Secure LLM-Wiki reconciliation

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

Reconciliation

Duplicate supersessions and distinct referents

Amendment: EPI-OBJ-claim-text is superseded by MEM-OBJ-claim-text. Both identify Claim.text: extracted natural-language candidate text retained in page/quarantine YAML-frontmatter Markdown, reloaded from frontmatter, and delivered to later consumers. Both place it within RT-RTE-ingest-admission. Identity evidence is the parser's text=item["text"] passage in RT-RTE-extraction-parsing, together with the serialization/deserialization passages in the memory object. The supersession connects the epistemic inventory, ledger and bounded conclusion to the memory object's storage and consumer account; it does not convert candidate text into verified propositions. The epistemic report's indeterminate extraction/answer relations remain its own findings.

Amendment: EPI-OBJ-claim-embeddings is superseded by MEM-OBJ-claim-embeddings. Both identify UUID-keyed dense vectors in memory and embeddings/<claim_id>.json, not provider weights. Both retain the same JSON-write passage and describe reload without a claim-status/trust join or model-version marker. The affected epistemic inventory, similarity ledger and retention findings concern the same numerical artifact the memory report traces into future admission. Neither record establishes vector compatibility, semantic correctness or observed effects. The producer RT-CMP-embed-model is not a duplicate of its output.

No supersession joins EPI-OBJ-claim-metadata with either MEM-OBJ-claim-metadata or MEM-OBJ-gate-records. The epistemic record is a valid broader container for control/lineage fields and diagnostic notes. Memory separates eligibility/provenance from gate progress and review diagnostics because their consumers differ. The combined record already distinguishes symbolic fields from natural-language notes and says those notes are not query-consumed policy. Its findings do not become wrong when the memory parts are separated; a split supersession is therefore not warranted. The memory account supplies the finer distinction: gate names are rendered to query, while review reasons and body-only detail are not.

Other overlaps are containment or cross-stage functions, not identity. EPI-RTE-security-checks and EPI-RTE-gate-disposition separate checking from consequential disposition inside RT-RTE-ingest-admission. They do not replace that full route. EPI-RTE-store-retention also covers direct storage outside ordinary CLI admission; EPI-RTE-text-render crosses ingestion and query. Neither is identical to either enclosing runtime route. EPI-OBJ-source-text and EPI-OBJ-query-answer are input and output parts, not duplicates of routes or retained claim text.

MEM-OBJ-trust-rules is retained policy state within RT-RTE-operator-maintenance, not the maintenance route itself. MEM-OBJ-wiki-history is an operation-trace artifact spanning admission and maintenance, not another claim store or automatic recovery consumer. MEM-ABS-query-feedback makes a bounded absence explicit; it does not supersede the positive RT-RTE-context-query route. Workflow declares no additional records; its annotations add workflow roles to these runtime routes.

The claims are also distinct. EPI-CLM-compiled-knowledge concerns persistent compilation and curation; RT-CLM-security-boundary concerns trust-boundary protection; EPI-CLM-quality-evidence concerns the warrant requested for selected context. The last is not identical to the policy artifact RT-OBJ-system-prompts or its consumer path RT-BAP-query-evidence. Sharing a passage does not make a claim, its carrier and its authority path one referent. No other cross-report record identities are established.

Independent convergence and integration dispositions

The workflow account independently traces the same durable handoff as runtime and memory: operator-selected source, extraction proposals, batch review, per-claim code admission, stored claims, then a separately invoked query. Its batch-review reuse and sequential existing-state updates elaborate coordination rather than contradicting RT-RTE-ingest-admission. Models propose or veto through text; Python owns progression and filesystem effects. Escalation ends in quarantine rather than awaiting an approver.

The reports agree on the consequential alternate paths: parser compatibility branches do not uniformly verify extraction nonces; direct storage bypasses ordinary admission; the actual loader can include pending pages while strict builders cannot; all shipped context paths retain an untrusted-label floor. RT-RTE-extraction-parsing, EPI-RTE-store-retention, RT-RTE-context-query and memory's annotations describe these as different control boundaries, not demonstrated live exploits. The epistemic rendering deduction adds a semantic-preservation limit to the marker stripping already described by runtime and memory; neither had asserted universal preservation.

Memory's integration issues require no factual correction to runtime. Its unresolved deployed authorship, effective trust and vector lineage questions remain local evidence limits. Its rationale issue connects MEM-OBJ-gate-records to the epistemic distinction between security admission and factual acceptance: retained failure reasons do not establish rationale-informed automatic learning. Its stale-vector warning agrees with EPI-RTE-security-checks, whose numerical-domain countercase adds detail without claiming observed mismatches.

Operational admission in RT-RTE-ingest-admission does not conflict with the lack of factual acceptance in EPI-RTE-gate-disposition. Their criteria and intended licenses differ. Likewise, evidence-labelled delivery through RT-BAP-query-evidence is compatible with unknown source warrant and unverified answer quality. Query output without persistence agrees across workflow, runtime and MEM-ABS-query-feedback; no report treats this bounded absence as proof about opaque provider learning.

All members remain within the implementation and doctrine/design evidence registered in the boundary. Wired delivery, uninspected operation and claimed warrant are separate findings, not statuses to rank or merge. No observed-run or causal evidence resolves model obedience, injection resistance, factual correctness or benefit. These are shared limitations, not disagreements. No cross-report disagreement was identified in the supplied versions, and no correction answers or declined objections were supplied. No source read was needed for reconciliation.