See the input → responsibility → output diagram.
What it does for you: It keeps what the local fixture worker observed for each check as a record named by its own contents and pinned to the exact Candidate, Run and Attempt. A record that has been altered on disk is refused when read. Assurance then reads the named records back and gives a Verdict: Verified, Failed or Inconclusive. A check that plainly failed gives Failed. Anything missing, altered or unclear never counts as a pass.

Meet the curator. Show me what happened. Keeps observations attributable and separate from the judgement made from them. Meet the whole Plumb family.
How it links: It takes observations from the Local fixture worker, pinned to Execution Runs and Attempts, and stores those of the Repository collector (accepted, not yet called). It gives Verdicts to Guarded verification / recovery, which re-checks them later against the same records.
Honest limits:
- It trusts the local collector. Anyone who can write the store's folder can add well-formed records, and the collector name is a label, not proof of identity.
- A Verdict covers only the records it was given, on local synthetic fixtures. It is not a release, production or customer claim, and survival after power loss is not tested.
Status: Accepted for trusted local fixtures. Acceptance and evidence.
W30 pre-seal host Evidence candidate
EvidenceStore.putImplementation/getImplementation retain separate sf-evidence/2 request/source/program/Attempt/dispatch observations through the same durable writer. Its checked brand cannot satisfy Candidate Assurance. Existing v1 bytes and brand are unchanged. Host Evidence precedes output interpretation and Candidate sealing; a later witness supplies result linkage. Author compatibility/native checks passed; independent final review remains open. Implementation Workflow.
