The cognitive plane has one currency: the universal Evidence object in tcsf/evidence/models.py. Scanners may produce stdout. Nothing above storage is allowed to reason on it.
What an evidence node is
evidence_type— port state, vuln match,hypothesis_support,experiment_result,llm_extraction, …observations[]— keyed values with extractor + confidencerelationships—supports/contradicts/derived_fromcontent_hash— SHA-256 of type + target + observations + metadata
LLM output is not a finding
LLM_EXTRACTION is a typed evidence class. The model may propose hypotheses; it cannot mint confirmed findings. AntiHallucinationGuard downgrades unsupported claims to hypothesis — it never deletes history.
Customer-facing rule
No confirmed finding without at least one non-LLM evidence node. Integrity is queryable via GET /api/tcsf/investigations/<id>/integrity. Persist path: tcsf_evidence in the TCSF SQLite store.
Why this matters
Autonomous security research only earns trust when the runtime can refuse to overclaim. This note documents a shipped invariant in Nexus / TCSF — the kind of detail practitioners and search engines both need to evaluate the system honestly.
FAQ
Why not reason on scanner stdout?
Layer A invariant: scanners emit normalised findings; engines above storage consume Evidence — typed, hashed, relational — not raw tool dumps.
Is LLM output a confirmed finding?
No. LLM_EXTRACTION is a typed evidence class. AntiHallucinationGuard downgrades unsupported claims to hypothesis.