Cortex report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The pinned reports support a bounded code-grounded account. I read all five reviewed members, the boundary, their types, the record contract and the supplied structural check. The structural check reports none; the four answer files are empty. There are no correction arguments to accept by default. I independently read source rather than treating analyst check statements as verification. Checkout HEAD equals the reviewed revision. No target code, tests, models or services were executed.

Records and consequential behavior

I checked every declaration's referent, source anchor, evidence layer, status and applicable authority/admission fields, every Part of: relation, and the reconciliation's three supersessions. The following groups summarize their substantive dispositions, including the annotations and prose depending on them.

Records and passages	Independent check and disposition
Runtime capture/direct-import routes, stage-resilience claim, workflow account, epistemic metadata generation/check/admission	SRC-1 src/cortex/pipeline/orchestrator.py, src/cortex/pipeline/normalize.py, src/cortex/pipeline/link.py, src/cortex/services/llm.py, src/cortex/pipeline/importer.py, src/cortex/transport/mcp/server.py and CLI capture/routing bodies support ingestion before processing, caller-summary bypass, explicit entity precedence, structural relationship gates and continuation after failed stages. The top-level capture type is initialized before normalization. Reprocessing supplies no original entity argument. Local suppression and fallback explain why complete is not semantic acceptance. The workflow's connected request-to-result account and rejection/recovery qualifications hold. See RT-RTE-capture-pipeline, RT-RTE-direct-import, RT-CLM-stage-resilience, workflow account, and epistemic records.
Runtime store, all memory object parts and corresponding epistemic objects	SRC-1 src/cortex/db/store.py, schema/update/delete/embedding/config/query methods in src/cortex/db/content_store.py, and src/cortex/pipeline/temporal.py support separate text, summaries, symbolic metadata/graph, vectors, counters, logs, weight configuration and historical rows. Raw Markdown differs between import branches. Snapshot-before-update is conditional on initialized temporal support; direct backing-store writes bypass it. Current-row deletion is not complete erasure. Summary reuse and missing vector regeneration on edit support the freshness qualifications. The transient diagnostic and narrative objects are outputs, not retained summary fields. See RT-OBJ-store, memory objects, epistemic objects, and RT-CLM-store-consistency.
Retrieval/delivery, promotion, learning claims and searched absence	Independent engine/presenter/server/learner reads establish ranking inputs, bounded candidate sets, summary projection, richer ID read, separate dossier/synthesis selectors, query logging and access-to-tier writes. Tier is displayed and preserved by enrichment, not used by the hybrid score. Source-wide helper/caller searches corroborate the missing shipped weight/miss/demotion bridge; direct CLI read also records access. The absence is correctly bounded to shipped wiring, not external library consumers. See RT-RTE-retrieve-context, RT-BAP-knowledge-delivery, MEM-RTE-access-promotion, MEM-ABS-adaptation-wiring, MEM-CLM-learning-scope, and EPI-CLM-retrieval-learning.
Maintenance, rule object/authority, epistemic derivation/retention/diagnosis/synthesis/health routes and associated claims	Complete reasoner, advanced-reasoner and presenter reads plus doctor bodies support conditional graph entailment, additive retention, advisory diagnostics, transient model synthesis and deterministic count fallback. Query failure can look like zero pending inference. Semantic contradiction comparison is docstring doctrine, not implemented by detect_contradictions. Pattern explanation exceeds its frequency evidence; the reports retain that distinction. Separate staleness predicates do not establish a universal invalidation rule. SRC-1 src/cortex/pipeline/reason.py, src/cortex/pipeline/advanced_reason.py, src/cortex/retrieval/presenters.py, src/cortex/pipeline/enrich.py, src/cortex/cli/main.py. See RT-RTE-maintain-knowledge, RT-OBJ-inference-rules, RT-BAP-rule-execution, MEM-RTE-issue-delivery, and epistemic ledger.
Model components, transport controls and install/release route	SRC-1 src/cortex/services/llm.py, src/cortex/services/embeddings.py and src/cortex/core/config.py support inference adapters without exact artifact pins or inspected parameter updates. Server registration, REST key verification and dashboard auth helpers support only path-specific controls; delete consumes no safety approval. Restore and systemd installation bodies plus release dependencies support attempted recovery and delivery wiring, not successful deployment or global atomicity. See runtime components, RT-RTE-transport-control, RT-RTE-install-release, and RT-RTE-maintain-knowledge.

Source anchors above were checked at SRC-1. Additional checks covered llms.txt, the README claim passages, MCP client session/call handling, facade call sites, deploy/cortex-mcp.service, .github/workflows/release.yml, benchmark fixtures and evaluator assertions. They support doctrine, implementation and evaluator design only. Synthetic embeddings, fixture membership and asserted F1 are not observed outcomes. The epistemic classification of model transformations as indeterminate is supported: source requests alone establish neither semantic preservation nor non-entailment. Formal-rule consumption supports localized content and executable dependence, not criticism or iteration of the rules themselves. The independently scoped learning, reflection, autonomy and benefit qualifications remain warranted.

Relations and reconciliation

All declared store parts are contained in the runtime store. Generation/checking, derivation/retention, synthesis and access-route parts identify narrower functions within their stated parents; records spanning pipeline and maintenance correctly avoid pretending to have one parent. Coarser overlapping objects remain valid containers rather than competing canonical values.

The reconciliation preserves the findings attached to the summary, access and diagnostic supersessions. The source implementations establish the same referents on each side of those amendments; the different lens fields do not make separate mechanisms. Document text versus raw Markdown, graph assertions versus the broader graph part, and combined classification versus individual metadata parts are correctly retained as overlaps rather than full identity. I found no conflicting analyst positions requiring an unresolved-conflict paragraph. Source-doctrine mismatches are consistently qualified, not silently settled as stronger implementation claims.

Omission questions and stopping decision

The frozen top-level tree and producer/consumer searches guided three consequential questions: whether shipped callers bypass the described admission/recovery gates; whether another shipped caller closes automatic ranking adaptation or host injection; and whether inference/diagnostics/synthesis confer warrant beyond structural matching and retention. CLI/direct/import, facade/server, learner/engine, prompt and reasoning traces support the existing qualifications. Shipped callers were inspected as Cortex responsibilities, not dismissed as external hosts.

Further optional expansion stopped after those questions had supported dispositions. No identified material contradiction remains without disposition. Detailed UI, evaluator and provider investigation would not establish observed operation from this source-only boundary. The two local inventory limits below do not invalidate the positive capture, retrieval, tier-adaptation or conditional-inference findings; they prevent treating those inventories as exhaustive.

Blockers

none

Limits

• The import route is represented coarsely in RT-RTE-direct-import and the epistemic coverage account. SRC-1 src/cortex/pipeline/importer.py additionally resolves wiki-link titles by exact, normalized or fuzzy matching and writes a supports relation. The epistemic ledger does not separately assess that mapping's warrant. Retain the supported admission and graph-reuse findings, but withhold a complete comparison of imported relationship semantics or their checking. Title similarity is not evidence that the linked content actually supports another assertion.

  rel_type="supports", --- src/cortex/pipeline/importer.py

• The memory coverage table explicitly leaves detailed diagnostic bodies uninspected. SRC-1 src/cortex/transport/mcp/server.py retains an in-process tracemalloc snapshot and compares a later snapshot against it, clearing state on start/stop. This is an additional cross-invocation diagnostic read-back surface, not retained knowledge, ranking adaptation or a provider-learning mechanism. The document/config/history findings remain supported, but the inventory cannot be used as an exhaustive account of every retained diagnostic state or read-back route.

  _tracemalloc_state["prev_snapshot"] = snapshot --- src/cortex/transport/mcp/server.py