expo-llm-wiki reconciliation

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

Reconciliation

Boundary and method

All four reports share the revision and run identity fixed by boundary. The reconciliation uses only the supplied members and their retained evidence passages; no source read was needed. No refusal or correction answers were supplied. The dispositions below connect findings without revising their declaring members or judging their support.

Identity and containment

Amendment: EPI-OBJ-fact-text is superseded by MEM-OBJ-entry-text. Both identify the natural-language fact title/body assertions persisted in SQLite entries, produced by extraction or caller admission and consumed by retrieval and maintenance. The epistemic record's pure.ts validation passage identifies the same title/body fields whose persisted insertion the memory record quotes from ops/apply.ts. The memory record additionally describes tags and lineage; this does not identify another fact-text store. The epistemic inventory and grounding/ops/heal assessments remain about the same assertion content, not the vectors or trust fields. Supersession connects identity only; it does not replace either report's findings about truth-aptness, semantic transformation or source warrant.

The other overlapping objects are not duplicate declarations of the same whole. EPI-OBJ-task-description isolates the natural-language description within MEM-OBJ-tasks, which also covers task status, priority and trust fields. EPI-OBJ-grounding-trust isolates lifecycle and verification fields within the broader access, source, confidence and validity metadata of MEM-OBJ-entry-metadata. Their schema/grounding passages support these narrower operative parts. Supersession is ruled out: the memory containers remain valid alongside these declared parts. EPI-OBJ-judge-verdict has no competing object declaration; the judge provider is its producer, not the verdict itself.

MEM-RTE-temporal-history is explicitly part of RT-RTE-direct-admission. Likewise, EPI-RTE-ops-classify and EPI-RTE-ops-disposition are declared parts of RT-RTE-ops-librarian. Checking and mutation are different functions with different immediate consumers; neither replaces the whole orchestration. No supersession is warranted.

EPI-RTE-heal-assess and EPI-RTE-heal-dispose separately describe assessment and application within MEM-RTE-heal-memory. The latter also includes scheduling, pre-model aging and prompt degradation. All three cite the heal implementation or prompt and describe compatible stages; the combined maintenance route is not made wrong by the separation. Keep the valid container and both declared functions, without supersession.

EPI-RTE-grounding-check and EPI-RTE-grounding-status distinguish a textual-occurrence check from metadata disposition across several writers. They are neither duplicate admission containers nor interchangeable stages. Their relationship to the metadata objects does not establish route/object identity.

EPI-RTE-judge-check, EPI-RTE-judge-record and WF-RTE-replay-cache declare distinct parts of RT-RTE-benchmark-answer. The first two separate evaluation from Boolean recording; the third prepares the retained database. WF-RTE-report-delivery adds the outer CLI spending gate and delivered aggregate. None duplicates or invalidates the runtime's per-question container.

MEM-RTE-graph-context and RT-RTE-native-traversal are parallel TypeScript/presenter and Rust routes, not established identical implementations. MEM-BAP-manifest-guidance overlaps RT-BAP-prompt-guidance at writer prompts but additionally traces persisted manifests to classifier choices and symbolic validation. Neither authority path replaces the other. The optional classifier, text provider, embedding provider and retained embedding vectors also have different operation or producer/output identities. Supersession is ruled out for these overlaps. No other duplicate identity was identified.

Integration dispositions and convergence

Ops rejection. The memory annotation on RT-RTE-ops-librarian and EPI-RTE-ops-disposition independently retain the same apply.ts branch: rejected or missing operations can still insert the candidate. Failed resolution and immutable-target cases can retain drafts; failed supersession linkage need not remove the newly inserted row. This supplies application detail the runtime explicitly left uninspected. It is not a conflict with the bounded orchestration finding and does not turn operation rejection into fact rejection.

Grounding and downstream authority. MEM-OBJ-entry-metadata and EPI-RTE-grounding-status agree that draft status does not by itself exclude later reads. Their retained RetrievalService passage makes exclusion opt-in. EPI-RTE-grounding-check bounds the check to quote occurrence; the memory report bounds what trust and reasons survive storage and formatting. The epistemic report additionally identifies an ops UPDATE exception. These are distinct limits, not conflicting strength assessments. EPI-CLM-grounding-draft is a narrower prompt promise than RT-CLM-longterm-memory, not a duplicate claim.

Oracle and delivery. Memory's benchmark annotation, EPI-RTE-judge-check and workflow's benchmark annotation converge on q.answer as the reference/rubric and prefix-yes parsing as the recorded decision. This resolves the runtime's expressly uninspected oracle-content field without contradicting it. Workflow further traces the public report projection: raw answer/judge text exists in internal question results but is omitted from the delivered aggregate. This narrows claims about delivered evidence, not the runtime record's immediate return. No member establishes a verdict-to-memory repair loop.

Selection and context. Memory's annotations extend RT-RTE-memory-read with formatting and access feedback. Workflow distinguishes its truthy-budget forwarding from core's valid zero-budget path. These are different call boundaries, so the driver omission does not refute the core packing counterexample. The memory graph presenter drops bodies; ordinary formatting omits full provenance/trust/history; maintenance degradation removes prompt copies rather than stored content. None licenses a claim that storage implies complete delivery or enforced model compliance.

Remaining integration limits. The memory report's unresolved imported-summary/vector lineage, ontology initialization/helpers, withdrawal/source-index cleanup and external hook limits remain local uncertainties. Its positive backfill and withdrawal findings do not conflict with the epistemic report's explicitly unassessed coverage of those operations. ScopeLab provider/storage mismatches remain carried limits on RT-RTE-scopelab-chat, not evidence that every local retention route fails. RT-RTE-remote-tools remains remote transport, and RT-RTE-outbox-sync retains its external-commit/local-acknowledgement gap. No local-engine guarantee transfers automatically to those external consumers.

Conflicts and prevented conclusions

No cross-report disagreement requiring an Unresolved conflict: disposition was identified. No required undeclared part was identified for a proposed split. Differing inspection coverage, unresolved provider/lineage facts and unobserved operation are not disagreements. The members retain implemented routes separately from observed operation, reliable semantic revision, learning and causal benefit; reconciliation does not promote any of those conclusions. Their unassessed branches also prevent converting local negative findings into repository-wide absence claims.