GBrain report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The supplied reports support publication as a bounded implementation account. Both previous memory blockers are repaired. The source HEAD matches the boundary, and its working tree is clean. The complete artifact, partial loop classification remains appropriate. No target execution, provider call or experiment was performed. The structural check reports none; runtime, epistemic and workflow answers are empty. Memory's two correction answers were assessed against the changed declarations and frozen implementation, not accepted as proof.

I reviewed every declaration and annotation in the four reports, their source anchors, status assertions, containment and identity comparisons, and the reconciliation's supersession and conflict dispositions. Independent source inspection concentrated on consequential transitions, alternate branches and the newly separated parts. Source evidence is implementation, except prompts, descriptions and comments, which establish doctrine rather than compliance. The grouped dispositions below retain those distinctions. Citation resolution and quotation occurrence are deterministic checks, not substitutes for semantic judgment.

Runtime and workflow

RT-RTE-capture-write and RT-OBJ-persistence-requests correctly separate proposal, durable admission, publication and receipt. CLI byte guards precede decoding; local-owner and remote branches differ. Admission compares request intent and current authority. Publication rechecks source/grants, claim ownership, page identity/revision and canonical bytes before applying state and recording settlement. Derived effects are queued separately. Database-only and pending returns qualify the guarantees rather than contradicting them. SRC-1 src/commands/capture.ts, src/core/ops/pages.ts, src/core/persistence/page-mutations.ts, src/core/persistence/journal.ts, src/core/persistence/authority.ts, src/core/persistence/coordinator.ts, src/core/persistence/service.ts.

RT-RTE-mcp-dispatch, RT-RTE-query-read, RT-RTE-context-assembly and RT-BAP-context-addition establish controls and delivery, not truthful content or host compliance. Dispatch checks the allowed surface and source binding; text and image query branches thread privacy/scope constraints into retrieval and output processing. Assembly supplies text and pointers without taking ownership of host compaction. Direct local handlers and untraced upstream filters remain outside any universal security guarantee. SRC-1 src/mcp/dispatch.ts, src/core/ops/search.ts, src/core/context-engine.ts.

The component findings distinguish model interfaces from retained outputs: RT-CMP-chat-model, RT-CMP-embedding-model, RT-CMP-reranker-model. Defaults, overrides and fallback wiring support configurable inference, not exact weight identity or parameter learning. The embedding-identity refusal is executable. The Claude CLI adapter disables native tools and uncontrolled MCP discovery; optional hermetic configuration and an external CLI contract do not establish OS isolation. SRC-1 src/core/ai/gateway.ts, src/core/ai/defaults.ts, src/core/ai/providers/claude-cli-language-model.ts.

WF-RTE-agent-submission, RT-RTE-subagent-loop and WF-RTE-agent-delivery form a connected conditional workflow. Submission persists the owner and delegation snapshot. The live grant intersects that ceiling before guarded tool execution. The gateway persists assistant and pending execution state before effects when callbacks are supplied; this caller supplies them. Settled outcomes short-circuit replay, while uncertain non-idempotent work refuses repetition. Pending write receipts escape without converting the tool ledger to failed. Worker retry/dead disposition and token-fenced completion differ from model termination. Polling filters by owner and subagent identity. Open submission does not request mandatory writes, and supplemental accounting is not semantic review. SRC-1 src/core/ops/jobs.ts, src/core/minions/delegated-policy.ts, src/core/minions/delegated-tools.ts, src/core/minions/handlers/subagent.ts, src/core/minions/handlers/subagent-persistence.ts, src/core/ai/gateway.ts, src/core/minions/worker.ts, src/core/minions/queue.ts. The workflow annotation and memory replay annotation preserve the runtime referent; neither implies successful recovery or learned strategies.

RT-OBJ-shared-skill-heads and RT-RTE-skill-publication retain a supported publication/activation distinction. Editor authority cannot silently widen disclosure; publication rechecks policy and heads and updates revisions/catalog state. Asset retrieval is described as data-only. The memory annotation does not assign instruction-body or asset execution to the catalog head itself. Downstream activation and improvement are not claimed. SRC-1 src/core/shared-skills/publication.ts, src/core/ops/skills-catalog.ts.

Memory and the two corrections

The fact declarations distinguish text, eligibility/lineage fields, vector output and fence representation: MEM-OBJ-facts, MEM-OBJ-fact-metadata, MEM-OBJ-fact-embedding, MEM-OBJ-facts-fence. Explicit confidence 1 is assigned metadata, not calibrated truth. MEM-RTE-remember-write and MEM-RTE-extraction-write have different producers/defaults, gates and DB-only alternatives. Inferred linking limits dedup without eliminating the vector. Session-loss wording is a declared operation-contract limitation, not observed rebuild evidence. SRC-1 src/core/persistence/memory-prepare.ts, src/core/facts/extract.ts, src/core/facts/backstop.ts, src/core/ops/facts.ts.

MEM-RTE-recall-delivery, MEM-RTE-context-pack, MEM-RTE-delta-delivery and MEM-RTE-core-delivery preserve their different selectors. Recall packs the redacted presentation; packs reserve core before item packing; failed delta-state reads refuse rather than initialize a new cursor. Core is default-disabled, designation-based and filtered at delivery. Notify policy permits remote content edits followed by notice, not prior human approval. These routes do not promise exact enclosing-token budgets or host uptake. SRC-1 src/core/ops/facts.ts, src/core/core-memory.ts, src/core/persistence/core-guard.ts, src/commands/hook.ts.

The checkpoint correction now separately declares MEM-OBJ-checkpoint-corpus and MEM-OBJ-checkpoint-manifest, each explicitly part of MEM-OBJ-session-checkpoints. Independent tracing confirms the distinctions. Rendering/redaction produces file input for extraction; existing content-addressed files are not rewritten. Corpus cleanup uses mtime and a longer un-ingested horizon. The SQL manifest merges verified page references, refuses a failed read as a merge seed, caps entries and normalizes titles. Later rendering selects ten pointers and instructs get_page, not transcript delivery. Its SQL GC is independent of corpus-file GC. SRC-1 src/core/context/corpus-segments.ts, src/core/context/session-state.ts, src/core/context/checkpoint-harvest.ts, src/core/context-engine.ts.

MEM-RTE-checkpoint-harvest preserves the supported raw-to-derived-to-pointer chain and its conditions. FIFO retains a receipt after manifest-write failure; the direct fallback writes ingested even when link publication fails. Thus neither banked corpus nor retained facts guarantees a later pointer. The new part declarations assign extraction-input authority to corpus and host pointer delivery to manifest, while preserving the valid container. Their containment is substantive, not just matching record kinds. The former blocker and its dependent coverage, core-idea, route, read-back and integration passages are repaired. Scheduling, complete cleanup and original host provenance remain expressly uninspected, not unsupported positive findings.

The feedback correction separately declares MEM-OBJ-retrieval-events, MEM-OBJ-retrieval-feedback and MEM-OBJ-ranking-weights, each part of MEM-OBJ-retrieval-weights. Event/client/page-hash state bounds rating eligibility; newly inserted history controls idempotent updates; later ranking reads weights, not history. The event foreign keys cascade to attribution children and feedback, but weights have no event foreign key and can survive their deletion. Weight reset/target pruning and revision damping are separate capabilities. SRC-1 src/core/feedback/store.ts, src/core/schema-migrations/v207-retrieval-feedback.ts, src/core/feedback/rate.ts, src/core/search/feedback-boost.ts.

MEM-RTE-feedback-update and MEM-BAP-ranking-weights retain the positive later-consumer witness. The hybrid stage invokes feedback after reranking, changes ordering keys without overwriting raw scores, and bypasses disabled/zero influence or read failure. SRC-1 src/core/search/hybrid/rank.ts, src/core/search/feedback-boost.ts. Raw attribution, rating history and derived authority are now independently assessable. The former blocker and dependent write/read-back and integration findings are repaired without weakening trace-fed update into mere logging. Benefit, link consumers, implicit signals and complete cleanup remain unestablished.

MEM-OBJ-fact-withdrawals, MEM-RTE-fact-withdrawal and EPI-RTE-memory-withdrawal converge on scoped fingerprint withdrawal, expiry, projection invalidation and canonical overlays. Supersession expiry and accepted-no-action review remain distinct. No universal erasure or paraphrase guarantee is emitted. SRC-1 src/core/facts/withdrawal.ts, src/core/facts/forget.ts, src/core/facts/proposal-review.ts, src/core/facts/withdrawal-overlay.ts, src/core/engine-sql/facts.ts.

RT-RTE-maintenance-run, MEM-OBJ-takes, MEM-RTE-facts-consolidation and EPI-RTE-fact-consolidation support diagnostic driving and deterministic existing-text promotion, not generated semantic warrant. The strict confidence reducer copies a claim; gate refusal and chronological validity writes qualify admission and later eligibility. The phase is correctly part of the driver's cycle action. SRC-1 src/commands/maintain.ts, src/core/cycle/phases/consolidate.ts. Memory's later inspected range does not contradict epistemic's local uninspected range.

Epistemic checks and reconciliation

EPI-OBJ-compiled-truth, EPI-OBJ-unverified-claims and EPI-OBJ-answer distinguish body, review retention and answer variants. EPI-RTE-dream-candidates, EPI-RTE-mechanical-check, EPI-RTE-mechanical-disposition and EPI-RTE-repair-retention correctly separate requests, checks, removal/repair and persistence. Prior normalized units can skip checking; failures remove units; repaired reimport updates projections. Post-write and fail-open paths prevent unconditional pre-use certification. SRC-1 src/core/cycle/synthesize.ts, src/core/cycle/synthesize-verify.ts.

EPI-CMP-grounding-provider, EPI-RTE-grounding-check and EPI-RTE-grounding-disposition retain selected-window, source-relative scope. Threshold/margin/coverage, acting versus shadow mode, drift and unavailable calibration qualify the result. Kept-on-error is not acceptance; probability is not source truth. SRC-1 src/core/cycle/grounding-decide.ts. The indeterminate synthesis classification and separated check/disposition/retention ledger are supported.

EPI-RTE-answer-check, EPI-RTE-answer-delivery and EPI-RTE-owner-confirmation do not overstate warrant. Question text can ground quotations, unmatched spans can be delivered with markers, and citations outside gather warn rather than reject. Confirmation changes trust tier under permission or interactive confirmation without requiring evidence-consuming factual review. SRC-1 src/core/think/index.ts, src/core/cycle/synthesize-verify.ts, src/core/verbs.ts, src/core/ops/trust.ts, src/core/trust/confirm.ts.

EPI-RTE-writeback-evaluation distinguishes injected gold from real extraction and bounds keyword/entity scoring and comparison-exit force. No score or causal effect is asserted. SRC-1 src/eval/brainbench/metrics/write-back.ts, src/commands/eval-brainbench.ts. RT-CLM-memory-control and EPI-CLM-transcript-grounding remain attributed doctrine with bounded implementation comparisons, not observed cross-host reliability or unconditional grounding.

The reconciliation correctly supersedes the duplicate consolidation route without rewriting its findings or superseding the maintenance container. Other overlap comparisons distinguish producer, selector, executor and consumer instead of merging them. All six explicit containment relations are supported: consolidation within maintenance, two checkpoint parts within their container, and three feedback parts within theirs. The two former missing-part conflicts are resolved by declarations; no current unresolved inter-report conflict was found. Local differences in inspected ranges are not disagreements. No evidenced-absence record or observed/causal route status is emitted. Annotations add lens fields without changing generic identity. No reconciliation blocker is needed.

Omission questions and stopping decision

The frozen top-level tree and scoped source-file inventory supplied the search basis, including shipped callers rather than treating API callers as external. Selected questions were whether nominally canonical memory always writes files, whether replay logs exhaust trace-fed behavior, whether checkpoint consumers receive raw corpus or selected pointers, and whether delegated completion certifies successful work. Independent branch and producer/consumer checks above dispose of each: DB-only paths exist; rating-derived weights have a later ranking consumer; corpus and manifest have different consumers and horizons; completion records a bounded handler result rather than semantic acceptance.

Hook dispatch also identifies Stop writeback, SessionEnd and PreCompact paths. Maintenance's phase table identifies concept, grading, pattern and skill-optimization alternatives. These remain included but incompletely assessed in the reports; their existence does not establish absence, benefit or a complete update inventory. SRC-1 src/commands/hook.ts, src/core/cycle/phase-table.ts; see memory coverage and epistemic coverage. Pinned-question writers, compiled-context consumers, remaining evaluator/adaptor internals and full withdrawal/cleanup scheduling were not exhaustively traced. The account supports the selected positive mechanisms, not comparisons requiring complete initialization, all-consumer suppression, optimization or whole-loop coverage.

Further optional omission expansion stopped because the selected consequential questions have supported dispositions, the repaired parts preserve the central memory distinctions, and no identified material challenge remains without disposition. The reports already withhold the broader conclusions those uninspected alternatives would affect. Missing execution evidence prevents exercised behavior and benefit conclusions without negating architectural wiring. These faithfully scoped uncertainties require neither a new blocker nor an additional publication limit.

Blockers

none

Limits

none