ClawVault reconciliation

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

Reconciliation

Boundary and evidence basis

The four reports use the same revision and the implementation and doctrine/design layers registered in the boundary. Their source inspections differ; none supplies observed operation or causal evidence. This reconciliation connects the pinned records without rereading source or changing their findings. No correction answers or declined objections were supplied. The memory report's Integration issues are addressed below. No disagreement between the supplied reports was identified; their local inspection limits do not become contradictions merely because another analyst inspected further.

Duplicate supersessions

Amendment: EPI-OBJ-extracted-facts is superseded by MEM-OBJ-fact-store. Both identify the same ExtractedFact records in .clawvault/facts.jsonl, with entity/relation/value, confidence, source/rawText, validity fields and the plugin's additional entity-graph serialization. Both cite SRC-1, src/lib/fact-store.ts and src/plugin/fact-extractor.ts; the epistemic declaration additionally cites the extraction type in src/lib/fact-extractor.ts. The memory record's save and validity passages and the epistemic record's facts-path and rawText passages establish one referent, not separate stores. The epistemic inventory and ledger therefore concern the memory record's object. Its scope-loss and warrant assessments remain the epistemic report's findings; supersession does not replace either report's values. Both IDs remain declared. Neither record establishes a later consumer for the separate graph snapshot merely by establishing its serialization.

Amendment: EPI-RTE-replay-import is superseded by MEM-RTE-history-replay. Both identify the shipped replay command's caller-selected historical import: normalize export messages, filter/group dates, use an Observer with threshold one and a date-specific clock, process/flush, then reflect over the span. Both cite SRC-1, src/commands/replay.ts; the epistemic record also identifies CLI dispatch in bin/register-maintenance-commands.js. Their retained passages establish the same importer rather than two successive replay stages. This connects the epistemic import assessment to the memory route's lineage/read-back account without claiming normalization fidelity, transactional completion or observed import. Both retain wired implementation and uninspected operation; both IDs remain declared.

Containers, parts and distinct mechanisms

MEM-OBJ-observations and EPI-OBJ-entity-profile are different document families within RT-OBJ-vault-documents, not duplicate identities of the entire vault. Their declared containment remains valid. MEM-OBJ-raw-transcripts identifies JSONL input strings, and MEM-OBJ-embedding-cache identifies retained numerical selector state; neither is the Markdown document referent. The cache is also not the inference component RT-CMP-hosted-embeddings: stored vectors and provider state have different substrates and consumers. No supersession among these records is warranted.

MEM-RTE-fact-context is a valid combined account of fact production, persistence and explicit read-back. EPI-RTE-fact-extraction, EPI-RTE-fact-supersession and EPI-RTE-fact-context separately individuate proposition construction, active-set disposition/retention and later context selection. None alone has the combined memory route's referent. Together they refine it without making its findings wrong, so the container is not superseded by a split. The epistemic extraction record adds a source-level fictional-scope counterexample; the memory report's unassessed extraction accuracy does not assert the contrary. Neither report turns that implementation deduction into an observed incident.

The fact-selection route EPI-RTE-fact-context also overlaps MEM-RTE-explicit-context, but is not identical to it. The memory route includes multiple context sources, strategy recall and injectable-item selection, whereas the epistemic route isolates the fact branch. Neither is identical to RT-RTE-prompt-recall, whose consumer is a host prompt-build event using its own vault retrieval. These overlapping accounts do not establish that the plugin automatically consumes the explicit builder's facts, profiles or token budget. No supersession is warranted.

EPI-RTE-router-admission, EPI-RTE-question-rewrite and EPI-RTE-fixture-scoring retain their declared part relations to RT-RTE-observe-session, RT-RTE-outbound-filter and RT-RTE-autoresearch-change, respectively. Category admission is not the whole observation loop; semantic rewrite is not the whole search/cancellation hook; score production is not the script's keep/discard decision. Their added checks and limits do not invalidate the containers, so none supersedes its parent. EPI-OBJ-outbound-message and EPI-OBJ-scoring-result are their distinct transient outputs, not retained vault documents.

EPI-RTE-entity-synthesis extends the synthesis stage mentioned in RT-RTE-capture-turn, but also covers independent refresh/read APIs. It is not identical to capture admission, and its output profile is not identical to the producing route. Its separate assessments of heuristic kind, copied summary and file-domain counts do not contradict runtime's previously uninspected synthesis internals.

MEM-RTE-document-patch changes an existing body rather than admitting a new file through RT-RTE-slot-store. MEM-RTE-observation-archive individuates movement/withdrawal also invoked by RT-RTE-reflect-week; it is not the weekly consolidation writer. Replay is an additional producer feeding observation and reflection, not a duplicate of either downstream route. These mechanisms remain distinct without supersession.

Independent convergence and workflow connections

The workflow account annotates RT-RTE-session-handoff without declaring another route. It confirms caller-owned handoff meaning, optional secondary observation/reflection, and return/printed context as the endpoint. Its source sequence adds failures outside the secondary-work catches, flag clearing before complete wake delivery, different recovery/recap selectors and a possible nonzero exit despite printed output. These qualify command completion without contradicting the runtime's lack of transactional or recovery-completeness guarantees. The workflow's comparison to RT-RTE-slot-store identifies shared core storage, not identity between the slot entry point and the sleep/wake route. The memory checkpoint annotation adds alternative durable recovery forms, not another declaration requiring supersession.

Runtime, memory and epistemic accounts agree that RT-RTE-capture-turn applies route-local quality proxies, not a vault-wide truth gate. Direct storage, patch, peer transfer and Python storage retain their separate admission paths. They also agree that selected snippets in RT-BAP-system-context constitute requested influence, not demonstrated model compliance. The epistemic question-rewrite record supplies previously uninspected helper semantics; it does not contradict the runtime container's bounded claim that the hook can request replacement or cancellation.

RT-RTE-maintain-inbox remains a combined maintenance route. Memory adds distinctions among reports, derived prose and processed hashes; epistemic adds distiller source linkage and janitor age-based actions. These agree with RT-BAP-worker-state: processed hashes have a demonstrated selector, while later agent use of a report remains unestablished. No duplicate maintenance route is declared.

RT-RTE-autoresearch-change and the epistemic fixture-scoring part agree on a bounded metric-driven development pathway, not demonstrated production improvement. The doctrine/script acceptance discrepancy is already a finding within both accounts, not a disagreement between them. The epistemic negation counterexample limits semantic interpretation of scores without disputing the existence of score arithmetic or scalar admission. Claimed autonomous iteration, afforded reflection and dormant self-improvement are not strengthened into exercised learning or theory criticism.

Integration-issue dispositions and retained limits

Memory issue 1 is additive coverage: its fact route and the epistemic fact records refine runtime's explicit inspection limit. Asynchronous router counts remain insufficient evidence of complete fact persistence. Issue 2's statement of no intentional duplicates holds relative to the runtime supplied to that analyst; the two cross-specialist identities above require supersession in the assembled set. The remaining proposed identity comparisons are ruled out by their differing referents or valid containment.

Issue 3 remains a limitation, not a conflict: MEM-RTE-observation-archive establishes core reindex withdrawal, not universal forgetting. Derivatives and independently enumerating consumers prevent a system-wide erasure conclusion. Issue 4 independently converges with EPI-RTE-fact-supersession: rewriting the FactStore map and same-ID replacement do not establish immutable audit history. A source header's append-only description is not an opposing analyst finding.

Issue 5 remains unresolved coverage rather than evidenced absence. The reports retain uninspected task/project/migration and related mutations, detailed matchers, checkpoint-history selection, SDK helpers, dashboard parsing and the separate evaluation adapter. Epistemic entity synthesis and workflow recovery inspection add local coverage; they do not settle every entity mutation or checkpoint-history path. These gaps prevent exhaustive admission, trust, parity, recovery and utility conclusions without invalidating the traced mechanisms.

Issue 6 adds an embedding producer and a bounded freshness countercase in MEM-OBJ-embedding-cache. Runtime expressly leaves document-vector freshness unproved, so the first-12,000-character hash limitation is compatible with its finding. Neither report supplies an actual stale-vector incident.

No required missing part or unresolved cross-report dependency contradiction was identified in these pinned accounts. Source-native claims remain distinct: RT-CLM-structured-memory concerns memory/context capacity, RT-CLM-autonomous-research concerns development iteration, and EPI-CLM-memory-answer concerns communication. Their differing warrant limits do not justify merging them or selecting a single system-wide status.