LLM Repo Wiki reconciliation

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

Reconciliation

Boundary and identity dispositions

The four analyst reports share the frozen boundary and Source register. They describe implementation and declared intent, not a supplied execution or causal experiment. The workflow annotates runtime records rather than declaring new identities. No correction answers or refusal were supplied for this reconciliation. No additional source reads were needed.

Amendment: EPI-RTE-wiki-lint is superseded by MEM-RTE-lint-wiki. Both identify the same scheduled/live deterministic linter: current accepted pages, five structural finding classes, completed-sync restriction on deleted-source checks, durable scheduled findings and live UI recomputation. Both cite SRC-1 src/wiki/lint.ts, src/wiki/jobs.ts and src/app/w/[wiki]/lint/page.tsx, with retained passages for deterministic checking and the log/live distinction. Their implementation statuses are wired and their operation statuses uninspected. This is identity, not a choice of stronger findings. The epistemic ledger's check target, advisory force and comparison with EPI-CLM-lint-contradictions remain assessments of that same route; neither declaration is removed.

Other overlaps do not establish duplicate identities:

  • MEM-RTE-index-page explicitly declares itself part of RT-RTE-revision-admission. Its narrower transformation is accepted Markdown to indexed sections. The runtime route also covers acceptance, rejection, direct editing and restoration. The larger route remains valid; no split supersession is needed. The workflow's downstream-indexing account and the epistemic ledger's indexing function refer to this connection without declaring another route.
  • MEM-RTE-search-knowledge spans selection used by RT-RTE-author-pages, RT-RTE-chat-turn and RT-RTE-external-consumption. It is not identical to any complete caller route. MEM-RTE-bench-author is a destructive caller of the author route, not that route itself. EPI-RTE-benchmark-checks is a valid three-driver diagnostic container; its author branch overlaps the memory benchmark route, but its chat and MCP branches prevent whole-record identity. Its ledger separates those evaluators. Neither container requires supersession.
  • EPI-OBJ-source-evidence describes retained and delivered source text/coordinates. MEM-OBJ-chunks also includes embedded text, vectors and access structures with different consumers. The delivered excerpts are projections of the stored content, not interchangeable with the complete retained object. EPI-OBJ-page-prose isolates Markdown within the broader revision state of MEM-OBJ-page-revisions. Reasons, status and revision pointers remain separately described in the memory record. These are content/containing-state distinctions, not same-referent duplicates requiring supersession.
  • EPI-OBJ-source-citations includes both source hrefs in prose and symbolic citation objects. MEM-OBJ-revision-sources identifies persisted revision-source tuples. The latter can diverge from inline links after repository removal. Neither record can replace the other: their overlap does not erase the body/metadata distinction.
  • EPI-OBJ-chat-content is a conversation-content container. MEM-OBJ-chat-messages separately describes whole UIMessage payloads, including tool and approval structures; MEM-OBJ-chat-summary separately describes derived prose and its cutoff. The epistemic record distinguishes their consumer timing and does not claim equivalent checkpoints. The memory declarations supply the separate retained objects, so no missing-part or split supersession is required.
  • EPI-OBJ-evaluation-output groups diagnostic output across lint, benchmarks and citation evaluation. MEM-OBJ-eval-results covers retained benchmark/citation files, while MEM-OBJ-wiki-log covers broader database events and accounting. These are overlapping, valid containers with different storage and coverage, not identical objects. EPI-RTE-citation-evaluation supplies the semantic judging route for part of those files; it is neither the files nor RT-RTE-explain-answer, which filters references in diagnostic answers.

The remaining model components, instruction/brief objects, behavioral-authority paths and claim records have no same-referent counterpart across the supplied declarations. In particular, RT-OBJ-wiki-schema is not wiki prose, MEM-OBJ-project-context is not the filesystem schema, and their respective instruction paths are not MEM-BAP-summary-context, which carries trace-derived chat text. The epistemic grounding, evaluation-result and contradiction claims are not RT-CLM-provider-routing: they concern different source claims and licenses.

Independent convergence and integration issues

Admission differs by route. Runtime's RT-RTE-author-pages guarantees proposals within that job, while RT-RTE-chat-turn allows immediate apply-mode creation and selected approval for other changes. Memory independently confirms those distinctions; the epistemic account uses them to bound EPI-CLM-wiki-grounding. The workflow deliberately selects the proposal/review path and explicitly excludes a universal proposal requirement. This is not a disagreement about whether every application write is human-reviewed. Proposal-only UI/MCP visibility further qualifies pre-use review, without contradicting export's accepted-only selection.

Grounding is not semantic acceptance. Runtime's author record retains exact-old-key admission, clipped evidence versus full hit bounds, and the nonzero-citation threshold. Memory's annotation agrees. The epistemic EPI-OBJ-source-citations adds a deterministic illustration of link removal without claim deletion. It does not report an executed draft. The workflow's mechanical-screening description and the epistemic distinction between operational acceptance and evidence-consuming factual acceptance preserve, rather than contradict, the runtime's findings.

Persistence and delivery have separate failure boundaries. RT-RTE-revision-admission, MEM-RTE-index-page and the workflow independently retain enqueue failure and the pre-embedding currentness check. MEM-RTE-search-knowledge adds the omission of revision/stale metadata during hydration. These findings connect acceptance to potentially lagging retrieval, not a guaranteed synchronized snapshot. The workflow adds queue handoff ordering and partial bulk-review/retry outcomes without changing the runtime route's identity or status.

Derived memory is not fully rolled back with messages. RT-RTE-context-fold and memory's annotation agree that regeneration does not repair the stored summary/cutoff. Memory adds the precise clipped transcript and later instruction consumer. The epistemic summary-transformation classification remains indeterminate because requested compression is not checked semantic preservation. A known trace-fed write and an unknown semantic result coexist; neither implies observed improved capacity.

Evaluation coverage differs, not conclusions. Runtime expressly leaves detailed author/chat/MCP and citation evaluator semantics uninspected. Memory supplies destructive benchmark effects and diagnostic resume; epistemic supplies complete semantic-judge inspection in EPI-RTE-citation-evaluation. Thus a semantic checker exists outside the production admission loop, while no automatic verdict-fed production acceptance is established. The narrower runtime account does not assert its absence. Memory's partial helper inspection is a local limit, not counterevidence to the epistemic route. Both specialists also retain apply-mode chat benchmark effects and whole-wiki author wiping; no report supports a general read-only-evaluation claim.

Reported performance remains attributed. EPI-CLM-evaluation-results separates reported results from supplied traces and bounds the semantic percentage to the agreed subset. RT-RTE-retrieval-eval scores expected paths, not factual answers; the benchmark record preserves heterogeneous targets. No report's missing execution evidence cancels its inspectable wiring, and no report's wiring establishes observed benefit. Provider and external-host opacity likewise remain local limits rather than system-wide absence claims.

No unresolved disagreement between the supplied reports was identified. Memory's integration questions about installed-state provenance, index races, summary rollback and provider/host behavior remain bounded uncertainties with explicit prevented conclusions, not competing findings. No dependent finding was found to require a supplied record to say something it no longer says. The sole duplicate supersession above leaves all other declarations and their independent statuses intact.