verbs.ts:60-162, src/core/persistence/memory-mutations.ts:47-94, src/core/persistence/memory-prepare.ts:28-90, src/core/facts/single-prepare.ts:13-55.

if (candidate && score >= 0.95) return { status: candidate.kind === input.kind ? 'superseded' : 'duplicate', candidate, }; --- src/core/facts/single-prepare.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

'provenance is REQUIRED (free text, e.g. "conversation 2026-06-12", "user said in chat", "import: notes.md"). ' + --- src/core/verbs.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

Amendment (memory lens): supersession is near-duplicate replacement, not contradiction handling. A contradicting fact below cosine 0.95 is inserted beside the old one; correcting it depends on an explicit forget or a near-identical restatement. Without an embedding provider, deduplication degrades to fingerprint matching. Affects the curation and epistemic readings of RTE-1 and RTE-9.

RTE-2 — Page write (put_page, capture). Implementation conclusion status: wired. Parses, chunks, stamps provenance, prepares timeline, fact and tag projections, writes the file through write-through, and embeds as a later persistence effect. Graph links are extracted only for trusted callers; a remote MCP put_page stores references as text and reports auto_links.skipped: remote. The fact backstop queues extraction of facts from the written page (RTE-9). Source: SRC-1, src/core/ops/pages.ts:275-304, src/core/persistence/page-prepare.ts:97-245, src/core/persistence/effects.ts:12-17.

const links = !noop && !targetDeleted && ordinaryPage && (row.authority.autoLinkTrusted ?? !row.authority.remote) && await isAutoLinkEnabled(engine) --- src/core/persistence/page-prepare.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-3 — Requested retrieval (recall, search, query, graph and entity reads). Implementation conclusion status: wired. recall gathers facts by entity, time window, session and supersession chains, optionally adds a page-search arm, updates last_retrieved_at, and packs facts first into a token budget. Remote callers see only world-visible facts and non-private pages. search runs keyword plus vector arms with reciprocal-rank fusion (k=60), post-fusion boosts (backlinks, salience, recency, compiled truth, source prefix, supersede down-rank) and a reranker in balanced and tokenmax modes; without an embedding provider it is keyword-only. query additionally expands the question with a model call. Next-step owner: the requesting agent. Source: SRC-1, src/core/ops/facts.ts:180-510, src/core/ops/search.ts:226-400, src/core/search/hybrid.ts:1-10,1226, src/core/search/mode.ts:441-619.

const visibility = ctx.remote === false ? undefined : ['world'] as ('private' | 'world')[]; --- src/core/ops/facts.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

• Pipeline: keyword + vector → RRF fusion → normalize → boost → cosine re-score → dedup *
• RRF score = sum(1 / (60 + rank_in_list)) --- src/core/search/hybrid.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-4 — Cited synthesis (think op, synthesize verb, gbrain think). Implementation conclusion status: wired. Gathers pages by hybrid search, takes by keyword and vector search, and graph neighbours of an anchor; asks the resolved model for JSON {answer, citations, gaps}; flags citations absent from the gathered set as CITATION_NOT_IN_GATHER. With --with-calibration (off by default) it injects the calibration profile and anti-bias rules. Remote callers cannot persist; local --save writes synthesis/<slug>-<date> with a ## Gaps section directly through engine.putPage, and --take writes a take. Gaps are returned, not queued as work; gap-driven follow-up rounds are not implemented. Source: SRC-1, src/core/think/index.ts:9-17,91-100,479-1000, src/core/think/gather.ts:155-211, src/core/think/prompt.ts:40-70, src/core/ops/takes.ts:187-280.

const remote = ctx.remote ?? true; // Codex P1 #7 + privacy: remote callers cannot persist via MCP. const safeSave = remote ? false : Boolean(p.save); --- src/core/ops/takes.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-5 — Automatic per-turn and per-session context push. Implementation conclusion status: wired (GBrain side); host activation: uninspected. Bootstrap installs Claude Code hooks for SessionStart, UserPromptSubmit, Stop, SessionEnd and PreCompact that call gbrain hook <event>. On a user prompt, the hook sends the prompt window over an authenticated local socket to gbrain serve, which assembles reflex pointers, volunteered pages (deterministic entity matching, confidence ≥ 0.7, at most 3 pages) and hot facts under a world-only context, wraps them in a data envelope, and returns up to 10,000 characters as additionalContext. Session start adds a file digest and a context pack; PreCompact banks standing entities. Independently, MCP tool responses can carry _meta.brain_hot_memory (top facts, 30-second TTL). The hook never opens the engine and fails open. Source: SRC-1, src/core/bootstrap/host-specs.ts:184-199, src/commands/hook.ts:1-33,471-530,1226,1291-1330, src/core/context/turn-context.ts:1-24,59-60, src/core/context/volunteer.ts:1-45, src/core/facts/meta-hook.ts:1-20.

export const TURN_CONTEXT_ENVELOPE = ''; --- src/core/context/turn-context.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

Amendment (memory lens; the specialist's SessionStart route is merged here). Selection details:

• User prompt: the selector reads the last four turns, prior injected blocks and the session ID. Entity pointers and volunteered pages come from lexical alias, title, surname and slug-suffix matching, with no embedding or model call. Hot facts are the session's facts (else the last 24 hours) ranked by decayed confidence, top 10. The default block budget is 8 KB, trimmed facts first.
• SessionStart (1.5-second deadline): a digest of three MEMORY.md sections up to 3 KB (OBJ-14), the newest live-buffer line (OBJ-15), status notes, then a context pack of this session's banked standing entities, up to four entity cards, hot facts, the delta since the last wake, and checkpoint links on compaction re-entry (OBJ-16). The wake cursor advances only when the pack is complete.
• _meta.brain_hot_memory is attached to MCP responses on both transports except recall, extract_facts and forget_fact.
• The IPC lane works for Postgres as well as PGLite (CLM-8).

Push paths are world-only. Source: SRC-1, src/core/context/turn-context.ts:180-330, src/core/facts/meta-hook.ts:65-165, src/commands/hook.ts:471-600, src/mcp/context-pack-handler.ts:137-160, src/core/context/volunteer.ts:1-24.

• Zero-LLM, deterministic, precision-biased: push noise is worse than pull --- src/core/context/volunteer.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

if (sessionId) { rows = await ctx.engine.listFactsBySession(sourceId, sessionId, { activeOnly: true, limit: topK, visibility, }); } // If no session-scoped rows, fall back to recent across the source. --- src/core/facts/meta-hook.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

rows.sort((a, b) => effectiveConfidence(b, now) - effectiveConfidence(a, now)); --- src/core/facts/meta-hook.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

export const DIGEST_SECTIONS = ['standing rules', 'open commitments', 'active context']; --- src/commands/hook.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-6 — Ambient writeback. Implementation conclusion status: wired; default off. With memory.auto_writeback set to salient or all, the Stop hook parses the transcript tail, applies a deterministic gate to the last user turn, banks it as a corpus file, and asks the server to schedule a checkpoint flush. The serve-side harvest queue (8 jobs, 60-second abort) runs the LLM fact-extraction pipeline with remote: false, deduplicates and writes facts; the maintenance sweep is the backstop when serve is down. Source: SRC-1, src/commands/hook.ts:1452-1535, src/core/context/checkpoint-harvest.ts:1-27,299-307,437-446, src/core/facts/extract.ts:519-577.

if (!wb.enabled) return 'wb_off'; --- src/commands/hook.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

Amendment (memory lens): the route has two lanes. Writeback is opt-in and, in salient mode, filters to medium-and-up notability. Compaction harvest is separate: PreCompact banks the window as a secret-scanned segment (OBJ-17), and the harvest is gated only by facts.extraction_enabled, default on. Extracted facts carry source hook:writeback or hook:compact and the session ID. The compaction lane also banks the checkpoint link manifest (OBJ-16), which the same conversation receives on post-compaction re-entry. Extraction runs as trusted local, outside the MCP dispatcher. Source: SRC-1, src/core/context/checkpoint-harvest.ts:40-130,284-360,379-465.

r = await runFactsPipeline(raw, { engine: job.engine, sourceId: job.sourceId, sessionId: job.sessionId, source: 'hook:compact', mode: 'inline', remote: false, --- src/core/context/checkpoint-harvest.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-7 — Withdrawal, supersession, TTL and page deletion. Implementation conclusion status: wired. forget/forget_fact strike the fence row, mark the fact withdrawn with a reason and keep history; a later remember of the identical claim fails with fact_withdrawn. Supersession sets expired_at and superseded_by and search down-ranks superseded content. TTL becomes valid_until and reads filter on it. delete_page is a 72-hour soft delete restorable by restore_page; autopilot purge then hard-deletes. Doctrine states that withdrawal is not erasure. Source: SRC-1, src/core/verbs.ts:343-383, src/core/persistence/memory-mutations.ts:96-148, src/core/facts/withdrawal.ts:11-86, src/core/ops/pages.ts:352-410, docs/guides/memory-boundaries.md.

if (decision.status === 'superseded') await tx.executeRaw(UPDATE facts SET expired_at=now(),superseded_by=$3 ---src/core/persistence/memory-prepare.ts@6040075c6cb95be5881cc2e1b76ef7d71f4e5d29`

RTE-8 — Dream-cycle page generation (synthesize, patterns, enrich_thin). Implementation conclusion status: wired; enrich_thin off by default. Trigger: gbrain dream or autopilot. Synthesize: a cheap-model triage selects transcripts; each passing transcript gets an LLM sub-agent (default Sonnet) whose put_page writes the orchestrator holds; the orchestrator then repairs quotes mechanically (exact, normalized or near match to a verbatim transcript span, otherwise strip the quotation marks) and stamps provenance. Patterns: one sub-agent writes or updates a pattern page per theme from recent reflections. Enrich_thin: grounded LLM synthesis from brain context, skipped when context is too thin, written with expected_revision and enriched_by. Admission: auto-applied; no review queue. Rejection: triage and the thin-context gate. Rollback: none general (ABS-3); the synthesize mode dial and page soft delete. Guidance shaping the proposal: fixed prompts in code, not revised by this route. Source: SRC-1, src/core/cycle/synthesize.ts:1-48, src/core/cycle/synthesize-verify.ts:1-30, src/core/cycle/patterns.ts:1-21, src/core/enrich/thin.ts:1-40, src/commands/enrich.ts:438-462.

• Ladder invariant: NEVER fabricate — every replacement is a verbatim slice of
• the source transcript; when nothing grounds, the span loses its quotation
• marks but keeps its text. --- src/core/cycle/synthesize-verify.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

Amendment (memory lens): the branches differ in lineage and gate. Synthesize is trace-fed from the transcript corpus (OBJ-17) and enabled by default only when dream.synthesize.session_corpus_dir is configured, which bootstrap does not do. Its pages carry dream_generated and raw_source, and are excluded from fact re-extraction. Patterns is compiled from reflection pages, not traces. Enrich_thin is off by default. Once enabled, all three are applied without review. Source: SRC-1, src/core/cycle/synthesize.ts:1519-1525,2935-2941.

// v2: enabled defaults to true when corpus dir is configured, false otherwise. --- src/core/cycle/synthesize.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-9 — Page-write fact extraction, fence reconciliation and consolidation (amended after the memory lens; the specialist's page-write backstop route is merged here). Implementation conclusion status: wired. The fact backstop is the single choke point for put_page, sync, file upload and code import. It queues LLM extraction over the page body, using the configured extraction model with Sonnet as fallback. Sync admits only high-notability facts. Dream-generated pages are ineligible (anti-loop guard). In the extraction lanes any matched candidate is treated as a duplicate, so extraction never supersedes. The LLM classifier in src/core/facts/classify.ts (duplicate, supersede or independent) has no non-test importer at this revision (CLM-7). The cycle's extract_facts phase runs no model: it reconciles DB rows from each page's ## Facts fence, which is canonical. consolidate is also deterministic. It takes buckets of at least three facts whose oldest is at least 24 hours old, clusters them greedily at cosine 0.85, and turns each cluster of two or more into a takes(kind='fact') row using the highest-confidence fact's text. The contributing facts are marked, never deleted. Confidence decays by a kind-specific half-life at read time; the decay ranks facts and deletes nothing. Source: SRC-1, src/core/facts/backstop.ts:1-40,261, src/core/facts/extract.ts:1-66, src/core/persistence/facts-prepare.ts:83-120, src/core/cycle/extract-facts.ts:1-25, src/core/cycle/phases/consolidate.ts:1-22,150-260, src/core/facts/decay.ts:1-33.

if (decision.candidate) { entries.push({ fact, duplicateId: decision.candidate.id }); continue; } --- src/core/persistence/facts-prepare.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

• Reconciles the facts DB index from the ## Facts fence on each

• entity page. --- src/core/cycle/extract-facts.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

• 1. For each cluster ≥ 2: pick the highest-confidence fact's text as

• the take claim (v0.31 ships without LLM synthesis to keep the --- src/core/cycle/phases/consolidate.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

• Formula: confidence × exp(-age_days / halflife_days). Clamped to [0, 1]. --- src/core/facts/decay.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-10 — Takes: proposal, grading and calibration. Implementation conclusion status: wired. propose_takes has an LLM write rows to a review queue only; a human accepts with gbrain takes propose --accept N. grade_takes collects evidence by hybrid search and asks an LLM judge for a verdict, cached with applied=false; auto-resolution is off by default and, when enabled, applies at confidence ≥ 0.95 with a three-model ensemble for the 0.6–0.95 band. calibration_profile computes Brier and accuracy scorecards plus narrative statements gated by a Haiku voice judge. Later consumers: think --with-calibration (off by default), stderr nudges on take commit, the brainstorm orchestrator. An undo removes auto-applied resolutions by wave_version. Source: SRC-1, src/core/cycle/propose-takes.ts:1-37, src/core/cycle/grade-takes.ts:1-33,513-533,772-785, src/core/cycle/calibration-profile.ts:1-24, src/core/think/index.ts:91-100,557-576, src/core/calibration/undo-wave.ts:1-30, src/core/calibration/nudge.ts:1-27.

• propose_takes only WRITES proposals to the queue. Nothing here mutates --- src/core/cycle/propose-takes.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

• On a fresh install, grade_takes runs and writes verdicts to the cache,

• but applied=false on every row. Operator reviews the queue, then flips --- src/core/cycle/grade-takes.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-11 — SkillOpt skill revision. Implementation conclusion status: wired. Trigger: gbrain skillopt <skill> (human-invoked), the admin-scoped MCP operation run_skillopt (requires a remote skill allowlist), or the dream-cycle skillopt phase (off by default; weekly per skill; $0.50 per-skill cap; no-mutate for every skill, bundled or not, despite a cycle.ts comment that limits this to bundled skills, CLM-9). Proposal: the target model runs the skill on training tasks; two optimizer calls read the current skill body, failure or success trajectories, the plain-English scoring criteria and the rejected-edit buffer, and propose at most eight surgical add/replace/delete edits, each with a one-sentence reason and each required to address a specific observed failure; frontmatter edits are refused. Decision: a judge scores each selection task three times; the candidate is accepted only if the median selection score exceeds the current best by more than 0.05; an optional held-out set refuses candidates that regress. Answer oracle: the benchmark, supplied by the skill author (or bootstrapped by the optimizer model and marked BOOTSTRAP_PENDING_REVIEW until a human removes the marker); judges are rule checks, LLM rubrics, or qrels expected page slugs. Admission: accepted text is committed in five ordered steps ending with SKILL.md; with --no-mutate, in the dream path, or for a bundled skill without --allow-mutate-bundled plus a held-out set, the candidate goes to proposed.md for human review. For a shared brain skill the accepted text is published through put_skill against the approved revision, after an editor authorization check. Rejection: rejected edits enter rejected.json with the gate reason and are withheld from later proposals against the same skill text. Rollback: crash recovery of a pending version only; no rollback command (ABS-2); older versions remain on disk. Rollouts run with brain read tools only. Source: SRC-1, src/core/skillopt/reflect.ts:1-120, src/core/skillopt/validate-gate.ts:160-190, src/core/skillopt/version-store.ts:1-60,207-245, src/core/skillopt/cycle-phase.ts:1-20,145-170, src/core/skillopt/rejected-buffer.ts:1-40, src/core/skillopt/rollout.ts:36-37, src/core/skillopt/benchmark.ts:79-85, src/core/shared-skills/optimizer.ts:1-158.

// D12 accept rule: strict > best + epsilon. Ties/sub-epsilon-gains are // rejected (paper-faithful — protects against noise-as-improvement). const threshold = opts.bestScore + VALIDATION_EPSILON; const accepted = selScore > threshold; --- src/core/skillopt/validate-gate.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

• Each edit MUST address a SPECIFIC failure pattern you observed.
• anchor / target MUST be uniquely identifiable in the skill body (exact match).
• Do NOT propose edits already in the rejected-edit history — those were tried and didn't help. --- src/core/skillopt/reflect.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

noMutate: true, // ALL dream-cycle runs are no-mutate by default --- src/core/skillopt/cycle-phase.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

const publication = await submitSharedSkillMutation(ctx, 'put_skill', { ...target, name: opts.skillName, files: input, --- src/core/shared-skills/optimizer.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

Retained reasons: each edit's one-sentence reason is kept in history.json with its selection score and delta, which only crash recovery reads later; SKILL.md itself carries no reasons. Rejected edits and their gate reasons are read by the next reflect call. Source: SRC-1, src/core/skillopt/types.ts:69-72,240-249, src/core/ops/skillopt.ts:127-130.

description: 'Run SkillOpt against a single skill. Admin scope and remote skill allowlist required. --- src/core/ops/skillopt.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-12 — Durable sub-agent (gbrain agent run, Minions subagent jobs). Implementation conclusion status: wired. Local trusted submission only: shell, subagent, subagent_aggregator and the synthesize, patterns and consolidate jobs are protected names MCP cannot submit. The sub-agent receives brain read tools plus put_page and add_timeline_entry fenced to wiki/agents/<subagentId>/ or a trusted-workspace allow-list, and each tool call runs with remote: true. Messages and tool executions persist in two phases so a killed worker resumes. Source: SRC-1, src/commands/agent.ts:1-13, src/core/minions/protected-names.ts:15-30, src/core/minions/tools/brain-allowlist.ts:14-18,51-78,242.

remote: true, // match MCP trust boundary for auto-link skip --- src/core/minions/tools/brain-allowlist.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-13 — Plugin build, publication and host install. Implementation conclusion status: wired (repository side); marketplace refresh and host loading: uninspected. The generator computes each lane as the OpenClaw skill list minus recorded base exclusions plus recorded additions, snapshots skills whose declared MCP operations exceed the starter surface, copies skills and shared dependencies, writes a stamped README, and emits self-contained persona variants with their own manifests and launcher. check-plugin-tree.sh regenerates into a temporary directory and fails on any byte difference; it runs in bun run verify and before release publication. The release job copies the manifests, launcher and plugin trees into a fresh repository and force-pushes one history-less commit to the codex-plugin branch. The Claude plugin's MCP server runs the launcher as serve --surface starter --source-guard; the launcher resolves an already installed gbrain binary and never installs one. Neither plugin manifest declares hooks (ABS-5). No code compares an installed plugin tree's stamp with the binary (ABS-1); freshness depends on the host marketplace refresh. Source: SRC-1, .claude-plugin/plugin.json, .claude-plugin/marketplace.json, .codex-plugin/plugin.json:21-22, .codex-plugin/mcp.json, .agents/gbrain-launcher:18-51, scripts/generate-plugin-tree.ts:17-30,96-318, scripts/check-plugin-tree.sh:36-60, .github/workflows/release.yml:285-354.

"skills": "./plugin/skills/", "mcpServers": { "gbrain": { "command": "${CLAUDE_PLUGIN_ROOT}/.agents/gbrain-launcher", "args": [ "serve", "--surface", "starter", "--source-guard" ], --- .claude-plugin/plugin.json @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

if ! diff -r "$TMP/plugin" "$ROOT/plugin" >/dev/null 2>&1; then echo "check-plugin-tree: FAILED — committed plugin/ tree is stale." >&2 --- scripts/check-plugin-tree.sh @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

git init -q -b codex-plugin --- .github/workflows/release.yml @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-14 — Workspace skillpack scaffold and reference lens. Implementation conclusion status: wired. gbrain skillpack scaffold copies selected skills, paired source files and shared dependencies into a workspace, never overwriting an existing file and recording no receipt. skillpack reference is read-only and reports identical, differs or missing with unified diffs; --apply-clean-hunks applies only hunks that apply cleanly; skillpack sync scaffolds only new skills and reports drifted ones. Without install hashes this lane cannot separate a local edit from upstream drift. The legacy managed AGENTS.md block installer and skillpack install/uninstall were retired. Third-party packs (skillpack.json, api_version: gbrain-skillpack-v1) pass a trust prompt and record a pinned commit or tarball digest. Source: SRC-1, src/core/skillpack/scaffold.ts:1-26,97-151, src/core/skillpack/reference.ts:1-14, src/commands/skillpack.ts:103-191,362-387, src/core/skillpack/skill-currency.ts:1-36, src/core/skillpack/scaffold-third-party.ts:1-16, src/core/skillpack/state.ts:1-57.

• 1. Refuses to overwrite existing files. Once a file lands, the
• user owns it. To update, run gbrain skillpack reference <name>
• and decide. --- src/core/skillpack/scaffold.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-15 — Harness bridge install, three-way update lens, and stub skills. Implementation conclusion status: wired; host loading of the copies: uninspected. scaffold --harness claude-code|codex|opencode|openclaw copies a persona's skills plus their shared-dependency closure into the harness's native skills directory after a frontmatter gate and target confinement, refusing to overwrite, and records a SHA-256 for every file it wrote. The reference lens compares installed hash, current file and current source: a file still matching its install hash is upstream_drift, anything else local_edit, and a file without a recorded hash unknown. Applying updates is permitted only for proven upstream drift and refreshes the stored hash; stubs, local edits and unknown files are refused. Removal deletes only ledger-recorded files. In --stub mode the frontmatter (name, description, triggers) is kept verbatim for discovery and the body is replaced by a marker line and an instruction to call the MCP get_skill operation, with gbrain skill <slug> as a local fallback; preflight refuses stub mode when skill publication is disabled or the skill cannot be served. get_skill is gated by mcp.publish_skills and serves from the brain host's configured or autodetected skills directory (remote callers never fall back to the binary's bundled skills), from a brain-resident pack, or from the shared catalog when active. The stub body is therefore as current as the serving host's skill source, not necessarily the upstream release. Source: SRC-1, src/commands/skillpack/harness.ts:1-16,271-348, src/core/skillpack/harness-bridge.ts:1-31,78,283-313,604-730,897-930, src/core/ops/skills-catalog.ts:23-99, src/core/skill-catalog.ts:155-210, src/mcp/skill-resources.ts:11-81.

export const STUB_MARKER = ''; --- src/core/skillpack/harness-bridge.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

This is a cold-pull stub installed by \gbrain skillpack scaffold`. The full\n+skill body is served by your gbrain brain, so it is always current.\n\n+To use this skill: call the gbrain MCP tool `get_skill` with\n+ ---src/core/skillpack/harness-bridge.ts@6040075c6cb95be5881cc2e1b76ef7d71f4e5d29`

const installedHash = stateEntry?.written[slug ?? SHARED_DEP_LEDGER_KEY]?.files[rel]; const differsKind: HarnessRefDiffersKind = installedHash ? sha256Hex(actual) === installedHash ? 'upstream_drift' : 'local_edit' : 'unknown'; --- src/core/skillpack/harness-bridge.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

• REMOTE callers use autoDetectSkillsDir (no install_path tier — a hosted
• gbrain must never serve its own bundled dev skills); LOCAL callers --- src/core/skill-catalog.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-16 — Shared brain skills: publication, following and native router. Implementation conclusion status: wired (brain and adapter); native activation: uninspected and declared unverified by the source. Canonical skills live in a brain source's content root. An authorized editor (skill_editor scope plus explicit put_skill access) publishes a complete file set against expected_revision with a request ID; a race returns a conflict. Human or Git edits to canonical files are proposals published through the local-only import_skill_proposal with expected hashes. Publication policy, audiences and following are separate skill_publisher grants; ordinary revisions within an approved policy can follow without per-install consent, while new permissions, file classes or requirements cannot. On the harness side the adapter calls sync_brain_skills, fetches each immutable revision and asset, checks hashes against the manifest, caches them, renders a router skill that instructs the agent to sync and fetch before using a shared skill, and refuses to overwrite any locally edited managed file (local_conflict). The native router installs as gbrain-shared-<hash24>/SKILL.md and refuses unowned or edited routers. A staged migration (inventory, ownership, projection, policy, members) moves existing installs; old harness scaffolds over an active shared brain return a pending plan instead of copying bodies. Source: SRC-1, src/core/shared-skills/adapter.ts:48-200, src/core/harness/native-router.ts:27-78, src/core/shared-skills/knowledge-guard.ts:27, src/core/shared-skills/migration.ts:12-40, src/core/skillpack/shared-brain-bridge.ts:87-105, docs/guides/shared-brain-skills.md.

if (current !== null && current !== receipt.owned_files[path] && current !== pending?.before && current !== pending?.after) fail('local_conflict', 'A managed artifact was edited; preserve it and resolve the native shadow copy before following updates.'); --- src/core/shared-skills/adapter.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

This paragraph is advisory, not an enforced invocation hook. Start a fresh session after updates; native activation is unverified.\n; ---src/core/shared-skills/adapter.ts@6040075c6cb95be5881cc2e1b76ef7d71f4e5d29`

The publisher validates policy and file hashes, publishes the canonical bundle, and seals its catalog projection. A race returns a conflict instead of last-write-wins. --- docs/guides/shared-brain-skills.md @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-17 — Bootstrap and in-agent install writes. Implementation conclusion status: wired; host activation: uninspected. gbrain bootstrap renders identity and instruction templates into a new workspace without overwriting (--force backs up), writes agent.json, and records created_paths and registrations in a machine-local receipt that uninstall reverses; a newer-version receipt refuses the run. Claude hooks are merged into settings by _gbrain marker keys (bootstrap-v1, bootstrap-harness-v1), atomically with a backup, aborting on unparsable settings. gbrain bootstrap harness registers an HTTP MCP server with a bearer token (Claude), a managed TOML block (Codex) or a managed JSONC entry (opencode), with a write-ahead receipt and mint-before-revoke token rotation. A marker-delimited ambient-writeback instruction block is spliced into the user-global CLAUDE.md or AGENTS.md; reruns replace only its interior and damaged markers throw. In-agent installs (Grok Bot, Muse) write a launcher, setup script, embedded router instruction and maintenance instructions through hash-receipted ownership; a modified owned file stops the rerun, and a different version requires --upgrade. Source: SRC-1, src/core/bootstrap/render.ts:1-39, src/core/bootstrap/format.ts:1-200, src/core/bootstrap/hooks.ts:1-25, src/core/bootstrap/host-specs.ts:184-239,441-443, src/core/bootstrap/harness.ts:1-41, src/core/bootstrap/instructions-block.ts:1-108, src/core/agent-install/setup.ts:120-298.

if (actual !== receipt.owned_files[path] && actual !== pending?.before && actual !== pending?.after) { throw new AgentInstallError('modified_owned_file', Preserve your changes at ${target} before rerunning setup; it will not be overwritten.); --- src/core/agent-install/setup.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-18 — Upgrade. Implementation conclusion status: wired. Every CLI call reads a cached latest version (the VERSION file on master) and, when newer than the running binary and stderr is not a TTY, prints a machine marker UPGRADE_AVAILABLE <cur> <latest> for the gbrain-upgrade skill; a stale cache triggers a detached refresh. gbrain upgrade swaps the binary and runs post-upgrade, which always applies pending migrations (TypeScript registry, resumable ledger, upgrade checkpoint bound to the database URL hash), prints feature notes, and sweeps the detected workspace's scaffolded skills, listing new and drifted skills without overwriting. Natural-language migration manuals in skills/migrations/ are read by the host agent when host work is pending. Scaffolded copies change only when the user or agent acts; stubs read current text from the serving host; shared-skill copies follow through RTE-16. Source: SRC-1, src/cli.ts:351-412, src/commands/check-update.ts:48-50, src/commands/upgrade.ts:198-219,336-468,800-915, src/commands/migrations/index.ts:1-11, src/commands/apply-migrations.ts:1-13, src/core/upgrade-checkpoint.ts:1-45, skills/_AGENT_README.md:121-131.

process.stderr.write(UPGRADE_AVAILABLE ${VERSION} ${latest}\n); --- src/cli.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

'New skills → gbrain skillpack sync. Drifted skills → gbrain skillpack reference <slug> (local edits are yours; nothing is overwritten).\nSee skills/_AGENT_README.md for what your agent should do on update.\nSkip this sweep: GBRAIN_SKIP_REFERENCE_SWEEP=1.', --- src/commands/upgrade.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

For independent scaffolded copies, the user runs gbrain upgrade. Those skill files DO NOT change automatically. --- skills/_AGENT_README.md @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-19 — Markdown repository sync. Implementation conclusion status: wired. gbrain sync admits changed repository files page by page through managed sync with per-page source authority and remote: false; git deletions become soft deletes. Edits to repository files therefore enter the brain without MCP scope, surface or visibility checks, while remember, put_page and delete_page write through to the files. Source: SRC-1, src/commands/sync.ts:653-660,1383, src/core/persistence/sync-run.ts:143-160.

RTE-20 — compile-context files. Implementation conclusion status: wired. Deterministic selection of pages by pinned prefixes, tags and recency within a source, scored with source boosts, filtered by a sensitivity scan and world visibility, written to .claude/gbrain-context.md, an AGENTS.md block or an OpenClaw file. No model call. Later consumer: the host harness loading those files (uninspected). Source: SRC-1, src/commands/compile-context.ts:1-27, src/core/context/compile-view.ts:1-25,262.

RTE-21 — Proposal-only maintenance: contradiction probe, advisor, schema suggestion, drift. Implementation conclusion status: wired. The suspected-contradiction probe pairs pages by hybrid search, asks an LLM judge, stores a run row and prints paste-ready supersession commands that it never runs. The advisor runs deterministic collectors and applies a finding only on --apply <id> with a TTY confirmation and a gbrain argv free of shell metacharacters. Schema suggestion writes candidates to an audit file; review-candidates --apply writes a delta the user must merge. Drift is report-only. Decider in all four: a human. Source: SRC-1, src/core/eval-contradictions/runner.ts:1-25, src/core/eval-contradictions/auto-supersession.ts:5-6, src/core/advisor/apply.ts:1-40, src/commands/advisor.ts:122, src/core/cycle/schema-suggest.ts:9, src/core/schema-pack/review.ts:66-90, src/core/cycle.ts:71-74.

• a paste-ready CLI command. The probe NEVER auto-applies; the user runs
• the command themselves. The proposal is descriptive, not directive. --- src/core/eval-contradictions/auto-supersession.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-22 — Operation authorization and trust lanes. Implementation conclusion status: wired. Remote MCP callers pass surface filtering, localOnly denial off stdio, source guard, parameter validation, a missing-source refusal, and required-scope and bound-client checks (HTTP: OAuth scopes at list and call time). Writes from local lanes require a verified local writer whose lane (cli or stdio) matches the transport. A database trigger guard covers inventoried legacy writers when enabled; manual SQL is outside the protocol. gbrain call dispatches any operation with remote: false. Guarantee strength: protocol, for the MCP and CLI paths it covers; alternate paths are direct DB access, Markdown sync and trusted local calls. Source: SRC-1, src/mcp/dispatch.ts:499-720, src/core/persistence/authority.ts:25-37, src/core/persistence/writer-guard-schema.ts:1, src/commands/call.ts:10-18, src/mcp/server.ts:455-468.

const verified = currentVerifiedLocalWriter() ?? await verifyLocalWriter(ctx.engine, await readLocalWriter(ctx.engine, ctx.remote === false ? 'cli' : 'stdio')); if (verified.remote !== (ctx.remote !== false)) deny('The local trust lane does not match this transport.'); --- src/core/persistence/authority.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

/* Defense in depth for inventoried legacy writers. Manual SQL is outside the protocol. / --- src/core/persistence/writer-guard-schema.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-23 — Brain health remediation. Implementation conclusion status: wired. gbrain doctor --remediate turns health recommendations into a dependency-ordered plan of queued jobs with a cost cap and a recheck of the health score after each step; autopilot forces a full cycle under stated thresholds (time since last run, plan size, estimated duration, score below 70). Checks are deterministic; the synthesize, patterns and consolidate phase handlers are protected from MCP submission. Source: SRC-1, src/core/remediation/run.ts:1-40, src/commands/autopilot-remediation-policy.ts:1-35, AGENTS.md.

RTE-24 — Session-end capture and sweep extraction. Implementation conclusion status: wired. The SessionEnd hook (and Codex session-end capture) writes the redacted transcript remainder to the corpus (OBJ-17). Pass 3 of the serve maintenance sweep runs the LLM fact pipeline over unprocessed corpus files, spend-gated and exactly once through sidecar files. The resulting facts reach later sessions through recall, entity cards, packs and push. This lane is also the fallback for compaction segments when serve is down. Source: SRC-1, src/core/sweep.ts:1-45,582-586,733, src/commands/hook.ts:1630-1810.

• 1. CORPUS INGEST [CX-P0.1] — LLM-backed, spend-gated. Unprocessed --- src/core/sweep.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-25 — Agent-authored workspace memory writes. The host agent edits MEMORY.md and the daily logs (OBJ-14) in the same turn, as the bootstrap AGENTS.md instructs; corrections become standing rules with the incident attached. GBrain supplies the doctrine and the loading, not a writer. Conclusion status for the writing: claimed (doctrine); for the loading: wired. Source: SRC-1, templates/bootstrap/AGENTS.md.template:28-44, templates/bootstrap/memory-README.md.template:1-21.

1. Corrections become standing rules in MEMORY.md, same turn, with the incident attached. --- templates/bootstrap/memory-README.md.template @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

RTE-26 — Requested context_pack and delta. Implementation conclusion status: wired. These are two of the seven memory verbs, which the harness calls at session start, after compaction or on a heartbeat. context_pack packs requested entities to a token budget. delta returns changes since a per-(source, client, session) cursor with at-least-once delivery. include_private is honoured only for trusted local callers. No model call. Source: SRC-1, src/core/ops/facts.ts:512-715, src/core/context/session-state.ts:1-20.

Claims

CLM-1 — Autonomous overnight improvement. The README claims the author's deployment ingests, enriches, fixes its own citations and consolidates memory overnight, so that "I wake up smarter than when I went to bed". Layer: doctrine and reported operation; production scale figures (155,795 pages, 66 cron jobs) are reported, not inspectable. Supporting routes: RTE-8, RTE-9, RTE-10, RTE-23 (wired). The claimed effect on the user or agent: uninspected. Source: SRC-1, README.md.

It fixes its own citations and consolidates memory overnight. I wake up smarter than when I went to bed — and so will you. --- README.md @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CLM-2 — SkillOpt keeps only edits that measurably score higher. Layer: doctrine. Supporting route RTE-11 implements the rule against the benchmark oracle (wired); no run in this boundary shows a measured gain.

watch the optimizer propose edits and keep only the ones that measurably score higher. --- README.md @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CLM-3 — Graph retrieval benchmark. A 240-page BrainBench run reported P@5 49.1%, R@5 97.9% and +31.4 points P@5 over a graph-disabled variant, explicitly scoped. Layer: reported operation; scorecards live in a separate repository (excluded). Conclusion status: claimed.

CLM-4 — Stub skills are "always current". Layer: implementation string in the stub body (RTE-15). Supported reading: the body is fetched at use time from the serving brain host's skill source; currency relative to upstream releases depends on that host's skills directory or shared catalog. Conclusion status: afforded, with that qualification.

CLM-5 — Memorable revisions: a new procedure revision gets a trial window and "the one with the better track record wins". Layer: doctrine about a closed-source external component. Conclusion status: claimed; scoring uninspected.

surfaces whichever revision the evidence favours — a new one gets a short trial window, then the one with the better track record wins. --- docs/memorable-agents.md @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CLM-6 — Sharing boundary. The README states that remote clients are constrained by source and operation grants plus visibility filters, that local files and shared database credentials are a different trust boundary, and that tests exercise specific paths, not a universal no-leak guarantee. Layer: doctrine, consistent with RTE-22 and RTE-19.

The authorization tests exercise specific access paths, not a universal no-leak guarantee. --- README.md @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CLM-7 — The extract_facts operation description advertises a "cosine fast-path + classifier dedup pipeline". Layer: implementation string (operation description). A commit-wide git grep for classifyAgainstCandidates finds only its definition in src/core/facts/classify.ts and test/facts-classify.test.ts. Classifier conclusion status: absent from runtime routes; the description overstates the pipeline. Consequence: contradictions below the 0.95 threshold accumulate (RTE-1, RTE-9).

calls the configured extraction model (key-aware: any servable provider — OpenAI or Anthropic key both work), runs the cosine fast-path + classifier dedup pipeline, INSERTs into facts. --- src/core/ops/facts.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CLM-8 — The push-context guide says the hook lane is quiet on Postgres, but the code keys an IPC socket for Postgres from a hash of the database URL. Layer: doctrine versus implementation; the guide is stale. Implementation conclusion status for Postgres hook IPC: wired.

The hook lane rides the PGLite serve's IPC socket: on a Postgres brain or a --- docs/guides/push-context.md @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

if (cfg.engine === 'postgres' && cfg.database_url) { return localIpcSocketPath(join(ipcRunDir(), ${kind}-${hash12(cfg.database_url)}.sock)); --- src/core/context/resolve-ipc.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CLM-9 — A comment says the dream cycle never auto-mutates bundled skills; the implementation passes noMutate: true for every skill (RTE-11). Layer: implementation comment versus code. The comment's scope is narrower than the behavior.

// Bundled-skill safety: dream-cycle NEVER auto-mutates bundled skills. // For bundled skills we set --no-mutate; the user reviews proposed.md // at their own cadence. --- src/core/skillopt/cycle-phase.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

CLM-10 — Session targeting of hot memory silently degrades for fence-backed facts: a rebuilt fence row has no source_session, so selection falls back to 24-hour recency (OBJ-2, RTE-5). Layer: implementation inference from the fence schema and the selector. Conclusion status: wired.

Evidenced absences

ABS-1 — No runtime freshness check for installed plugin trees. Conclusion status: absent. Searched: git grep -n 'plugin-tree-stamp' over src/ at the reviewed commit; no match. Supports: the stamp is written only into generated READMEs and checked in CI; an installed plugin's currency depends on the host marketplace. Does not exclude a host-side check.

ABS-2 — No SkillOpt rollback command. Conclusion status: absent. Searched: git grep -n -iE 'rollback|restore|revert' in src/commands/skillopt.ts; no match. Supports: reverting an accepted skill version is manual (old versions stay in versions/); crash recovery covers only a pending write.

ABS-3 — No general dream-cycle rollback. Conclusion status: absent. Searched: git grep -n -iE 'rollback|undo|revert' in src/core/cycle.ts, src/commands/dream.ts, src/commands/autopilot.ts; no match. Supports: auto-applied phase writes (RTE-8) are undone only through ordinary page operations; src/core/calibration/undo-wave.ts is a narrower undo for auto-applied take resolutions.

ABS-4 — auto_think is not a cycle phase. Conclusion status: absent. Searched: git grep -n "auto_think\|'auto-think'\|auto-think" in src/core/cycle.ts; no match. Supports: configured recurring think questions do not run in the cycle at this revision. The memory lens independently reached the same finding from the phase's own "not yet wired" comment (src/core/cycle/auto-think.ts:24-28).

ABS-5 — No hooks in plugin manifests. Conclusion status: absent. Searched: git grep -n -E '"hooks"|hooks' in .claude-plugin/plugin.json, .codex-plugin/plugin.json, .codex-plugin/mcp.json; no match. Supports: plugin installs deliver skills and an MCP server only; per-turn push (RTE-5) requires bootstrap.

ABS-6 — No model-weight updates issued by GBrain. Conclusion status: absent. Searched: git grep -n -iE 'fine[-_ ]?tun|finetun|training_file|lora' over src/; the only implementation hit is a comment naming the provider's fine-tune identifier format in src/core/ai/recipes/openai.ts. Supports: CMP-5 parameter fixity at GBrain's boundary; does not exclude provider-side model changes.

ABS-7 — No retained test of dependence on recalled memory. Conclusion status: absent. Searched (memory lens): the evals/ directory listing, and docs/ and CHANGELOG.md for ablation and A/B evidence. Retrieval benchmarks (BrainBench P@5 and R@5, docs/architecture/RETRIEVAL.md) measure ranking, not dependence. The think A/B table, SkillOpt receipts and volunteer statistics would be deployment-local and are not in the source. Supports: faithfulness_tested is no within this boundary. Does not exclude evidence held in a deployment.

Behavioral-authority paths

BAP-1 — Delivered skills and resolver to the host agent. Consumer: host model (CMP-13). Channel: harness skill loading from plugin, scaffolded, bridge or stub copies (OBJ-5, OBJ-6, OBJ-7), or on-demand get_skill text. Force: instruction (advisory to the host; the resolver says the frontmatter triggers: are the authoritative routing signal). Horizon: every session in which the host loads the skill, until the copy changes. Activation: uninspected.

BAP-2 — Pushed per-turn context. Consumer: host model. Channel: hook additionalContext and MCP _meta.brain_hot_memory (RTE-5). Force: knowledge (the envelope labels it data, not instructions). Horizon: the current turn.

BAP-3 — User-global instruction block and bootstrap identity files. Consumer: host agent in every session reading the user-global CLAUDE.md/AGENTS.md or the workspace files (OBJ-9). Channel: always-loaded instruction files. Force: instruction. Horizon: persistent until removed.

BAP-4 — SkillOpt-accepted skill versions. Consumer: later executions of that skill, locally or, for shared skills, by every following installation through RTE-16 (OBJ-10, OBJ-8). Channel: SKILL.md. Force: instruction. Horizon: until the next accepted version or manual revert. Also covers followed shared-skill revisions from manual publication. The reasons for each edit stay in history.json, which no later consumer reads; SKILL.md does not carry them.

BAP-5 — Calibration profile in synthesis. Consumer: think --with-calibration and take-commit nudges (OBJ-3). Channel: prompt injection and stderr. Force: instruction on the framing of the question, with the scorecard (Brier score, per-domain accuracy) delivered as the reason. Horizon: per call; off by default, and reachable only through the CLI flag; the MCP think operation exposes no such flag.

\nCalibration-aware mode (v0.36.1.0): the user's calibration profile is included as <calibration> below the retrieval blocks. Apply it to the QUESTION FRAMING, not the evidence:, --- src/core/think/prompt.ts @ 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29

BAP-6 — Operation grants and surfaces. Consumer: remote MCP callers. Channel: dispatcher checks (RTE-22). Force: enforcement. Horizon: per call, within the covered paths.

BAP-7 — Supersession, withdrawal, TTL and decay flags. Consumer: recall, search and push selection (OBJ-2). Channel: eligibility filters and ranking. Force: ranking and eligibility. Horizon: until the fact is re-asserted or history is purged. The withdrawal ledger (OBJ-18) additionally refuses re-assertion of a withdrawn claim (enforcement).

BAP-8 — MEMORY.md standing rules. Consumer: host model in every main Claude Code session of a bootstrapped workspace. Channel: @MEMORY.md import plus the SessionStart digest (OBJ-14, RTE-5). Force: instruction. Horizon: until the agent or user edits the rule. The incident that motivated a rule is attached by doctrine and delivered with it.

Runtime account

Ordinary invocation (plugin lane, Claude Code). The user installs the gbrain plugin from the repository marketplace. The host loads skills from plugin/skills/ and starts the launcher, which execs an already installed gbrain serve --surface starter --source-guard over stdio (RTE-13). A user message triggers the host model (principal: the user; next-step owner: the host model). The host may select a GBrain skill by its triggers and call MCP tools. A remember call enters CMP-2 with remote: true; the dispatcher filters by surface, blocks writes when the source is ambiguous under --source-guard, validates parameters and builds the context; the persistence coordinator admits the write, embeds when a provider exists, decides insert, duplicate or supersede by similarity, writes the fence, file and rows, and returns a status (RTE-1, RTE-22). A later recall or search returns world-visible facts and pages ranked by the hybrid pipeline (RTE-3). State persists in the database and Markdown files; the host's own context is lost at session end unless the host or GBrain hooks write it back. Terminal output is the host's reply. Stdio serve also runs best-effort startup and idle maintenance sweeps; HTTP serve does not.

Material alternate paths.

• Bootstrap lane: hooks push context every turn and at session start (RTE-5); compaction harvest and session-end capture extract facts from transcripts, and ambient writeback does so per turn when opted in (RTE-6, RTE-24). Plugin installs have no hooks (ABS-5).
• Trusted local lane: the CLI, gbrain call, serve-side harvest and Markdown sync run with remote: false, so world-only filters, save refusals and auto-link skipping do not apply (RTE-19, RTE-22). The MCP guarantees cover only remote MCP calls.
• Direct database access: outside the protocol by the source's own statement.
• Thin client: forwards operations to a remote server, which enforces its own grants; not a bypass.
• Background lane: dream cycle and autopilot (RTE-8, RTE-9, RTE-10, RTE-23) and durable sub-agents (RTE-12) act without a user turn.
• Provider-native tools and host shell access belong to the host and are outside GBrain's enforcement.

Forcing cases traced.

1. Remote write into a multi-source brain from a plugin serve. The source guard refuses the write with source_binding_required unless GBRAIN_SOURCE binds a source; reads continue. Guarantee owner: CMP-2 dispatcher; enforcement point: before validation; strength: protocol; covered path: MCP served with --source-guard; alternate paths: CLI and sync. Implementation conclusion status: wired. Source: SRC-1, src/mcp/dispatch.ts:550-590.
2. Upstream release changes a skill that a user edited locally. Workspace copies: reference reports differs without distinguishing the edit (RTE-14). Harness copies: install hashes classify local_edit, and apply refuses it (RTE-15). Shared-skill cache: the adapter raises local_conflict and stops following (RTE-16). Plugin copies: replaced by the host marketplace refresh (uninspected). Implementation conclusion status for the three GBrain-side cases: wired.
3. Nightly SkillOpt on a skill with a benchmark. The phase runs one epoch under a cost cap with noMutate: true; an accepted candidate lands in proposed.md, not SKILL.md, and a human decides. Implementation conclusion status: wired (RTE-11).
4. Stub skill used where the MCP surface is too narrow. The stub instructs the host to ask for a wider surface or use gbrain skill <slug>; the scaffold preflight refuses stub mode when publication is disabled. Implementation conclusion status: wired (RTE-15); host behavior: uninspected.

Decision roles on admitting routes. Explicit memory writes: the caller proposes; a similarity rule decides supersession; the user can withdraw (RTE-1, RTE-7). Dream page generation: an LLM sub-agent proposes, code repairs quotes and auto-admits; no human veto before admission (RTE-8). Takes: an LLM proposes, a human accepts; grade auto-resolution is opt-in (RTE-10). SkillOpt: the optimizer model proposes; the judge and the numeric gate decide; a human triggers the run and, on the nightly and bundled paths, decides adoption; for shared skills an authorized editor's grant is required (RTE-11, RTE-16). Maintenance proposals: humans decide (RTE-21). Remediation: deterministic checks decide; autopilot executes (RTE-23).

Answer oracle. Only SkillOpt uses one at runtime: the author-supplied benchmark (rule checks, rubric, or expected slugs), in the bounded-experiment mode of a SkillOpt run. propose_takes reports prompt tuning against a labelled corpus offline (doctrine in the source comment), with no runtime oracle. Take grading, contradiction probing and takes-quality evaluation use model judges without expected answers.

Improvement triggers and operating modes. Open requests (ordinary agent use) change memory. Scheduled cycles and autopilot run maintenance and derivation. SkillOpt runs are bounded experiments over a benchmark, human-triggered or scheduled as proposal-only.

Execution preflight. No dynamic check planned. Checks considered: running gbrain skillopt on a bundled skill (would need API keys and spend, and a single run would not establish improvement); running the plugin generator to compare trees (CI already encodes that check, and static reading of the script and workflow was sufficient for the delivery findings); exercising scaffold --harness --stub in a scratch home (would show file writes already legible from the code, while host activation, the unresolved question, would remain uninspected). Static evidence was sufficient for every conclusion drawn; the unresolved questions concern host activation and measured benefit, which none of these checks would settle.

Lens scoping

Memory/context scope

Trigger evidence: OBJ-1, OBJ-2, OBJ-3, OBJ-16, OBJ-17, OBJ-12, RTE-1, RTE-3, RTE-5, RTE-6. Inspected boundary: retained material accumulated or changed through use and its write, curation and read-back routes at SRC-1, including brain-resident shared skills and SkillOpt versions; statically shipped skills and plugin trees are delivery artifacts. Pointed-to records: CMP-4, CMP-5, CMP-6, CMP-7, OBJ-1, OBJ-2, OBJ-3, OBJ-4, OBJ-8, OBJ-10, OBJ-11, OBJ-12, RTE-1, RTE-2, RTE-3, RTE-4, RTE-5, RTE-6, RTE-7, RTE-8, RTE-9, RTE-10, RTE-11, RTE-16, RTE-19, RTE-20. Depth: full. Rationale: memory is GBrain's primary claimed work, with several automatic write and push routes. Executed by a fresh specialist from the frozen input.

Epistemic scope

Trigger evidence: CLM-1, CLM-2, RTE-8, RTE-10, RTE-11, RTE-21. Inspected boundary: routes that produce or change truth-apt content, check or dispose candidates, or adapt behavior from evaluation at SRC-1. Pointed-to records: OBJ-2, OBJ-3, OBJ-10, OBJ-12, RTE-1, RTE-4, RTE-7, RTE-8, RTE-9, RTE-10, RTE-11, RTE-21. Classify-only: RTE-3, RTE-5, RTE-13, RTE-14, RTE-15, RTE-16, RTE-17, RTE-18 (transport, delivery and retrieval without content change). Depth: full, compact. Rationale: GBrain makes synthesis, consolidation and self-improvement claims, and has several admission rules with different evaluators. Executed locally in overlay mode.

Lens outputs

Memory/context lens

Executed by a fresh specialist from the frozen input memory-input.md (SHA-256 26c2ad92635b4da79478f177aa1c085436ebe6e2274a01010a0118410e046402), under method SHA-256 7e86ed242caadc095d7f20ccd7fcbcb837b9d62e94fca50656b782c65cb0d675. The report records report-status: complete and worker model claude-opus-5-5[1m]. The integrated findings below use canonical IDs, and their supporting quotes sit on those records.

Inventoried. Pages (OBJ-1), facts and their fences (OBJ-2), takes and calibration (OBJ-3), graph edges (OBJ-4), shared skill revisions (OBJ-8), SkillOpt stores (OBJ-10), derived pages (OBJ-12), workspace memory files (OBJ-14), the Stop live buffer (OBJ-15), session context state (OBJ-16), the transcript corpus (OBJ-17), the withdrawal ledger (OBJ-18) and access metadata (OBJ-19). Routes: explicit writes (RTE-1, RTE-2, RTE-19, RTE-25), trace-fed writes (RTE-6, RTE-24, the synthesize branch of RTE-8, RTE-11), automatic operations over retained material (RTE-9, RTE-10, patterns and enrich_thin in RTE-8), maintenance (RTE-7), pull (RTE-3, RTE-4, RTE-26) and push (RTE-5, RTE-20). Static bundled skills, plugin trees and bootstrap templates are excluded as delivery artifacts. Memorable's procedure extraction is excluded.

Findings.

1. Facts are the unit most memory routes produce and push, and the Markdown fence is canonical. Every fact lane runs one pipeline that ends in a fence row plus a DB row. Provenance is required but is free text that nothing verifies (RTE-1, RTE-9).
2. Curation handles near duplicates, not contradictions. remember supersedes only a same-entity, same-visibility, same-kind candidate at cosine ≥ 0.95. Extraction lanes treat any match as a duplicate. The LLM contradiction classifier is not wired (CLM-7). Consolidation copies the highest-confidence text of a cluster into a take and makes no model call (RTE-9).
3. Push is deterministic and labelled as data. Per-turn and session-start context is chosen by lexical entity matching, session identity and recency, never by an embedding or model call. It is world-visible only and wrapped as "data, not instructions" (RTE-5).
4. Compaction is a memory boundary. PreCompact banks the window. The serve process extracts facts from it as trusted local and banks link manifests and standing entities. The same conversation receives them on re-entry (RTE-6, OBJ-16, OBJ-17). This is the only route whose task horizon is established as within-task.
5. A second, agent-maintained tier lives outside the database. In a bootstrapped workspace the host agent is told to keep standing rules, commitments and active context in MEMORY.md, which the harness imports every session and the SessionStart hook digests (OBJ-14, RTE-25, BAP-8). GBrain supplies only the doctrine and the loading.
6. Trace learning is wired. Automatic trace-fed writes produce retained artifacts that later sessions receive. Facts come from compaction segments, writeback turns and session-end transcripts (RTE-6, RTE-24). Dream synthesize writes pages from transcripts, opt-in by corpus configuration (RTE-8). SkillOpt writes skill versions from rollout trajectories (RTE-11).
7. Instruction authority reaches the host model through three retained paths only: MEMORY.md standing rules (BAP-8), SkillOpt and shared skill revisions (BAP-4), and the calibration block (BAP-5). Everything else arrives as knowledge.

Retained reasons. MEMORY.md rules carry their incident, which is delivered every session. Facts carry provenance and context, which are delivered in packs. SkillOpt keeps per-edit reasons in history.json, which no later route reads. Rejected-edit reasons are read by the next reflection. Dream pages keep a pointer to the source transcript, not a rationale.

Comparison profile. The frontmatter memory-comparison block is the specialist's proposed profile with records mapped to canonical IDs; every axis is known at basis wired. faithfulness_tested is "no" (ABS-7). read_back_signal is coarse, identifier and inferred-lexical; inferred-embedding is excluded because the push selectors make no vector call.

Conclusions prevented.

• Activation: whether host harnesses honour additionalContext, _meta and @ imports.
• Any observed or causal status.
• Whether the OpenClaw context-engine lane (src/core/context-engine.ts, read only up to its facts call) selects by embedding. If it does, read_back_signal would gain inferred-embedding for that branch.
• The details of emotional-weight derivation, the pack-gated extract_atoms and synthesize_concepts phases, and connector ingestion.

Epistemic lens

1. Source-and-claim boundary

GBrain at SRC-1 (6040075c6cb95be5881cc2e1b76ef7d71f4e5d29); declared scope as in Boundary and evidence, excluding host harnesses, providers, Memorable and the evals repository. Question: which routes produce, check, accept, retain and use truth-apt content or adapt behavior. Assessed route families: explicit memory writes, fact extraction and curation, derived-page generation, synthesis, takes, contradiction probing, SkillOpt. Unassessed: company-brain, connectors, code-intel, Google open-loop extraction, schema-pack typing internals; these prevent any whole-system completeness claim about knowledge production. Missing evidence: no observed runs, so every observed candidate state is no instance observed. Claims: CLM-1, CLM-2, CLM-3, CLM-5.

2. Epistemic-object inventory

Object	Candidate truth-apt content	Producer → consumer	Evidence	Gap
OBJ-2 facts	Claims about entities, with provenance string	caller or LLM extractor → recall, push, think	SRC-1 src/core/persistence/memory-prepare.ts	Provenance is caller-supplied text; not verified against a source
OBJ-12 derived pages	LLM synthesis over transcripts, reflections, brain context	dream sub-agents, think --save → search, think	SRC-1 src/core/cycle/synthesize.ts	Quote repair checks quoted spans, not the synthesis claims
OBJ-3 takes, grades, profiles	Held positions with conviction; verdicts; accuracy statistics	LLM proposer and judge, human → think, nudges	SRC-1 src/core/cycle/grade-takes.ts	Grade verdicts rest on retrieved brain evidence only
OBJ-10 skill versions	Non-truth-apt procedure text; benchmark judges carry expected outputs	optimizer model → skill executions	SRC-1 src/core/skillopt/types.ts	Benchmark representativeness uninspected

3. Authority-route ledger

Route	Function	Architectural status	Content/update relation	Check target → evaluator	Result and force	Epistemic authority	Operational authority	Behavioral-authority path
RTE-1	retention and disposition	implemented	truth-apt transformation: acquisition/import	new fact vs same-entity facts → embedding similarity ≥ 0.95	insert, duplicate or supersede; enforcing on recall eligibility	none beyond caller's assertion; supersession is similarity, not truth	changes which fact recall returns	BAP-7
RTE-9	content transformation and dedup	implemented	truth-apt transformation: indeterminate (LLM extraction may paraphrase or infer)	turn or page → LLM extractor; duplicates by fingerprint and cosine (the LLM classifier is unwired, CLM-7)	automatic retention	preserved source link, unverified extraction	changes recall and push	BAP-7
RTE-8	content transformation	implemented	ampliative conjecture	quoted spans → verbatim transcript match	auto-admitted pages; quotes repaired or unquoted	quote fidelity only	pages enter retrieval	BAP-2 via push, requested reads
RTE-4	content transformation, check	implemented	ampliative conjecture	citations → membership in gathered set	answer returned; persisted only on local save	citation presence, not support	none unless saved	requested read
RTE-10 (propose)	disposition/acceptance	implemented	ampliative conjecture	proposed take → human	accept or leave queued	human judgment, criterion unspecified	take becomes active	BAP-5
RTE-10 (grade)	check/evidence production	implemented	no content change until applied	take → LLM judge over retrieved evidence	cached verdict; resolution only when opted in	verdict relative to brain evidence	none by default	BAP-5
RTE-21	check/evidence production	implemented	no content change	page pairs → LLM judge; health → collectors	report and paste-ready commands	suspected contradiction only	none; human applies	none automatic
RTE-11	behavior/policy adaptation	implemented	non-truth-apt policy/content update: SKILL.md procedure revision	candidate skill → judge against author benchmark on selection tasks	accept if median > best + 0.05; instruction force on later runs	benchmark performance within its tasks	replaces or proposes skill text	BAP-4

4. Per-object lifecycle disposition

OBJ-12 (RTE-8, RTE-4): ampliative conjecture. Observation: transcripts and reflections as input, implemented, no instance observed. Conjecture: LLM sub-agent synthesis, implemented, no instance observed. Derived consequence: no route found within boundary. Test/evidence: quote-span verification against the source transcript, implemented, no instance observed; it tests quotation fidelity, not the synthesized claims. Acceptance: no evidence-consuming acceptance transition; pages are admitted on write (no route found within boundary). Lifecycle integration: not reached as post-acceptance integration; retention and retrieval use happen before any acceptance.

OBJ-3 takes (RTE-10): ampliative conjecture. Conjecture: LLM proposal, implemented, no instance observed. Acceptance: human accept, implemented; evaluator human, criterion unspecified, intended use later reasoning and calibration; no instance observed. Test/evidence: grade_takes judges against retrieved evidence, implemented, no instance observed; auto-applied only when enabled. Integration: calibration profile and think --with-calibration, implemented, no instance observed.

OBJ-2 facts (RTE-1, RTE-9): explicit facts are acquisition/import, discovery lifecycle not applicable; warrant is the caller's provenance string, preserved as text and unverified. LLM-extracted facts are indeterminate: classifications still possible are non-ampliative reshaping or ampliative inference; lineage to the source page or turn is preserved; evidence needed is a fidelity check between extracted fact and source text, not found.

No lifecycle record for OBJ-10: no candidate truth-apt output for this object; relevant direct-adaptation routes: RTE-11.

5. System-claim versus route comparison

Claim	Layer	Implemented routes	Observed or causal support	Supported conclusion
CLM-1	doctrine, reported operation	RTE-8, RTE-9, RTE-10, RTE-23	none in boundary	Overnight citation repair, consolidation and enrichment routes exist; "smarter" is unsupported at this boundary
CLM-2	doctrine	RTE-11	none in boundary	The acceptance rule matches the claim within the benchmark; measured gains are not observed
CLM-3	reported operation	RTE-3	external repository excluded	Claimed only
CLM-5	doctrine about external component	none inspectable	none	Claimed only

6. Bounded conclusion

GBrain retains, retrieves and reshapes large volumes of content and generates ampliative content automatically (RTE-8, RTE-9) and on request (RTE-4). Its checks are fidelity and membership checks: quote spans against transcripts, citations against the gathered set, and similarity thresholds for duplicates and near-duplicate supersession. Contradiction detection between facts is not wired (CLM-7), and consolidation is deterministic. None of them tests the truth of the synthesized claims. The only human acceptance gates on truth-apt content are take proposals and the proposal-only maintenance routes. The one route with an external expected-answer oracle is SkillOpt, and it adapts procedure text rather than producing truth-apt claims. Imported facts keep caller-supplied provenance as unverified text. The overnight "smarter" claim remains unsupported at this boundary for lack of observed or causal evidence.

Reconciliation

Specialist proposal-ID mappings.

Proposal	Disposition	Canonical ID
MEM-OBJ-1	registered	OBJ-14
MEM-OBJ-2	registered	OBJ-15
MEM-OBJ-3	registered as one part of the OBJ-11 split	OBJ-17
MEM-OBJ-4	registered	OBJ-18
MEM-OBJ-5	registered	OBJ-19
MEM-RTE-1	merged: SessionStart push was already part of RTE-5	RTE-5
MEM-RTE-2	registered	RTE-24
MEM-RTE-3	merged: the page-write backstop was already part of RTE-9	RTE-9
MEM-RTE-4	registered	RTE-25
MEM-RTE-5	registered	RTE-26
MEM-CLM-1	registered	CLM-7
MEM-CLM-2	registered	CLM-8
MEM-CLM-3	registered	CLM-9
MEM-CLM-4	registered	CLM-10
MEM-ABS-1	merged, independent convergence	ABS-4
MEM-ABS-2	registered	ABS-7
MEM-BAP-1	registered	BAP-8
MEM-BAP-2	merged	BAP-4
MEM-BAP-3	merged	BAP-5

Profile references to OBJ-11 in the report denote session context state and were mapped to OBJ-16.

Amendments.

• OBJ-11 split. Superseded value: one record for session state and corpus files. Replacement: OBJ-16 (session context state) and OBJ-17 (raw transcript corpus); OBJ-11 is marked superseded. Evidence: src/core/context/session-state.ts, src/core/sweep.ts. Affected findings: storage and lineage axes, RTE-6.
• RTE-9. Superseded value: "a Haiku classifier decides duplicate, supersede or independent" and a cycle extract_facts that extracts. Replacement: the classifier has no non-test importer, extraction lanes never supersede, and the cycle phase reconciles DB rows from fences without a model. Evidence: the coordinator ran git grep classifyAgainstCandidates, which finds only src/core/facts/classify.ts and test/facts-classify.test.ts, and read src/core/cycle/extract-facts.ts. Affected findings: the epistemic ledger row for RTE-9, CLM-7, the curation axis, and the Synthesis mechanism (a).
• RTE-1. Superseded value: supersession described without its limits. Replacement: near-duplicate replacement only, with contradictions inserted beside existing facts. Evidence: src/core/facts/single-prepare.ts. Affected findings: epistemic bounded conclusion.
• RTE-8. Superseded value: synthesize treated as generally active. Replacement: synthesize is opt-in through corpus configuration, and branches are distinguished by lineage. Evidence: the coordinator confirmed src/core/cycle/synthesize.ts:1522 (session_corpus_dir). Affected findings: the lineage and trace-learning axes, Synthesis.
• RTE-11. Superseded value: nightly no-mutate stated without the MCP trigger. Replacement: adds the admin-scoped run_skillopt (the coordinator confirmed src/core/ops/skillopt.ts:127-128) and CLM-9. Affected findings: decision roles.
• RTE-5. Superseded value: session-start contents summarized. Replacement: detailed selector inputs and the Postgres IPC lane (CLM-8). Affected findings: the read-back signal axis.
• OBJ-2 amended with the session loss on fence rebuild (CLM-10).
• BAP-4 amended for shared-skill revisions and reason retention.
• BAP-5 amended from knowledge to instruction on question framing, reachable only through the CLI flag.

Independent convergence. ABS-4 (auto_think not wired) was found by the coordinator's grep of src/core/cycle.ts and separately by the specialist from the phase's own comment. The no-mutate behavior of the nightly SkillOpt path was found by both, from the same line.

Integration-issue dispositions. The specialist listed twelve issues. Items 1–8 were adopted as the amendments above. Items 9–11 were adopted through the mappings above. Item 12 (activation of _meta, @ imports and additionalContext) was retained as the Limitations row on the host harness. No conflict required return to the specialist.

Ownership check. RTE-5 and RTE-9 now carry merged specialist routes and keep their generic runtime identity. Shared routes RTE-11 and RTE-16 are owned by the runtime account, with memory classifications only in the profile.

Cross-lens ownership: RTE-11 appears in both lenses. The epistemic lens classifies it as a behavior/policy adaptation; the memory lens owns its trace-learning classification. Neither lens changes the canonical route record. RTE-16 is a delivery route in the runtime account; the memory lens treats shared skills as retained material only because use (editor publication, optimization) changes them.

Scout reports used during the runtime baseline were source-only locators; every load-bearing finding above was re-read or quote-checked against SRC-1 by the coordinator. One scout excerpt of src/mcp/surface.ts contained a scout-inserted ellipsis and was not quoted.

Bounded synthesis

Evidence basis and boundary. Static reading of the GBrain repository at 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29 (v0.54.1.0), cutoff 2026-09-23. The boundary is a complete artifact inside a partial loop: GBrain does not own the agent loop, so every claim about what a host does with delivered skills, hooks or context stays uninspected.

Architectural characterization and claimed work. GBrain is a memory and context service for agents that already exist. It stores pages and provenance-carrying facts in PGLite or Postgres with Markdown files as the canonical form for file-backed pages, exposes one operation contract through a CLI and MCP servers, and adds three layers around that core: retrieval and cited synthesis, automatic context push through harness hooks, and background maintenance through a nightly cycle and a durable job queue. It also ships a large skill set and treats the distribution of that skill set as a product surface in its own right.

Runtime map. Explicit writes (RTE-1, RTE-2) and trusted imports (RTE-19) fill the store; requested reads (RTE-3, RTE-4) and pushed context (RTE-5, RTE-20) return it. Compaction harvest (on by default in the bootstrap lane), opt-in ambient writeback (RTE-6), session-end sweep extraction (RTE-24) and the page-write backstop (RTE-9) extract facts automatically. The dream cycle derives pages and takes (RTE-8, RTE-10) and curates facts (RTE-9). Authorization (RTE-22) separates remote MCP callers from trusted local lanes; the local lanes, Markdown sync and direct SQL fall outside the remote-caller guarantees, which the source documents.

Delivery mechanisms (the caller's focus). GBrain ships its agent-facing parts through six mechanisms with different update semantics:

1. Plugins (RTE-13): one generated skill tree per lane, persona variants as self-contained plugin roots, byte-exact drift gating in CI, and a history-less codex-plugin branch rebuilt on every release. Every lane addition or exclusion carries a written reason. The plugin supplies skills and a launcher-started MCP server, but no hooks, and nothing checks an installed plugin's version at runtime.
2. Workspace scaffold (RTE-14): copy once, never overwrite, then compare. Without install hashes the lens reports differences but cannot tell an edit from upstream change.
3. Harness bridge (RTE-15): the same copy with per-file install hashes, giving a three-way classification that allows safe application of upstream drift and refuses local edits.
4. Stub skills (RTE-15): frontmatter stays local for discovery; the body is fetched at use time through get_skill or gbrain skill. Currency is relative to the serving host's skill source.
5. Shared brain skills (RTE-16): skills become versioned brain content with compare-and-set publication, separate editor and publisher grants, hash-checked delivery into a harness cache, and a router skill that tells the agent to sync before selecting. The source itself labels router activation as advisory and unverified.
6. Bootstrap and in-agent installs (RTE-17): marker-delimited instruction blocks, marker-keyed hook entries, and receipts recording owned files and hashes so that reinstall, upgrade and uninstall touch only what GBrain wrote.

Upgrade (RTE-18) connects them: a machine marker announces new versions to an upgrade skill, post-upgrade applies migrations, and a sweep lists new and drifted scaffolded skills without overwriting. Across all six, the consistent rules are: never overwrite a file the user may have edited, record ownership by hash where GBrain wrote the file, and treat activation in the host as unverified.

Memory. Facts with required but unverified provenance are the main retained unit. The Markdown fence is canonical and the database mirrors it. Automatic fact capture runs from compaction segments, opt-in writeback turns, session-end transcripts and page writes (RTE-6, RTE-24, RTE-9). These are wired trace-learning routes whose extracted facts reach later sessions through requested reads and deterministic push. A bootstrapped workspace adds an agent-edited MEMORY.md tier that the harness loads as instructions every session (OBJ-14, BAP-8).

Discriminating mechanisms. (a) Supersession by embedding similarity rather than any truth check. It covers only near duplicates: contradicting facts below cosine 0.95 accumulate side by side, because the LLM contradiction classifier is not wired (RTE-1, RTE-9, CLM-7). (b) Automatic ampliative page generation, admitted after quote-span repair, with no acceptance step (RTE-8). (c) A world-only view for remote callers across recall, push and synthesis (RTE-3, RTE-4, RTE-5). (d) SkillOpt's benchmark-gated revision of instruction text, which can propagate through shared-skill publication to every following installation (RTE-11, RTE-16).

Scenario-relative assessment. For a single user adding memory to an existing coding agent, the plugin or MCP path gives requested memory with provenance, correction and withdrawal, and requires no model API for keyword recall. Push context and automatic capture need the bootstrap lane and opt-in configuration. For a team brain, remote grants and visibility filters constrain MCP clients; shared local files and credentials do not isolate agents, as the source states. For skill distribution across many harnesses, shared brain skills provide versioned, authorized and hash-checked delivery; the plugin lane provides the simplest install but no GBrain-side currency check.

Learning, reflection and self-improvement. The strongest supported contribution is SkillOpt (RTE-11), and its components hold at different strengths:

• Formulated theory. A skill's text states a procedure for a task. Formulation status: wired.
• Operative use. The target model runs that text on benchmark tasks. Operative-use status: wired.
• Content-directed criticism. The optimizer reads the text, failure trajectories and the stated scoring criteria, and must tie each edit to an observed failure with a stated reason. Criticism status: wired, though the optimizer's reasoning itself is inaccessible model processing. The benchmark tests stated consequences of the skill.
• Revision. Edits are accepted only when they clear a numeric margin over the current best on selection tasks. Revision status: wired.
• Retention. Rejected edits and their gate reasons persist and are read by later reflection against the same skill text. Accepted edits persist with their reasons in history.json, which no later route reads except crash recovery.
• Improved capacity. Attributable improvement is what the gate tests, but no run within this boundary shows a gain, so the improved-capacity claim is uninspected. Addressability is high over one skill body: edits target exact anchors and frontmatter is fixed.

Whether criticism of the operative skill text improved the system's future capacity therefore remains unestablished for want of run evidence.

Reflection holds at two selected aspects, each wired:

• Doctor and remediation represent brain health; the represented state changes with the brain, and plans mediated by that representation change later state (RTE-23).
• SkillOpt represents a skill's performance as benchmark scores, and those scores gate the text that governs later runs (RTE-11).

No route was found that revises GBrain's own theory-building organization; the reflective-builder claim is not established.

Self-improvement at the declared boundary is wired but gated by humans:

• Local SkillOpt runs commit accepted versions.
• The nightly path only proposes.
• Bundled skills need a held-out set and an explicit flag.
• Shared-skill publication needs editor authority.

The dream cycle's automatic page generation and fact curation change the knowledge the system serves; it does not revise the procedures that produce it. Wired trace learning (facts from session traces) is a memory property; it does not establish conjectural learning. The extracted facts are not theories whose criticism is recorded, and no route tests whether recalling them improves later work (ABS-7).

What would change this assessment. Retained SkillOpt run records with selection and held-out scores would move improved capacity from uninspected toward observed. Host-side traces of skill selection and hook context would settle activation. A check of synthesized claims against sources, rather than of quoted spans, would change the epistemic reading of RTE-8. A runtime comparison of installed plugin versions would close ABS-1.

Limitations

Limitation	Affected IDs	Inspected boundary	Conclusion prevented	Evidence that would resolve it
No dynamic execution	SRC-1, all RTE-*	Static source at the reviewed commit	Any observed or causally supported status	Instrumented runs of the named routes
Host harness excluded	CMP-13, BAP-1, BAP-2, BAP-3, RTE-5, RTE-13, RTE-15, RTE-16, RTE-17	GBrain side of delivery and hooks	Activation of skills, hooks and instruction blocks	Host transcripts showing skill selection and injected context
Provider internals excluded	CMP-5	Gateway call sites and defaults	Model version resolution for undated aliases; output quality	Provider version logs per call
Memorable implementation not in repository	CLM-5	GBrain hook, consent gate and docs	Procedure extraction and revision scoring	Memorable source or run records
Evals repository and production deployment excluded	CLM-1, CLM-3	README claims	Benchmark and scale verification	gbrain-evals scorecards at a pinned revision
Unread subsystems	SRC-1	company-brain, connectors, code-intel, Google loops, schema-pack internals, admin SPA not read	Completeness of the write-route and epistemic inventories	Targeted reading of those subsystems
SkillOpt outcomes not observed	RTE-11, CLM-2, OBJ-10	Implementation and docs	Whether any skill improved	Retained history.json, receipts and held-out scores from real runs
Marketplace refresh behavior external	RTE-13, ABS-1	Release workflow and manifests	How quickly installed plugins track releases	Host marketplace documentation or observation

Verification and blockers

Semantic verification

Checked after reconciliation against the result type's comparison rules:

• Profile scope versus records. The profile scope names the same objects and routes as OBJ-1, OBJ-2, OBJ-3, OBJ-4, OBJ-8, OBJ-10, OBJ-12, OBJ-14, OBJ-15, OBJ-16, OBJ-17, OBJ-18, OBJ-19 and the listed routes. Static skills, plugin trees and templates (OBJ-5, OBJ-6, OBJ-9) are excluded explicitly. Every axis record resolves to a declared canonical ID. Superseded OBJ-11 is not referenced by any axis.
• Trace-fed writes against the learning criterion. Compaction harvest and writeback (RTE-6), session-end sweep (RTE-24), the synthesize branch of RTE-8, and SkillOpt (RTE-11) each have automatic trace-fed production, retention and a named later consumer. Their source, scope, timing and form are carried into the trace-learning axes. RTE-9 backstop extraction works over retained pages, not traces, and is excluded from trace learning.
• Push signals. Each push signal has its trigger, selector input and selected part named on RTE-5 and RTE-20. The identifier signal rests on session-ID selection of facts, banked entities and checkpoint links; a catalog entry alone is not counted. Requested reads (RTE-3, RTE-4, RTE-26) are counted as pull only.
• Amendments. Every amendment stays attached to its original canonical ID and referent. The OBJ-11 split uses new IDs.
• Overlays. The epistemic overlay cites canonical IDs without redefining them.
• Specialist corrections. The coordinator independently verified four: the unwired classifier (grep), synthesize gating, the cycle extract_facts phase, and run_skillopt.
• Load-bearing guarantees. Each names owner, enforcement point, strength and alternate paths (RTE-22, forcing case 1).
• Quote occurrence. Every quote block was checked for whitespace-normalized occurrence in its pinned blob with a local script before publication.
• Assessment. No known assessment is unsupported by its records. Remaining uncertainty is scoped in Limitations.

Deterministic validation

commonplace-validate --full kb/reports/state/agentic-system-analysis/AAS-2026-09-23-gbrain-01/result.md was run on the final bytes before publication and passed.

Blockers

none