LLM-Wiki-v3 reconciliation

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

Reconciliation

Evidence boundary and record identity

The members share the frozen boundary and its Source register. Their implementation findings concern SRC-1 at the same commit. None supplies an observed extraction, admission or downstream answer trace. The source-authored baseline remains reported operation, not a reproduced run or causal comparison. This reconciliation uses the pinned members only; no source reread or new evidence layer is introduced.

Amendment: EPI-OBJ-page-body is superseded by MEM-OBJ-page-body. Both identify the natural-language Markdown body crossing proposed and retained page phases, distinguished from frontmatter and provenance fields. Both use the retained Ada page as a candidate artifact without treating it as observed production or acceptance. The memory record supplies lineage and later consumers; the epistemic record supplies the truth-apt role. Identity concerns the body, not an upgrade from candidate claims to accepted knowledge. Epistemic extraction, checking, retention and selection findings referring to this object retain their own scope and statuses.

Amendment: EPI-OBJ-query-labels is superseded by MEM-OBJ-eval-queries. Both identify active/default-bootstrap JSONL questions, expected page IDs and query types. The shared implementation witness assigns the page stem to each nonempty generated line; the harness later consumes expected IDs. The epistemic relevance-assertion finding and memory's unresolved initialization history concern the same dataset objects and coexist. Bootstrap and retrieval-measurement findings retain the distinction between candidate label creation, possible active-path writing and actual later use.

Full identity is ruled out for EPI-OBJ-raw-source and either MEM-OBJ-raw-source or MEM-OBJ-raw-meta alone. The epistemic record explicitly includes source bytes plus meta.json; memory separates the evidence text from the symbolic sidecar whose existence suppresses re-storage. These are a valid container and its distinct parts, not competing accounts of one indivisible object. The byte-write passage supports their overlap, while the metadata early-return passage supports the separation. No split supersession is needed: the container does not claim that metadata itself is source prose or that the loader verifies retained bytes.

Full identity is likewise ruled out between EPI-OBJ-page-metadata and either MEM-OBJ-page-routing or MEM-OBJ-page-sources alone. The epistemic container includes typed fields, source pointers and relations. Memory distinguishes namespace/currentness/graph selectors from span and citation pointers, with different consumers and omissions. The container remains valid alongside those parts; it neither supplies semantic endorsement nor says that all metadata reaches later consumers intact.

EPI-OBJ-eval-results and MEM-OBJ-eval-results overlap at saved retrieval reports but are not identical containers. The former includes ingest measurement on stdout; the latter groups persisted baseline/component reports with the index build manifest. The epistemic record's containment in RT-RTE-eval-report does not extend to the build manifest. Neither container's findings require replacement when these scopes are kept separate. MEM-OBJ-ingest-manifest, MEM-OBJ-audit-log and the lexical, graph and vector objects (MEM-OBJ-bm25, MEM-OBJ-graph, MEM-OBJ-vectors) have no matching independent declarations in the other members.

Route containment rather than duplicate replacement

The epistemic subroutes do not supersede runtime's end-to-end routes. Within RT-RTE-source-ingest, EPI-RTE-raw-import, EPI-RTE-page-extraction, EPI-RTE-quote-location, EPI-RTE-proposal-selection and EPI-RTE-pending-retention distinguish import, generation, checking, disposition and retention. Their declared containment preserves the runtime route's progression. EPI-RTE-schema-check and EPI-RTE-span-check span several runtime routes and standalone or measurement consumers; neither is identical to the ingest route or exclusively contained in it.

EPI-RTE-review-disposition and EPI-RTE-knowledge-retention are decision and move functions within RT-RTE-page-review. EPI-RTE-query-selection is the selection/consumption function within RT-RTE-index-query, not a replacement for its rebuild-and-query account. EPI-RTE-ingest-measurement and EPI-RTE-retrieval-measurement separate functions within RT-RTE-eval-report. EPI-RTE-query-bootstrap isolates content transformation within RT-RTE-query-bootstrap; its similar name does not make the function identical to the enclosing command route. No route supersession follows from these containments.

The workflow member declares no independent records. Its annotations add retry, human-handoff and terminal-display consequences to the same runtime routes. Memory likewise annotates those routes rather than declaring replacements. The epistemic claims about EPI-CLM-faithfulness-support, EPI-CLM-eval-gates and EPI-CLM-provenance-trust are distinct from runtime's RT-CLM-provenance-policy and RT-CLM-history-policy: they concern numerical support, promotion requirements and later reliance respectively. Shared source themes do not establish claim identity.

Independent convergence and integration dispositions

Proposal selection and retry. Runtime's RT-RTE-source-ingest, memory's annotation, workflow's annotation and epistemic EPI-RTE-proposal-selection agree: first nonempty all-valid output wins; otherwise the final attempt supplies the pages considered for staging. Workflow makes explicit that a valid subset from an earlier mixed attempt is not merged into escalation. Memory adds first-ID deduplication and the absence of critique delivery to the retry. These are compatible consequences of the same predicates, not conflicting acceptance rules. Manifest “accepted” remains pending eligibility, not knowledge acceptance.

Admission and warning delivery. RT-RTE-page-review, RT-RTE-curated-seed, memory's write-side account and epistemic EPI-RTE-review-disposition converge on path-specific controls. The interactive decision is not a universal human authorization boundary; direct acceptance and checked seed writes differ. Edited acceptance has no pipeline revalidation, and the optional hook acts after movement. Memory's integration issue concerning MEM-OBJ-ingest-manifest is a delivery limitation, not a correction: ingestion prints changed-body warnings, but review enumerates Markdown rather than consuming the manifest. Workflow says the CLI reports those flags at ingestion, not that the later reviewer necessarily sees them. No disagreement follows.

History and withdrawal. Runtime's RT-CLM-history-policy, memory's MEM-OBJ-page-routing, epistemic EPI-RTE-knowledge-retention and the workflow account agree that same-name moves or seed writes can replace content without automatic reciprocal supersession updates. Current-load exclusion and physical deletion are separate effects. This is convergence about a doctrine/implementation distinction, not a conflict between analyst reports.

Read-back, currentness and lineage. Memory's index-query annotation, RT-RTE-index-query, epistemic EPI-RTE-query-selection and workflow's query annotation agree that returned dictionaries contain IDs, scores, paths and source IDs, not page bodies or an answering prompt. Current-page loading does not universally remove stale fused IDs; active rerank filters candidates, while fallback does not. Memory's provenance-omission issue is also compatible with the epistemic trust comparison: MEM-OBJ-page-sources retains spans that query output reduces to IDs, while bootstrap supplies body text without the source array. Stored lineage cannot be projected unchanged into either channel.

Distinct evaluation consumers. Memory's eval-report annotation clarifies rather than contradicts RT-RTE-eval-report. Direct baseline evaluation loads current corpus text without persisted indexes; component adapters use index-backed query paths. Ingest evaluation has a different file/raw-availability selection and no current-page filter. Thus query freshness limitations do not automatically describe the baseline loader, and a baseline current-page filter does not establish currentness for every evaluation route. The epistemic EPI-RTE-retrieval-measurement bounds metric authority to dataset-label agreement without requiring those consumers to be identical.

Lexical support and generated labels. RT-CLM-provenance-policy already distinguishes marker/overlap checks from all-assertion truth. EPI-RTE-ingest-measurement adds the unmarked-body maximum-score countercase; it does not contradict runtime's narrower lexical account. RT-RTE-query-bootstrap, MEM-OBJ-eval-queries and EPI-RTE-query-bootstrap agree that requested review is not enforced for every output path and caller-selected active-dataset writing can reach the evaluator without a separate merge. This establishes wiring, not semantic label validity or observed use.

Operation and improvement limits. Runtime and the specialists distinguish wired mechanisms from unobserved operation. Epistemic “implemented” architectural assessments do not upgrade runtime's operation status. Memory's retained audit/report objects and workflow's recovery account supply no automatic durable policy update or transactional rollback. The epistemic bounded conclusion keeps learning, reflection, autonomy and self-improvement independently unestablished on the named routes rather than deriving one global negative from missing execution evidence.

Conflict disposition

No unresolved cross-report disagreement is identified in these pinned members. No correction answers or declined objections were supplied for this attempt. The memory integration questions above remain either resolved cross-report distinctions or shared evidence limits: dataset initialization, seed-specification correctness, most corpus content, retained raw-byte lineage, external model artifacts, actual review/adoption and downstream answering use are not established. Those uncertainties prevent the corresponding trust, corpus-quality, exercised-use and benefit conclusions; they do not contradict the known transformations. No additional required undeclared part is identified by these connections. The supersessions above state identity only and change no analyst-owned finding or status.