Governed Evidence. Data eligible to ground a conclusion: carrying provenance, freshness, evidence approval state, citability, and sensitivity labels. Ungoverned data is not evidence.
That eligibility is contextual. Evidence eligibility asks whether this material may ground this conclusion for this actor and use, not merely whether the system has a copy of it.
When ungoverned material is presented as if citation alone made it defensible, the system has laundered evidence instead of grounding a conclusion.
When sources are merely displayed beside an answer instead of constraining the answer, the system has citation theater, not governed evidence.
Governed evidence crosses a secure fetch boundary before it enters a reasoning context: retrieval must preserve source identity, freshness, citation metadata, and the source snapshot instead of treating downloaded content as an automatic admission decision.
When the admitted evidence is used for model output, it should feed retrieval-grounded reasoning: evidence first, reasoning second, preserved derivation after.
The parametric memory boundary protects that order: learned model associations can help compose an answer, but they do not become governed evidence.
Governed evidence also depends on source health: parser success, crawl drift, freshness, and corpus approval decide whether a source remains eligible to ground reasoning after it was first admitted.
Governed evidence also needs a retention policy, because a conclusion cannot be replayed if the evidence that grounded it was expired by a storage default.
When governed evidence supports a conclusion, the system should preserve the derivation, not only the cited sources.