Cortex reconciliation

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

Reconciliation

Duplicate identities

Amendment: EPI-OBJ-note is superseded by MEM-OBJ-notes. Both identify the private repository's retained Markdown bodies, including imported notes, consumed through corpus and direct-context routes. Their evidence is SRC-1, scripts/ingest.mjs and lib/brain.ts, including the import warning retained by EPI-RTE-note-import. The epistemic inventory and admission ledger concern the same retained content, not a second store. This identity does not confer known source warrant or establish the private propositions actually stored; the memory record retains unresolved initial authorship and distinct parser-consumed metadata.

Amendment: EPI-OBJ-note-edges is superseded by MEM-OBJ-note-edges. Both identify Postgres note_edges rows, symbolic endpoints/kinds/weights and human-readable evidence consumed by handoff and diagnostics. The declarations and EPI-RTE-edge-rebuild retain SRC-1 evidence from lib/edges.ts, lib/handoff.ts and supabase/migrations/20260909032252_structural_edge_identity.sql. The memory declaration additionally covers freshness state and read-path limits; that detail does not change the shared row referent. Epistemic count, retention and ranking assessments remain separately scoped functions, not upgraded semantic-relevance claims.

Amendment: EPI-RTE-retirement-nomination is superseded by MEM-RTE-retirement-nomination. Both trace the same scripts/lifecycle-sweep.ts optional --propose request, propose_deletions RPC, nomination storage and undecided-candidate report through lib/lifecycle.ts. Both retain the passage that the CLI does not delete notes. Memory additionally inspected the SQL predicate and score-view interaction. Epistemic uncertainty about eligibility reflects its narrower inspection, not a contrary predicate. This supersession identifies the route without converting its implementation wiring into observed nominations, human decisions or removal.

EPI-RTE-note-import and EPI-RTE-description-backfill describe two mechanisms within MEM-RTE-local-curation. The latter also covers project-page splitting and subsequent publication requirements. These are valid narrower routes beside a grouped account, not interchangeable identities. The supplied findings do not require splitting away or superseding the container. They are also distinct from RT-RTE-trusted-write: local script admission bypasses the MCP writer, and local file changes need publication before production consumption.

EPI-RTE-edge-rebuild overlaps MEM-RTE-usage-selection but is not identical. The former follows graph publication and its consumers; the latter includes raw event acquisition, score-view transformation, router bands and pins as well as graph selection. MEM-OBJ-note-access, MEM-OBJ-note-scores and the edge object are distinct stages, not aliases. MEM-BAP-selection-metadata identifies their deterministic consumer channel, unlike the model-interpreted RT-BAP-answer-instructions.

EPI-OBJ-reader-reply includes parsed and rendered current-response state. MEM-OBJ-cached-answer identifies retained rendered replies and replay metadata. Their rendered text overlaps, but neither whole object replaces the other. EPI-RTE-answer-cache, EPI-RTE-answer-stamp and EPI-RTE-quote-check explicitly identify separate parts of RT-RTE-reader-answer. The complete route remains valid beside those parts; its guest specialization, RT-RTE-guest-mediation, additionally owns scope, grants and spending controls.

MEM-CLM-evaluation-gate and EPI-CLM-prediction-gate converge on the prediction adoption rule but are not identical claims. Memory compares prediction and graph-hop alternatives with serving consumers; epistemic isolates the prediction-policy statement and attributed historical verdict. EPI-RTE-prediction-evaluation, EPI-RTE-retrieval-evaluation and RT-RTE-reader-evaluation have different targets, reference outcomes and outputs. None supersedes another or establishes automatic successor installation.

Convergence and integration dispositions

The reports independently retain the narrow license of RT-CLM-quote-verification: quote occurrence is not answer entailment or source truth. The epistemic checking and stamping parts explain why a visible answer can remain unproven; the memory cache object explains why replay adds no fresh check. Direct delivery through RT-RTE-context-delivery is not subject to Ask's verdict. No conflict follows from those different consumer paths.

Memory's integration issues refine context coverage without contradicting the runtime. MEM-OBJ-router describes month aggregation, cold omission and section clipping after the runtime's router candidate set. MEM-OBJ-mirror and the edge record preclude treating a corpus SHA as a snapshot guarantee for graph and bubble state. Runtime already retains separate-input limits. These findings prevent full-body, complete-router and universal-freshness inferences; no inconsistent live delivery is established.

The lifetime-read wording issue is source commentary versus the bounded score view, not an inter-report disagreement. Memory's retirement record establishes that reads=0 uses a 180-day aggregation; epistemic leaves that predicate unassessed. Likewise, memory's annotation on the trusted-write route establishes a console reverify-request write but not its consumption by RT-RTE-groundskeeper-maintenance. No report establishes that extra consumer guarantee.

The maintenance-mode filter concerns recorded labels, not authenticated work origin. Memory's usage-selection record notes that direct reads default to read mode and the template does not explicitly request maintenance=true. Epistemic's edge-rebuild exclusion of maintenance applies to tagged rows; it cannot be extended to every actual maintenance session. Neither report supplies evidence of contaminated production data.

Workflow's groundskeeper annotation adds health-failure report-only delivery, conditional digest/no-write branches, non-atomic completion and the dated-versus-undated absorption-marker mismatch to the runtime route. Memory describes requested absorption markers and epistemic describes requested integration, neither asserting exact-match replay suppression. The mismatch therefore limits a literal completion guarantee without contradicting their bounded accounts. RT-BAP-maintenance-instructions supplies procedural force, not server-wide enforcement. All three specialist accounts distinguish the template's human-only proposal decisions from the broader trusted authority of RT-RTE-proposal-admission.

EPI-RTE-watch-checks, EPI-OBJ-watch-items and EPI-CLM-correction-chain add a diagnostic consumer and its explanation-versus-predicate limit. Memory leaves the full inbox inventory unassessed; it offers no incompatible semantic-contradiction finding. These additions are not aliases of proposals, retirement candidates or accepted note content.

No unresolved disagreement between the supplied reports is established. No correction answers or declined objections were supplied. Their shared absence of run evidence, together with local inspection limits, prevents conclusions about exercised maintenance, improved capacity, accepted ampliative production or exercised self-improvement; these uncertainties are not conflicts. No frozen-source reads were needed for this reconciliation, and no analyst findings or statuses have been replaced by an integration verdict.