An Evidence row is a specific edge
Evidence connects one Requirement to one Document fragment. It stores the locator, verbatim snapshot, content hash, factual source claim, status, confidence, and verification timestamps.
satisfied, concern, and insufficient describe what
the source currently demonstrates. Confidence describes how direct that demonstration is.
A source version changes
A user may upload a replacement, or Elevator may refresh a linked canonical source from Drive, Dropbox, Notion, or a URL. The new content is staged as a pending upload through the same intake contract as a manual upload.
Change detection compares the document version and the evidence pin. It does not assume that a familiar filename means the old claim still holds.
The dependent edge is invalidated
If the pinned fragment changed, disappeared, or belonged to an archived Document, the Evidence
becomes stale with a concrete reason.
Stale is not a final assessment. It means “this edge can no longer be trusted until recomputed.” Unresolved stale rows are deterministically returned to the evidence evaluator.
Recompute: re-pin, re-point, or kill
The evaluator re-reads active source material. It may re-pin the claim in the same Document, re-point the edge to a replacement Document, or kill the edge when the claim is gone.
Every path records stale_resolved_at and a resolution. Re-pointing is guarded:
a vaguely related document is not enough; the replacement must preserve the exact claim.
The consequence propagates to readiness
Here the replacement accounts still mention the capital account, but the balance has fallen.
The refreshed edge becomes a concern, with new rationale and a user-facing explanation.
Because the evidence determination changed, the linked Requirement receives a fresh readiness judgement in the same round. Deliverable status and draftability can change downstream.
Confidence guides scrutiny; events drive work
High confidence means a direct quotable passage. Medium means clear contextual implication. Low means inference across sources and cannot be the sole basis for coverage.
Confidence informs scrutiny, but the production work queue is also driven by concrete events: changed sources, archived documents, unresolved stale edges, user disputes, and downstream readiness timestamps. Drift is detected by dependency, not folklore.