GBrain reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

Boundary and status agreement

The reports share the frozen boundary, SRC-1 revision and complete artifact, partial loop scope. Added inspection does not change that boundary or establish complete loop coverage. Runtime, memory and workflow separate wired implementation from uninspected operation; epistemic's implemented architectural status addresses implementation, not observed operation. These are compatible layers, not competing statuses. No member supplies an observed run or causal experiment. This reconciliation uses only the supplied members; no source read was needed. No separate correction answers or declined objections were supplied. Memory's correction dispositions are stated in its Integration issues.

Duplicate identity and independent convergence

Amendment: EPI-RTE-fact-consolidation is superseded by MEM-RTE-facts-consolidation. Both declare the same consolidate-phase progression from eligible fact clusters through the strict confidence reducer, averaged metadata, managed submission or local take insertion, write-gate refusal and contributor marking. Both retain the same reducer and gate passages from SRC-1 src/core/cycle/phases/consolidate.ts. This is identity, not merely agreement about two routes. Affected findings are epistemic's non-ampliative ledger classification and consolidation example, and memory's take lineage and temporal eligibility account. The amendment replaces no analyst field: memory's inspected temporal writes coexist with epistemic's uninspected validity-writeback range. The phase's declared containment in RT-RTE-maintenance-run leaves that valid driver intact; a phase does not supersede its container.

EPI-CMP-grounding-provider is not RT-CMP-chat-model: decide grounding and chat inference are different interfaces and roles; a possibly shared backend does not establish identity. RT-CMP-embedding-model is not MEM-OBJ-fact-embedding: provider inference and retained vector output have different substrates and consumers. RT-CMP-reranker-model is not MEM-OBJ-ranking-weights: external model scores and localized retained feedback weights are separate state. None of these comparisons establishes parameter learning or exact pinned weights.

RT-CLM-memory-control and EPI-CLM-transcript-grounding are distinct claims: general explicit memory/cross-agent availability versus fidelity of generated transcript-derived pages. Epistemic's comparison of grounding doctrine with conditional enforcement does not contradict runtime's attribution of the README claim as doctrine.

Content admission and delivery

MEM-RTE-remember-write and RT-RTE-capture-write are not identical: remember decides fact insertion, duplication or supersession and optional fence publication; capture publishes a page. Shared journal/coordinator controls explain common recovery without identifying their outputs. MEM-OBJ-facts, MEM-OBJ-fact-metadata and MEM-OBJ-facts-fence are not RT-OBJ-persistence-requests: active claim text, eligibility/lineage, page representation and write-intent settlement have different consumers. Capture's database-only branches and MEM-RTE-extraction-write's unfenced fallback independently qualify universal Markdown-publication claims. Software admission establishes neither truth nor mandatory human review.

MEM-RTE-recall-delivery overlaps RT-RTE-query-read in page search but additionally selects fact arms by entity, session, question, supersession or recency. They are not route duplicates. EPI-RTE-answer-check and EPI-RTE-answer-delivery separately check and deliver composed answers; neither duplicates evidence retrieval. Their lexical checks, unverified markers and warn-only citation handling do not turn evidence delivery into accepted truth. Fact metadata's session-loss warning limits cross-rebuild session recall without reversing query wiring.

MEM-RTE-core-delivery explains a selector within RT-RTE-context-assembly, not an identical host callback. Default-disabled core, source intersections, sensitivity/withdrawal filters and character budgets qualify independent supply. Owner designation does not imply prior approval of every edit: notify policy permits publication followed by notice. MEM-RTE-context-pack is a separate explicit/hook-triggered operation; MEM-RTE-delta-delivery retains delivery-position state. Neither aliases OpenClaw assembly, and their budgets establish no universal enclosing-token guarantee. RT-BAP-context-addition describes interpretive text supply, not observed compliance.

Checkpoint correction disposition

MEM-RTE-checkpoint-harvest produces material later consumed by context assembly; it does not duplicate that consumer route. Banking may precede or occur without extraction, and derived facts need not yield manifest links. FIFO receipt repair and inline fallback's best-effort links remain distinct recovery contracts. These positive findings partly resolve runtime's uninspected producer fields without making its scoped uncertainty a disagreement.

The previous missing-part issue is resolved in the supplied memory report. MEM-OBJ-session-checkpoints remains a valid container, with separately declared MEM-OBJ-checkpoint-corpus and MEM-OBJ-checkpoint-manifest, each explicitly part of that container. No split supersession is warranted: the container now preserves rather than conflates its independently assessed parts. Corpus files retain rendered/redacted extraction input and sidecar retry state; the SQL manifest retains verified, capped page pointers. The manifest's quoted existence check, failed-read refusal and ten-link rendering distinguish its admission and later host delivery from corpus banking. Corpus mtime cleanup and SQL session GC are different capabilities, with scheduling still uninspected. Neither retained files nor extracted facts guarantee a surviving retrievable pointer. The former conflict's prevented object-level lineage, retention and consumer account now has named records; complete cleanup and host uptake remain limitations, not residual disagreements.

Withdrawal and maintenance connections

MEM-RTE-fact-withdrawal and EPI-RTE-memory-withdrawal overlap without identical progression. Memory traces the withdrawal transaction and projection invalidation; epistemic includes forget/review disposition, supersession distinctions and active-read overlays. Their different-stage passages independently converge on source/visibility/subject-scoped withdrawal, not universal erasure. MEM-OBJ-fact-withdrawals is persistent negative eligibility rather than journal settlement. New wording, incomplete semantic review, previously delivered bytes and untraced purge residuals prevent stronger suppression claims. Epistemic's accepted-no-action review branch further separates a recorded verdict from an enacted merge/link. EPI-RTE-owner-confirmation separately promotes trust tier; it is neither withdrawal nor general RT-RTE-mcp-dispatch admission, and imposes no mandatory evidence-consuming truth criterion.

EPI-RTE-dream-candidates is an internal synthesis phase, not the maintenance CLI driver or the fact-extraction route. EPI-RTE-mechanical-check, EPI-RTE-mechanical-disposition, EPI-RTE-grounding-check, EPI-RTE-grounding-disposition and EPI-RTE-repair-retention are distinct consequential functions despite shared implementation. Checking, disposition and retention are not duplicates. EPI-OBJ-compiled-truth is a dream body, not the entity facts fence; EPI-OBJ-unverified-claims is separate review retention. Post-write checking, optional/shadow grounding and fail-open repair qualify unconditional grounding without contradicting the driver account. Memory's locally uninspected synthesis fields do not deny epistemic's positive findings.

Feedback correction disposition

MEM-RTE-feedback-update retains outcome-derived weights and supplies a later-ranking witness; it does not duplicate query evidence delivery or subagent replay. MEM-BAP-ranking-weights exercises numerical selection force, unlike model-facing text addition. Epistemic's unassessed feedback and unresolved exercised-learning conclusions do not deny this wired pathway. Neither report establishes observed benefit or successful self-improvement. EPI-RTE-writeback-evaluation is a distinct gold-relative scorer with report/comparison-exit consumers, not production rating admission or maintenance health diagnostics. Its injected-gold and real-extractor modes support different assessments.

The previous feedback missing-part issue is also resolved. MEM-OBJ-retrieval-weights remains a valid mechanism container. MEM-OBJ-retrieval-events, MEM-OBJ-retrieval-feedback and MEM-OBJ-ranking-weights each declare containment and independently assess attribution, admitted history and derived numerical state. No split supersession is warranted. Event ownership and current page hashes bound rating eligibility; feedback insertion controls idempotent weight update; later ranking reads weights rather than history. The quoted cascade schema and weight-stage passages support the critical horizon distinction: event/history pruning can leave derived weights operative without their complete attribution. Revision changes damp deviation from neutral, and disabled influence/read failure can bypass ranking effect. These declarations resolve the former inability to address raw versus derived state and each consumer independently; link consumers, implicit signals, reset authority and cleanup scheduling remain scoped uncertainties.

Delegated workflow and remaining limits

WF-RTE-agent-submission, RT-RTE-subagent-loop and WF-RTE-agent-delivery connect as task admission, execution and settlement/polling. They are distinct from one another and from surrounding MCP dispatch. Workflow adds live-grant intersection, stable tool-write receipt waiting, worker retry/dead disposition and token-fenced completion to runtime's bounded loop. These resolve selected uninspected recovery fields without establishing complete fleet coverage. Memory's replay annotation concerns retained context/recovery, not learned strategies. Optional generic gateway callbacks remain distinct from this durable caller's supplied callbacks.

Completed job status records a normally returned handler result, whose stop reason may still be refused or limited. Supplemental write accounting is not independent semantic review, and the inspected submission does not require writes. Owner-filtered polling establishes delivery, not exact post-redaction bytes or a final host answer. These distinctions agree with runtime's effect boundaries and epistemic's separation of continuation from warrant.

RT-RTE-skill-publication and RT-OBJ-shared-skill-heads keep canonical publication and selection identities. Memory's annotation distinguishes heads from linked instruction bodies/assets; epistemic assesses policy/CAS disposition rather than improvement. No separately declared identical skill record requires supersession. Downstream activation and optimization remain unresolved.

The reports' remaining coverage limits are complementary: pinned-answer writers, compiled instruction consumers, connectors/drains, complete purge/GC propagation, optimization, other evaluators and untraced producer modes are not closed by selective positive findings. No current inter-report disagreement was established beyond the duplicate identity handled by the amendment. The two former missing-part conflicts are resolved by declarations, not by deleting supported routes or choosing stronger statuses. No unresolved conflict paragraph is retained for those now-supplied parts.