Cortex reconciliation

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

Reconciliation

Boundary and basis

All four reports use the revision and service boundary in boundary. The runtime supplies the request, storage, maintenance and transport baseline; memory separates retained parts and later consumers; epistemic separates generation, checks, disposition, retention and reliance; workflow annotates one connected capture request. None supplies observed target execution or a causal comparison. This reconciliation uses the pinned members only; no source reads or analyst correction answers were supplied or undertaken. Record passages cited below supply the source evidence, under SRC-1 in the Source register.

Duplicate supersessions and distinct identities

Amendment: EPI-OBJ-summary is superseded by MEM-OBJ-summary. Both identify the same retained natural-language summary field, its caller/model/fallback origins, existing-summary reuse and briefing consumption. Identity evidence is SRC-1, src/cortex/pipeline/normalize.py and src/cortex/retrieval/presenters.py, with the existing-summary and graph-summary passages in the epistemic declarations and the projection passage supplied by RT-RTE-retrieve-context. The epistemic inventory and generation/admission ledger refer to this same field. Their fidelity and acceptance qualifications remain; supersession does not replace either report's findings or make the summary independently warranted.

Amendment: EPI-RTE-access-adaptation is superseded by MEM-RTE-access-promotion. Both trace read/positive feedback to durable counters, the ten-access threshold, a possible reflex-tier write and later tier presentation. Their threshold/write passages identify the same implementation in SRC-1, src/cortex/retrieval/learner.py. The affected epistemic adaptation ledger and retrieval-learning comparison retain their distinction between metadata change, truth license and improvement. Memory additionally traces later enrichment and rendering order; that detail does not create another automatic access pathway. Neither route is identical to its valid parent RT-RTE-retrieve-context, which also covers retrieval, selection and delivery.

Amendment: EPI-RTE-state-diagnostics is superseded by MEM-RTE-issue-delivery. Both identify requested AdvancedReasoner/AlertPresenter findings delivered through reason/status, without automatic persistence, correction or withdrawal. Identity evidence is their shared structural-check and status-alert passages from SRC-1, src/cortex/pipeline/advanced_reason.py and src/cortex/transport/mcp/server.py, together with their presenter findings in src/cortex/retrieval/presenters.py. The epistemic ledger's separate structural-check and conjectural-interpretation functions concern this same diagnostic route; those interpretations remain attached to their original declaring report. This is not identity with RT-RTE-maintain-knowledge, whose scope also includes mutation and recovery, or with retrieval as a whole.

The remaining apparent object matches do not establish whole-record identity:

  • EPI-OBJ-document-text covers title/content assertions, whereas MEM-OBJ-documents also distinguishes raw_markdown and its branch-dependent source preservation. These overlap within RT-OBJ-store, but their full referents differ. Neither replaces the valid store container. Evidence: SRC-1, src/cortex/db/store.py and src/cortex/pipeline/importer.py, including the memory record's raw-markdown branch passages.
  • EPI-OBJ-graph-assertions isolates predicate/endpoint propositions. MEM-OBJ-relationships also covers graph objects, entities, mentions and descriptive properties. The assertion subset is not identical to that broader graph part, and neither is RT-OBJ-inference-rules: data/products and shipped query rules differ. Evidence: SRC-1, src/cortex/pipeline/link.py and src/cortex/pipeline/reason.py in those records.
  • EPI-OBJ-classification groups type/project/entity associations, confidence, tier and access configuration. MEM-OBJ-tier, MEM-OBJ-access-count and the graph part above expose different portions and consumers. No one memory record is identical to the combined epistemic referent. The memory parts and epistemic route distinctions permit the grouping to remain without treating all its fields as one authority or update mechanism. Evidence: SRC-1, src/cortex/services/llm.py, src/cortex/pipeline/enrich.py and src/cortex/retrieval/learner.py in the cited declarations.

EPI-OBJ-diagnostic-result is the transient output, not the diagnostic route. EPI-OBJ-period-narrative is likewise the transient synthesis output, not MEM-OBJ-summary, a retained input. Their passages in SRC-1, src/cortex/retrieval/presenters.py distinguish return from persistence. EPI-RTE-period-synthesis is a part of the broader runtime retrieval route, not a duplicate of that whole route.

The generation and structure-check parts EPI-RTE-metadata-generation and EPI-RTE-structure-check do not replace RT-RTE-capture-pipeline. EPI-RTE-metadata-admission spans pipeline and explicit classify writes, rather than duplicating either complete runtime route. EPI-RTE-rule-derivation and EPI-RTE-rule-retention distinguish query products from persistent additions within maintenance. Their source passages in SRC-1, src/cortex/services/llm.py, src/cortex/pipeline/link.py, src/cortex/pipeline/normalize.py and src/cortex/pipeline/reason.py identify different functions, not interchangeable routes. No split supersession of the valid runtime containers is required.

EPI-RTE-health-check and EPI-RTE-health-repair add direct doctor paths. Memory's maintenance annotation independently mentions repair, but declares no duplicate route. Checking and repair are not identical to each other or to runtime's combined maintenance/recovery route. Evidence: their distinct doctor invocation and repair passages in SRC-1, src/cortex/cli/main.py.

EPI-CLM-retrieval-learning overlaps the learning portion of MEM-CLM-learning-scope, but the latter also assesses proactive-alert doctrine. These are not identical complete claims. EPI-CLM-semantic-checks isolates a docstring claim only discussed inside runtime maintenance; EPI-CLM-formal-intelligence concerns product doctrine, not the runtime stage-resilience or store-consistency claims. Their respective passages in SRC-1, src/cortex/retrieval/learner.py, src/cortex/retrieval/presenters.py, src/cortex/pipeline/advanced_reason.py, llms.txt and README.md rule out whole-claim identity. No other duplicate declarations are established; workflow declares none.

Convergence and integration dispositions

Admission and completion. RT-RTE-capture-pipeline, memory's capture annotation, the epistemic generation/check/admission records and workflow converge on ingestion before optional processing. Existing summaries and disabled processing are alternate paths, not conflicting descriptions. Workflow adds caller-entity precedence, endpoint validation limits and the initial capture-type return. These qualify handoff guarantees without contradicting runtime's structural-gate account. The collected-status meaning of complete and partial remains distinct from semantic acceptance and storage consistency; see RT-CLM-stage-resilience and RT-CLM-store-consistency.

Freshness and recovery. Memory's summary/vector freshness and historical-erasure issues connect MEM-OBJ-summary, MEM-OBJ-embeddings and MEM-OBJ-document-versions to runtime maintenance. Workflow's reprocessing account agrees that rerun is not guaranteed reclassification or replay of original entity arguments. These are qualifications of separate retained parts, not evidence that runtime promised universal refresh, historical accuracy or erasure. Query trails and access keys are separately retained, as MEM-OBJ-query-log and MEM-OBJ-access-count establish.

Learning and diagnostics. MEM-ABS-adaptation-wiring gives a bounded searched absence of a shipped ranking-adaptation bridge. Epistemic's lack of established tuning and runtime's constructor-default account do not oppose that absence or strengthen it to all external integrations. Tier adaptation remains wired; operation and benefit remain unobserved. Memory's three different staleness predicates and epistemic's structural-versus-explanatory diagnostic distinction concern different tests and output meanings, not a unified expiry policy. Semantic-contradiction and proactive labels are source-doctrine discrepancies recorded consistently across the reports, not disagreements between analysts.

Authority and evidence. RT-BAP-knowledge-delivery bounds host-facing informational output; RT-BAP-rule-execution bounds executable rule authority. Memory and epistemic preserve that difference. Conditional derivation and retained triples do not establish premise truth, criticism of the rules, or a criticism-to-successor chain. Runtime's separately assessed theory-builder properties are not contradicted by epistemic's narrower absence of established acceptance in its inspected routes. Memory's evaluator-design qualification and epistemic's benchmark discussion both preserve the boundary's lack of measured results.

No unresolved cross-report disagreement is established by these pinned inputs. No dependent finding was found requiring a supplied record to say something it no longer says. The memory integration issues remain the bounded limits above; the epistemic coverage gaps remain local limits rather than system-complete negatives. There are no supplied correction or decline dispositions to adjudicate. No required missing part is identified beyond the parts already declared and connected here; this does not assert an exhaustive source audit.