llm-wiki reconciliation

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

Reconciliation

Shared boundary and evidence

All four reports use the same boundary and Source register. SRC-1 supplies implementation and doctrine/design, not observed runs or causal experiments. Runtime's wired helper paths and epistemic's implemented functions do not conflict with their uninspected operation fields. Semantic workflows interpreted by the external host remain prescribed behavior. Reconciliation selects no stronger operation, reliability or benefit status. The cited members supply the evidence passages; no additional source reads were needed. No analyst correction answers or declined objections were supplied to this attempt.

Identity dispositions

Amendment: EPI-RTE-lesson-extraction is superseded by MEM-RTE-extract-lessons. Both identify the same claude-plugin/commands/ll.md conversation-to-lesson workflow: model extraction, structured raw note, append-only article Rule and separately approved instruction proposals. Both retain the prohibition on automatic agent-instruction edits. Epistemic's content-transformation and behavior-adaptation ledger rows concern functions of this same route. This identity does not replace either analyst's semantic-relation or theory assessments, or establish exercised learning.

Amendment: EPI-RTE-source-refresh is superseded by MEM-RTE-refresh-sources. Both identify claude-plugin/commands/refresh.md: sampled key-fact comparison, user skip/update/flag/retract choice, raw overwrite and recompilation on update, and later use of changed files/dates. Epistemic's checking, disposition and acquisition rows concern this same route. The delegated-visibility disagreement below survives this identity; supersession chooses neither field.

Other overlaps do not establish interchangeable whole records:

• EPI-OBJ-raw-articles combines raw claims and symbolic frontmatter. MEM-OBJ-raw isolates source bodies, while MEM-OBJ-frontmatter covers fields across raw and article files. The metadata record is not wholly contained in the epistemic raw object. These are overlapping inventories, not whole-record identity. No supersession or split is asserted: separation does not show the container's findings wrong.
• EPI-OBJ-wiki-articles includes compiled articles and thesis files with verdict fields. MEM-OBJ-wiki names synthesized concept/topic/reference bodies, and MEM-OBJ-frontmatter separates document fields. Ordinary article content overlaps, but the thesis-bearing container is not identical to the body-only record. No supersession is asserted.
• EPI-OBJ-feedback-candidates includes candidate, status/index and promoted-note surfaces. MEM-OBJ-feedback distinguishes candidates from promotion status; promoted raw content also falls within MEM-OBJ-raw. The broader container remains alongside those distinctions. Neither object is identical to the session digest.
• EPI-OBJ-trust-reports is retained diagnostic material, not its producing MEM-RTE-review-memory route. EPI-RTE-trust-review traces semantic checking and advisory verdicts within memory's broader health-review, resume and proposal progression. Shared audit/librarian passages do not make those whole route scopes identical. No supersession is asserted.
• EPI-RTE-feedback-classification and EPI-RTE-feedback-promotion separate candidate transformation from explicit raw-note retention within MEM-RTE-feedback-candidates. Neither function is identical to that valid combined progression. Nor is candidate promotion identical to RT-RTE-session-bridge, whose promotion copies a digest body. No supersession is asserted.
• EPI-RTE-privacy-retract covers implemented literal mutation/rescan. MEM-RTE-withdraw-memory also includes the separate model-operated source-path dependency cleanup. Literal replacement is not identical to that combined withdrawal account and does not establish semantic removal of unsupported claims. No supersession is asserted.

The explicitly declared research parts EPI-RTE-source-review, EPI-RTE-source-selection and EPI-RTE-thesis-verdict remain parts of RT-RTE-research-paths, not duplicate end-to-end routes. EPI-RTE-fixture-evaluation identifies delivery/check functions within RT-RTE-query-evaluation. Its separately inspected JS assertions are comparison evaluators, not claimed benchmark components. Neither valid container requires supersession.

Memory's MEM-OBJ-session-state and MEM-OBJ-session-digest are explicitly retained parts of RT-RTE-session-bridge. Raw/article/index objects are inputs or outputs of runtime routes, not duplicate routes. MEM-OBJ-schema and MEM-BAP-topic-guide distinguish a guidance object from its authority path. Neither is identical to RT-OBJ-query-profile, RT-BAP-query-instructions or RT-BAP-specialist-methods: their text, selectors and consumers differ.

The remaining memory inventory, dataset, project and research-provenance objects have no interchangeable runtime or epistemic declaration. Source admission precedes compilation; review, refresh, lessons and withdrawal have separate triggers/effects from lint and compile. Workflow declares no new records and annotates the existing research route. Epistemic's knowledge, audit and feedback claims are distinct assertions from RT-CLM-concurrent-writes. These relationships require no additional supersessions.

Disagreements and changed connections

The prior reconciliation's research-owner conflict no longer applies to the supplied records. EPI-RTE-source-review and EPI-RTE-source-selection now distinguish single-path foreground review from plan-mode path-agent review and ingestion. Their progression, visibility, selection and effect fields, ledger and bounded conclusion place plan-path raw writes before foreground compilation. Their retained plan-mode passages assign the full search → credibility → ingest pipeline to each path-agent and require the later compiler to read already-ingested sources. This agrees with RT-RTE-research-paths and its workflow annotation. Foreground cross-path compilation is not a pre-ingest source-admission gate. No current dependent finding requires the earlier undifferentiated ownership account. This connection does not extend the selected workflow to thesis mode or establish actual execution.

Unresolved conflict: EPI-RTE-source-refresh calls delegated visibility inapplicable because its inspected protocol uses foreground fetch/comparison and user choice. MEM-RTE-refresh-sources calls it uninspected because the same command declares Agent without specifying a concrete delegated packet. Both cite claude-plugin/commands/refresh.md and retain the same action protocol. Their disagreement is whether delegation is outside this route or an unresolved alternative, not merely shared uncertainty about operation. It prevents a settled account of refresh context boundaries. Reconciliation neither infers a worker from a tool declaration nor excludes one through supersession.

Integration dispositions and independent convergence

Raw preservation and rebuildability. Memory's integration issue adds refresh's explicit overwrite passage to runtime's retraction and compile-migration exceptions. Epistemic treats this as a qualification of runtime's already non-universal account, not a contradiction between reports. MEM-OBJ-raw and the refresh records retain the source-level inconsistency between collection version preservation and authorized overwrite. MEM-OBJ-wiki and RT-RTE-compile-wiki independently retain conversation authorship. These qualify the concurrency claim's raw-only rebuild rationale; they establish no universal immutability or lossless reconstruction.

Session admission and delivery. Memory's annotation on RT-RTE-session-bridge reports enabled/balanced defaults despite the full skill's opt-in language. Runtime says configuration controls enablement but asserts no contrary default. This supplements the entry account; it retains a doctrine/default discrepancy rather than a cross-report contradiction. Runtime and memory distinguish state-derived summary/pointer delivery from digest-body delivery, and epistemic retains the operational/canonical distinction. Separate promotion does not certify factual warrant or install an instruction. Manual-summary sanitation, regex classification and omitted prior-result context remain bounded by MEM-OBJ-session-digest, MEM-OBJ-feedback and the epistemic feedback records; the runtime entry trace does not certify them.

Query and index policy. MEM-OBJ-index and memory's query annotation distinguish the explicit read-only query command from ambient full-skill repair/logging instructions. RT-RTE-query-wiki already scopes compact read-only behavior instead of claiming a whole-system invariant. The explicit command is an entry-specific qualification, not a contrary finding about that route. Indexing doctrine's rebuild-on-read instruction remains inconsistent with the command's no-write policy. Count-only freshness does not certify updated summaries; index repair does not preserve overwritten article edits. Launcher tool requests are not inspected host isolation.

Dates and semantic review. MEM-OBJ-frontmatter and EPI-OBJ-wiki-articles retain compilation's instruction to set today's verified date. Refresh and trust-review records show why that date cannot uniquely establish human confirmation. Runtime makes no contrary human-verification finding. Epistemic's source-selection, thesis-verdict and trust-review doctrine adds semantic checks without contradicting runtime's narrower structural lint account. Memory's broader health-review progression and epistemic's diagnostic functions agree that findings can persist without silently rewriting knowledge. Acceptance-like thesis doctrine is not an evidenced accepted ampliative instance, and article retention/linking alone is not post-acceptance integration.

Bounded executable checks and guidance. Runtime, memory and epistemic retain distinct targets for adapter-plan/receipt matching, checkpoint hashes/scanner/word floor, specialist structural admission, and fixture grading. None gives those checks universal factual authority. Checkpoint semantic selection, final publication and import remain separate from sealing. Enabled method delivery does not grant tools or establish method quality. Benchmark reports can inform maintainers; the inspected path does not automatically install a winning profile. These are convergent scope limits, not conflicting statuses.

Recovery, delivery and remaining coverage. Workflow adds nested delegation, per-path source counts, filename-prefix variation, an unspecified failed-path skip transition and one-session-per-wiki recovery limits. These qualify runtime's selected coordination protocol without negating its sequential compiler. MEM-OBJ-session-provenance independently distinguishes ephemeral crash state from durable provenance. Retained files, reasons and terminal reports establish neither successful replay nor semantic acceptance. Available but uninspected dataset, collection, output, package-generation and evaluation branches remain local coverage limits in the analysts' inventories, not evidenced absences. External payloads, hosts and real histories remain unavailable evidence. These limitations do not erase the supported routes or justify a system-complete negative.

The reports keep retained guidance, source criticism, proposed revision and evaluation capacity separate from learning, reflection, autonomous theory building and exercised self-improvement. Route-specific theory conditions and evidence limits remain the declaring analysts' findings; reconciliation neither combines their statuses into a score nor strengthens one account through another's uncertainty.