LLM Wiki v2 reconciliation

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

Reconciliation

Boundary and evidence alignment

The four reports share the frozen revision and the knowledge-system boundary in boundary. They distinguish inspected implementation, shipped instructions and attributed benchmark results from observed operation and causal support. The workflow selects classic filing followed by a separately requested question; it does not narrow the other reports' classic/CROW, maintenance, relay or evaluation coverage. No analyst answers or refusal were supplied. This reconciliation uses only the supplied members; no source reread was needed.

Record identity and containment

No whole-record supersession is established. The following overlaps do not make the records interchangeable:

• MEM-RTE-source-reduction is an explicitly declared part of RT-RTE-source-filing. The latter retains preparation, admission, persistence and enrichment beyond reduction. EPI-RTE-merge-guard, EPI-RTE-guard-choice, EPI-RTE-revision-check and EPI-RTE-revision-admit declare other parts, separating evidence computation from its consequential consumer. None replaces the whole filing route.
• EPI-OBJ-note spans candidate/draft bodies and stored main bodies. MEM-OBJ-note-body describes the retained main-body portion and its later consumers. The candidate-to-stored container remains useful and is not wrong merely because the retained portion has its own record. Whole-record identity is ruled out; neither record supersedes the other.
• EPI-OBJ-source concerns assertions in supplied Source.text and raw bodies. MEM-OBJ-raw describes persisted extracted text together with origin, sequence, hash and removal metadata. Imported assertion content overlaps, but the input/assertion scope and the persisted mixed-content object are not identical.
• EPI-OBJ-answer is a transient output object, not RT-RTE-question-answer, its producing route, or MEM-OBJ-note-body, a retained input. MEM-RTE-note-read delivers a requested note without the question route's relevance selection or answer generation; these are alternate consumers, not duplicate routes.
• MEM-OBJ-answer-cache and MEM-OBJ-replay retain outputs rather than the parameters described by RT-CMP-classifier and RT-CMP-llm. Their keys, covered model kinds and persistence differ. MEM-RTE-answer-reuse connects the two reuse implementations without identifying their stores.
• MEM-BAP-retained-knowledge concerns delivered retained evidence. RT-BAP-model-instructions concerns rendered role guidance; RT-BAP-client-instructions asks an external client to use memory and follow relay restrictions. A common model consumer does not make these channels or authority sources identical.
• EPI-RTE-route-audit and EPI-RTE-audit-apply distinguish classifier proposals from consequential moves. MEM-OBJ-proposals is instead the durable caller-supplied area queue. EPI-RTE-health-check consumes log counts, whereas the log's split-attempt consumer in MEM-OBJ-log suppresses later attempts. Shared storage or move primitives do not establish route identity.

Convergence and integration-issue dispositions

Delivery is not semantic acceptance. The workflow's annotations use the runtime's existing filing and question records rather than declaring duplicates. Its structural rejection, post-write relating failure and caller-owned retry account agrees with RT-RTE-source-filing. The epistemic check/disposition parts add consequential heuristic checks without turning the runtime's factual-preservation claim into an enforced guarantee: see RT-CLM-factual-preservation, EPI-RTE-step-check and EPI-RTE-guard-choice. These are complementary scopes, not conflicting statuses.

Full text means the delivered display, not the serialized file. Memory integration issue 1 qualifies runtime and workflow shorthand. MEM-OBJ-note-body retains the full_text construction; MEM-OBJ-note-metadata identifies omitted control fields and See-also links. MEM-RTE-note-read and RT-RTE-question-answer deliver that display through different routes. Neither the runtime nor workflow requires that all frontmatter or linked targets reach the answering writer. No cross-report disagreement follows.

Removal, purge and rebuild are different operations. Memory integration issue 3 extends the runtime's statement that removed-source notes remain. MEM-OBJ-raw and the memory maintenance annotation describe marking, reconstruction and mixed-source refiling; EPI-RTE-origin-purge is a declared withdrawal part of RT-RTE-manual-maintenance. Their conclusions agree that lineage-based deletion does not establish universal semantic erasure or rollback. The memory annotation's placement under maintenance does not relocate the epistemic revision parts out of filing.

Evaluation is an offline consumer, not an online oracle. Memory integration issues 2 and 6 connect MEM-OBJ-evaluation-results, MEM-OBJ-replay and EPI-RTE-fact-score. Stored results, response substitution and lexical scoring are different referents. The runtime's retrieval-evaluator account tests selected-note identity without generating an answer. EPI-CLM-benchmark-results retains attributed aggregate results and the unavailable contradiction/revision harness. These limits reinforce, rather than contradict, memory's narrower evaluator coverage. None supplies observed factual correctness or component-level causal attribution.

Caller and provider uncertainty remains local. Memory integration issues 4 and 5 leave no whole-route duplicate and preserve unresolved authorship and approval provenance. Runtime relay and maintenance findings do not depend on a verified human principal: see RT-RTE-client-relay and RT-RTE-manual-maintenance. Provider identity limits remain in the component records. The epistemic bounded conclusion and workflow's lack of an autonomous critic-author cycle do not negate implemented storage, checks or read-back, and do not establish absent internal reasoning.

Unresolved record separation

Unresolved conflict: EPI-OBJ-check-results groups value sets, probabilities, route proposals, health records and fact-match booleans, while MEM-OBJ-evaluation-results groups retained evaluator outputs. The epistemic container explicitly assigns different warrant and consequences through EPI-RTE-merge-guard, EPI-RTE-revision-check, EPI-RTE-conflict-check, EPI-RTE-route-audit, EPI-RTE-health-check and EPI-RTE-fact-score. Their retained passages locate the differing producers and consumers in SRC-1 src/llmw2/agents/steps_classic.py, src/llmw2/agents/revisions.py, src/llmw2/agents/researcher.py, src/llmw2/wiki.py, src/llmw2/audit.py and benchmarks/note_facts.py. The routes separate these functions, but corresponding operative-object declarations for the individual check-result parts are missing. This prevents complete part-level reconciliation of transient check results versus retained metric instances; it does not establish contradictory evaluator findings. The container cannot be superseded by undeclared parts, and reconciliation allocates no IDs. The verifier must address the required separation to the declaring analyst rather than infer a common warrant, retention horizon or authority from the container.