Evidence packet — RUN-GBRAIN-20260820

Prepared once by the orchestrator under step 2.4 of analyse-agentic-system. Lens workers consume this packet plus targeted reads inside the frozen boundary. No source reacquisition, no boundary widening, no revision change, no parallel ID namespace.

ERRATUM — added after lens return; the text below is left exactly as the workers received it. The memory/context lens inspected the checkout and found three assertions in this packet to be wrong. They are corrected centrally in result.md §4.4, not here, so that this file remains a faithful record of what was frozen and dispatched. Do not rely on the following without reading that correction: 1. "43 skills" (§3 boundary table, SRC-16, CMP-16) — the tree holds 53 directories and 51 SKILL.md files. Four incompatible counts exist across the repo; preserved as conflict C1. 2. CMP-16's skillpack description — skills/manifest.json declares 50 skills at version: 0.32.3.0 while VERSION is 0.42.25.0; the manifest is stale relative to the tree. 3. Table anchoring — facts, takes, and query_cache are defined only inside the MIGRATIONS array of src/core/migrate.ts, not in src/schema.sql. An inventory keyed on schema.sql alone misses the three most memory-relevant tables in the system.

1. Run identity

• Run/result ID: RUN-GBRAIN-20260820
• Target system: GBrain (github.com/garrytan/gbrain)
• Staging identity: kb/work/analyse-agentic-system/trials/gbrain/
• Analysis cutoff: 2026-08-20 (inspection date); frozen revision below.

2. Frozen revision

• Checkout: /home/zby/llm/commonplace/related-systems/gbrain
• Revision: 9a0bae8d62cdd1e0dd6655e24e082fe6c69c5dac
• Commit date: 2026-06-03
• Commit subject: v0.42.25.0 fix(pricing): unify chat-model pricing into one canonical source; add Opus 4.8 (#1819) (#1827)
• VERSION file: 0.42.25.0
• Working tree: clean (git status --porcelain empty)
• Inspection mode: read-only. No fetch, pull, or refresh.

3. Declared boundary

Included by function — components whose scheduling, context selection, retained state, action execution, checking, acceptance, or authority decisions produce or constrain GBrain's deployed behavior:

Function	Included material
Scheduling	src/core/minions/{queue,worker,supervisor,child-worker-supervisor,rate-leases,quiet-hours,lock-renewal-tick}.ts; src/core/cycle.ts phase orchestration
Model-call loop	src/core/minions/handlers/subagent.ts, src/core/minions/system-prompt.ts, src/core/minions/tools/brain-allowlist.ts, src/core/ai/gateway.ts
Context assembly	src/core/search/* (hybrid, mode, token-budget, query-cache, graph-signals, rerank, expansion), src/core/think/{gather,prompt,index}.ts, src/core/context-engine.ts
Retained state	src/core/facts/*, takes tables + src/core/cycle/{propose-takes,grade-takes,calibration-profile}.ts, src/core/cycle/phases/consolidate.ts, src/schema.sql, src/core/migrate.ts, brain git repo as system of record
Action execution / boundary	src/core/operations.ts (contract), src/mcp/{dispatch,server,http-transport,rate-limit}.ts, OperationContext.remote trust boundary, source isolation
Checking / acceptance	src/core/eval-contradictions/*, src/core/cycle/grade-takes.ts, src/core/cycle/nightly-quality-probe.ts, src/core/skillopt/* (validate-gate, score, held-out), src/core/guardrails.ts
Instruction artifacts	skills/ (43 skills, RESOLVER.md, manifest), templates/*.template, INSTALL_FOR_AGENTS.md, AGENTS.md/CLAUDE.md for the GBrain repo itself
Doctrine	README.md, DESIGN.md, docs/ (architecture, ethos, eval, takes-vs-facts, contradictions, guardrails)

Excluded (named, not analysed): admin/ dashboard UI; test/ and tests/ (1,244 files) except as named pins; scripts/ CI guards; release/versioning tooling; binary self-update; embedding-provider recipe internals; evals/ harness internals; docs tutorials as walkthrough narrative.

External dependencies (outside boundary, named): Anthropic Messages API and other model providers via the gateway; ZeroEntropy embeddings + reranker (default); OpenAI/Voyage/etc. embedding providers; Postgres/Supabase or PGLite (Postgres 17 WASM); the host agent platform (OpenClaw, Hermes, AlphaClaw, Claude Code, Codex, Cursor, ChatGPT, Perplexity) that runs the outer conversational loop; git; S3/Supabase storage for attachments; Twilio/OpenAI Realtime for the voice recipe.

Boundary character: whole-system for the GBrain repository, but subsystem-only with respect to the advertised agent loop. README's "signal → search → respond → write → auto-link → sync" loop places the signal detector and the respond step inside the host agent platform, which is not in this checkout. Conclusions about the end-to-end agent loop are therefore bounded; conclusions about GBrain's own scheduling, retrieval, retention, and synthesis machinery are not.

4. Overall evidence tier

code-grounded. Every material loop recorded in the runtime account rests on inspected TypeScript implementation in this checkout. Doctrine sources are cited as doctrine/design and never upgraded.

No observed run and no causal experiment evidence exists in this run: nothing was executed. Benchmark numbers in README.md (P@5 49.1%, R@5 97.9%, +31.4 P@5) and production counts (146,646 pages; 100,720 takes) are claimed — the underlying runs live in the sibling gbrain-evals repo, outside the boundary.

5. Source register

ID	Kind	Identity / location	Revision / capture	Evidence layer	Inspected scope	Access gaps
SRC-01	repo checkout	related-systems/gbrain (whole)	9a0bae8	implementation	tree listing, file inventory	none
SRC-02	doc	README.md	9a0bae8	doctrine/design + reported operation	full (442 lines)	production numbers unverifiable here
SRC-03	doc	CLAUDE.md	9a0bae8	doctrine/design	full (732 lines)	—
SRC-04	code	src/core/minions/handlers/subagent.ts	9a0bae8	implementation	full (1,275 lines)	—
SRC-05	code	src/core/minions/worker.ts	9a0bae8	implementation	lines 1–470 read; rest scanned by grep	tail of file not line-read
SRC-06	code	src/core/minions/queue.ts	9a0bae8	implementation	claim, handleTimeouts, schedule columns	~1,200 lines not line-read
SRC-07	code	src/core/cycle.ts	9a0bae8	implementation	header + CyclePhase + ALL_PHASES + lock/scope commentary	body of phase dispatch not line-read
SRC-08	code	src/core/think/prompt.ts	9a0bae8	implementation	full (227 lines)	—
SRC-09	code	src/mcp/dispatch.ts	9a0bae8	implementation	full (284 lines)	—
SRC-10	code	src/core/facts/meta-hook.ts	9a0bae8	implementation	full (133 lines)	—
SRC-11	code	src/core/facts/decay.ts	9a0bae8	implementation	full (63 lines)	—
SRC-12	code	src/core/facts/extract.ts	9a0bae8	implementation	header + config gates (lines 1–70)	body not line-read
SRC-13	doc	docs/takes-vs-facts.md	9a0bae8	doctrine/design + reported operation	full (94 lines)	2026-05-10 extraction run is reported, not observed here
SRC-14	code	src/core/skillopt/{apply-edits,audit}.ts	9a0bae8	implementation	headers	rest of skillopt/ not line-read
SRC-15	file	templates/{SOUL,HEARTBEAT,USER,ACCESS_POLICY}.md.template	9a0bae8	doctrine/design	SOUL + HEARTBEAT heads	USER/ACCESS_POLICY not read
SRC-16	dir	skills/ (125 files incl. RESOLVER.md, manifest.json, _brain-filing-rules.md)	9a0bae8	implementation + doctrine	inventory only	individual SKILL.md bodies not read
SRC-17	code	src/core/search/* (30 files)	9a0bae8	implementation	inventory only	hybrid.ts (1,870 lines) not line-read
SRC-18	code	src/core/operations.ts	9a0bae8	implementation	referenced via dispatch.ts + CLAUDE.md invariants	4,751 lines not line-read

Targeted reads inside the frozen boundary are permitted, but they must be reported back to the orchestrator for central registration; they invalidate affected downstream findings.

6. Canonical records registered so far

Components (CMP-*)

ID	Component	Substrate	Form
CMP-01	Minions job queue (minion_jobs + MinionQueue)	Postgres/PGLite tables	symbolic (SQL + TS)
CMP-02	MinionWorker (in-process concurrent poller)	Node/Bun process	symbolic (TS)
CMP-03	Minion supervisor / child-worker supervisor	OS processes	symbolic (TS)
CMP-04	Subagent LLM tool loop handler	Node/Bun process + provider API	mixed (TS + prompt)
CMP-05	AI gateway (src/core/ai/gateway.ts)	in-process module	symbolic (TS)
CMP-06	Brain tool registry / allowlist	in-process module	symbolic (TS + JSON schema)
CMP-07	Dream/maintenance cycle orchestrator (runCycle)	in-process module	symbolic (TS)
CMP-08	Hybrid retrieval stack (vector + BM25 + RRF + rerank + graph signals)	Postgres/pgvector + provider APIs	symbolic (TS + SQL)
CMP-09	think synthesis pipeline	in-process module + provider API	mixed (TS + prompt)
CMP-10	Facts hot-memory subsystem	facts table	symbolic + natural-language rows
CMP-11	Takes cold-storage subsystem	takes table	symbolic + natural-language rows
CMP-12	Brain repo (markdown, system of record)	git repo on disk	natural-language
CMP-13	Engine layer (BrainEngine: PGLite / Postgres)	database	symbolic
CMP-14	Operation contract + MCP dispatch	in-process module	symbolic (TS)
CMP-15	Schema packs (gbrain-base-v2 etc.)	config + files	symbolic (declarative pack)
CMP-16	Skillpack (43 skills + RESOLVER)	markdown files in agent workspace	natural-language (prompt-consumed)
CMP-17	SkillOpt optimizer	in-process module + provider API	mixed
CMP-18	Contradiction eval (eval-contradictions)	in-process module + provider API	mixed
CMP-19	Calibration profile builder	in-process module + provider API	mixed
CMP-20	Audit/JSONL trail (~/.gbrain/audit/*.jsonl)	filesystem	symbolic

Operative objects (OBJ-*)

ID	Object	Substrate	Form	Producer	Consumer
OBJ-01	Page (markdown + frontmatter + timeline)	git repo + pages table	natural-language	user, agent, dream cycle	retrieval, think, sync
OBJ-02	Chunk + embedding	chunks table / pgvector	distributed-parametric + text	chunker + embedding provider	vector search
OBJ-03	Typed graph edge (attended, works_at, …)	links table	symbolic	auto-link on every page write (no LLM)	graph query, graph signals
OBJ-04	Fact row (hot memory)	facts table	natural-language + typed fields	per-turn Haiku/Sonnet extractor	recall, MCP _meta.brain_hot_memory
OBJ-05	Take row (cold, attributed claim)	takes table	natural-language + typed fields (kind, holder, weight, since)	LLM extraction from pages; consolidate bridge	think, takes CLI/MCP, grading
OBJ-06	Synthesis answer (+ citations + gaps)	think output, optionally a saved page	natural-language	think LLM call	user / calling agent
OBJ-07	Structured citation (page_slug, row_num)	synthesis_evidence	symbolic	think structured output	citation persistence, cite-render
OBJ-08	Gap statement	think output field	natural-language	think LLM call	user; --rounds N follow-up
OBJ-09	Take verdict (graded)	takes grading rows	symbolic + natural-language	judge model in grade_takes	calibration profile
OBJ-10	Calibration profile (pattern statements + bias tags + Brier)	profile store	natural-language + symbolic	calibration_profile phase	think prompt (<calibration> block)
OBJ-11	Suspected-contradiction record	contradictions cache	symbolic + natural-language	query-conditioned LLM judge	dream cycle report, agent
OBJ-12	Subagent conversation turn	subagent_messages	natural-language blocks (JSONB)	subagent loop	crash replay
OBJ-13	Tool execution record	subagent_tool_executions	symbolic	subagent loop	crash replay
OBJ-14	SKILL.md	markdown file	natural-language (prompt)	maintainers; SkillOpt	host agent platform
OBJ-15	SOUL.md / HEARTBEAT.md / USER.md / ACCESS_POLICY.md	markdown file	natural-language (prompt)	soul-audit skill	host agent platform
OBJ-16	Schema pack	declarative pack file	symbolic	maintainers; schema detect/suggest/review-candidates	parse, extract, route, cache key
OBJ-17	Job row	minion_jobs	symbolic	submitters	worker claim
OBJ-18	Audit event (JSONL)	filesystem	symbolic	handlers, skillopt, batch-retry	gbrain doctor
OBJ-19	Query cache row	query_cache	symbolic + vector	hybrid search	hybrid search
OBJ-20	Atom / concept page (lens packs)	pages	natural-language	extract_atoms, synthesize_concepts phases	retrieval

Routes (RTE-*)

ID	Route	Kind	Endpoints
RTE-01	Job submit → claim → handler → complete/fail/dead	control	submitter → minion_jobs → MinionWorker → handler
RTE-02	Subagent turn loop (LLM call → tool dispatch → persist → next turn)	control + action	handler → provider API → brain tools → DB
RTE-03	Cron/daemon → dream cycle → ordered phases	control	cron/autopilot/autopilot-cycle job → runCycle → phase functions
RTE-04	Query → hybrid retrieval → ranked chunks	context	caller → search stack → pages/chunks/links
RTE-05	think: gather → merge → synthesize	context + truth-apt	question → pages/takes/graph/trajectory/calibration → LLM → answer+citations+gaps
RTE-06	Conversation turn → fact extraction → facts table	state (write)	host agent turn → extractor → DB
RTE-07	Any MCP tool call → _meta.brain_hot_memory injection	context (read-back, push)	dispatcher → facts → tool response
RTE-08	recall op → facts	context (read-back, pull)	agent → facts
RTE-09	Facts → dream consolidate → takes	state (promotion)	facts → LLM consolidation → takes
RTE-10	Pages → propose_takes → review queue → accepted takes	truth-apt	markdown → LLM proposal → human gate → takes
RTE-11	Unresolved takes → grade_takes (evidence retrieval + judge) → verdicts	checking	takes → retrieval → judge model → verdicts
RTE-12	Verdicts → calibration_profile → profile → think prompt	state + context	verdicts → aggregation → <calibration> block
RTE-13	Retrieval pairs → contradiction judge → conflict records	checking	pages/takes → date pre-filter → judge → cache
RTE-14	Page write → auto-link (regex, no LLM) → typed edges	state (write)	put_page → link extraction → links
RTE-15	Brain repo git → sync → DB (soft deletes)	state	git → import → pages/chunks
RTE-16	MCP request → scope/localOnly/trust/source-isolation gate → op handler	authority	client → transport → dispatch → operations
RTE-17	Skillpack install → host agent workspace → prompt context	context (static)	installer → skills/ in workspace → host agent
RTE-18	SkillOpt: benchmark → proposed edits → score → accept/reject → SKILL.md	checking + state	skill + benchmark → optimizer → validate-gate → version store
RTE-19	Schema pack activation → parse/extract/route/cache-key behavior	control + state	pack → every read and write path
RTE-20	Handler events → JSONL audit → gbrain doctor checks	observability	handlers → files → doctor

Behavioral-authority paths (BAP-*)

Definition in force: one consumption path's consumer, channel, and force, plus this run's horizon field.

ID	Consumer	Channel	Force	Horizon
BAP-01	Host agent platform LLM (OpenClaw/Claude Code/Codex)	skillpack markdown loaded into workspace prompt	advisory instruction (model may deviate)	every session where the skillpack is installed
BAP-02	Host agent platform LLM	SOUL.md/HEARTBEAT.md identity + cadence files	advisory instruction	persistent until regenerated by soul-audit
BAP-03	Subagent loop model	buildSystemPrompt(toolDefs, data.system) system block	advisory instruction	one subagent job
BAP-04	Subagent loop model	tool schemas (input_schema) in the API request	advisory shape; enforced server-side on execution	one subagent job
BAP-05	Brain tool executor	filterAllowedTools / allowed_slug_prefixes / PROTECTED_JOB_NAMES	binding enforcement (code refuses)	one subagent job
BAP-06	MCP operation dispatcher	scope: read/write/admin + localOnly + OperationContext.remote	binding enforcement (pre-handler)	every remote MCP call
BAP-07	Read-side operations	sourceScopeOpts(ctx) source isolation, takesHoldersAllowList	binding enforcement (SQL predicate)	every read op
BAP-08	think synthesis model	THINK_SYSTEM_PROMPT_BASE hard rules (cite everything, surface conflicts, name gaps, never instruct)	advisory instruction	one think call
BAP-09	think synthesis model	<take> contents declared DATA, not instructions; INJECTION_PATTERNS sanitizer	advisory instruction + partial code enforcement	one call
BAP-10	Human operator	propose_takes review queue; schema review-candidates; SkillOpt bundled-skill gate (proposed.md)	binding gate (no write without accept)	per proposal
BAP-11	Worker scheduler	quiet hours, rate leases, budget tracker, RSS watchdog, lock fencing	binding enforcement	continuous
BAP-12	Downstream agent	_meta.brain_hot_memory on tool responses	advisory context (client may ignore unknown _meta)	per tool call, 30s cache

7. Vocabulary in force

• Conclusion statuses: absent, inapplicable, uninspected, claimed, implemented, observed, causally supported.
• Never upgrade: context presence → activation; implementation → observed operation; observation → causality; operational continuation → warrant.
• Memory read-back = material accumulated or changed through use returns to a later invocation or action. Static shipped material (skills, tool specs, docs) and ordinary current-run state are retained state, not read-back.
• Activation = evidence that delivered material changed behavior, not merely that it entered context.
• Truth-apt = capable of truth or falsity.
• Keep behavioral, epistemic, and operational authority separate.