llm-wiki reconciliation

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

Reconciliation

Record identities and containment

Amendment: EPI-OBJ-wiki-pages is superseded by MEM-OBJ-wiki. Both identify retained source/entity/concept and saved-answer page content in filesystem Markdown, with host/backend producers and host/MCP consumers. Their shared contradiction-preservation passage comes from SRC-1 CLAUDE.md. Memory also describes frontmatter and links as tooling support; epistemic separates quality metadata from the page bodies. The identity concerns the derived content, not a merger with quality metadata. The epistemic inventory, ledger and conclusions retain their findings about candidate propositions, unknown fidelity and unestablished theory production; no value changes through supersession.

Amendment: EPI-OBJ-quality-metadata is superseded by MEM-OBJ-quality-metadata. Both identify confidence/lifecycle metadata associated with page frontmatter and helper calculations in SRC-1 llmwiki/confidence.py, llmwiki/lifecycle.py and llmwiki/mcp/server.py. Epistemic explicitly includes transient helper values; memory describes the computations and unresolved durable update route. Supersession retains that transient-versus-persisted distinction and does not make returned values persisted facts. Both distinguish candidate status from MCP lifecycle. Inventory, selection and helper findings remain unchanged; neither record establishes synchronization, calibration or a production writer installing computed values.

EPI-OBJ-reader-overview explicitly declares itself part of RT-RTE-reader-build. It is the optional prose output of that route, not a duplicate route. Its evidence in SRC-1 llmwiki/build.py connects model stdout to render_index and reader index.html. It is distinct from the curated wiki content identified above: storage and later consumers differ. The runtime container already distinguishes metadata-only reader prose from source-page synthesis. Neither the object nor the route is superseded. The epistemic ledger and bounded conclusion now cite this separate output rather than treating it as ordinary wiki-page content; no dependent lineage conflict remains in these supplied members.

EPI-OBJ-raw-sessions is the transcript product of RT-RTE-session-conversion, not a duplicate of its producer route. Memory annotates conversion without declaring a second raw-data object. MEM-OBJ-candidate-archive similarly identifies retained products and optional reasons, not RT-RTE-candidate-admission's transitions. EPI-OBJ-lint-issues identifies returned reports rather than checking or command-disposition routes. These product/route connections establish no record identity.

MEM-OBJ-navigation, RT-OBJ-source-page, RT-OBJ-pending-prompts and MEM-OBJ-ingest-queue remain distinct. Their referents are, respectively, retained navigation/scaffolds, a static synthesis template, rendered pending tasks with UUID coordination, and a JSON path queue. Shared storage or foreground consumption does not make them interchangeable. In particular, unresolved queue wiring does not supersede the implemented pending-prompt handoff.

EPI-RTE-metadata-selection declares itself part of RT-RTE-mcp-access. The detailed metadata suboperation does not replace the broader MCP container. MEM-RTE-quality-inspection overlaps these MCP views but also covers independent CLI lint. It is not identical to either MCP record. Its diagnostic coverage overlaps EPI-RTE-semantic-proxies, EPI-RTE-freshness-check and EPI-RTE-lint-return, which distinguish issue production, age warnings and command-status disposition. The grouped report-only account remains alongside the detailed operations; no split supersession is warranted by their findings.

MEM-RTE-wiki-maintenance groups update and reflection, whereas EPI-RTE-page-update and EPI-RTE-wiki-reflection separately describe them. Neither individual operation is identical to the grouped route. The container preserves their different admission rules, so it is not superseded by the separate records. Their implementation-status disagreement is retained below.

MEM-RTE-graph-query adds CLI retrieval from retained graph metadata, distinct from MCP page/snippet retrieval and reader publication. MEM-RTE-self-learn concerns human-approved method-guidance revision, not ordinary RT-RTE-host-ingest or wiki-page reflection. These additions do not duplicate the selected runtime routes.

Claim identities also remain separate. EPI-CLM-source-fidelity concerns traceability rather than RT-CLM-history-compilation's compilation claim. EPI-CLM-semantic-checks concerns lint descriptions; EPI-CLM-lifecycle-states concerns quality-state doctrine. MEM-CLM-archive-visibility compares archive inactivity with actual consumers, while EPI-CLM-candidate-trust records human-approval doctrine for the trusted layer. Related consequences do not make those claims identical. Workflow declares no records and annotates its supplied runtime routes. No further cross-report identity is established.

Cross-report disagreement

Unresolved conflict: MEM-RTE-wiki-maintenance assigns implementation conclusion status afforded to update/reflection; EPI-RTE-page-update and EPI-RTE-wiki-reflection assign claimed to those procedures. Memory treats shipped foreground instructions as revision/review affordances. Epistemic retains the claimed status and assesses the architecture as doctrine only. Both rely on SRC-1 .claude/commands/wiki-update.md and .claude/commands/wiki-reflect.md, agree that compliance is unobserved, and distinguish requested local edits from approval-gated reflection changes. Their implementation assessments differ despite that common evidence. A single implementation-status conclusion for the procedures is prevented; reconciliation selects neither finding and leaves judgment to report verification.

Independent convergence and integration dispositions

The workflow account connects RT-RTE-backend-synthesis to RT-RTE-agent-completion without adding another route identity. Its staging, serial foreground pickup, rejection and delivery account agrees with memory's pending-prompt annotation and the epistemic ledger: the backend writes durable tasks rather than launching a worker; synthesis state can advance on placeholder persistence; UUID equality checks association, not truth. Successful completion writes retained content, while later retrieval, answer dependence and publication remain separate.

RT-CLM-append-preservation is consistently treated across the reports as a source docstring/implementation mismatch. Completion writes preserved frontmatter plus the supplied body, rather than automatically retaining prior appended edits. Workflow adds pre-write rejection and best-effort cleanup details without extending the body-preservation guarantee. This is neither a disagreement between reports nor an observed data-loss finding.

MEM-CLM-archive-visibility converges with the epistemic comparison of EPI-CLM-candidate-trust against RT-RTE-mcp-access: recursive Markdown selection can return candidate/archive text without adjacent rejection reasons. This limits downstream withdrawal, not RT-RTE-candidate-admission's file transitions. Runtime already treats the candidate mechanism as optional and leaves initial proposal identity unresolved. Memory's candidates-first prose versus direct host writes is consequently a source-level tension, not a cross-report conflict.

The quality findings likewise converge. MEM-OBJ-quality-metadata, EPI-RTE-lifecycle-helper, EPI-RTE-metadata-selection and EPI-RTE-freshness-check distinguish computation, display and warnings from durable updates or query exclusion. Unresolved production persistence does not negate implemented helpers. Memory's structural-proxy account is extended by EPI-RTE-semantic-proxies's callback-presence/skip findings. EPI-RTE-lint-return adds command consequences without licensing proposition truth.

Memory's schema-divergence issue qualifies folder-context selection in RT-RTE-host-ingest: Claude doctrine contains the pre-read step, while the agent-agnostic schema omits it. Runtime specifically discusses the Claude source and does not require every host schema to include that step. The qualification therefore leaves no cross-report conflict. MEM-OBJ-navigation describes the retained selection inputs without establishing automatic injection.

The unresolved queue, hot-refresh, Auto Dream and SessionStart surfaces in MEM-OBJ-ingest-queue and MEM-OBJ-navigation are consistent with the boundary's missing-hook limit. They do not erase known direct reads, transformations or deferred completion. The added graph and self-learn routes extend the runtime's explicitly selective maintenance coverage; epistemic's unread variants are inspection limits, not negative findings about these additions.

All members retain the Source register's implementation and doctrine/design evidence boundary. Their uncertainties about actual generated propositions, host compliance, exercised acceptance, consumer dependence and improvement do not themselves disagree. No correction answers or declined objections were supplied here; the separate reader-overview object and its revised dependencies were assessed from the supplied epistemic member, not inferred from an earlier connection. No source reads beyond the supplied members were needed.