Implemented
PRs #14, #21, and #22 added content-addressed fleet artifacts, a frozen benchmark manifest, exact image identity checks, and paired development evidence.
Remaining work
Seal the complete supervisor-owned evaluator snapshot, including evaluator code, configuration, commands, base commit, and selected output artifacts. Detect drift before selection and resume, fail closed on changed inputs, and report integrity status. The existing hashes do not establish the original cryptographic snapshot and resume criteria. The proof sequence #15–#20 does not by itself close this gap.
Acceptance criteria
- Evaluation snapshots include evaluator code/configuration, commands, base commit, and artifact hashes.
- Resume refuses or clearly flags a changed snapshot.
- Reports make integrity status visible to a human reviewer.
Implemented
PRs #14, #21, and #22 added content-addressed fleet artifacts, a frozen benchmark manifest, exact image identity checks, and paired development evidence.
Remaining work
Seal the complete supervisor-owned evaluator snapshot, including evaluator code, configuration, commands, base commit, and selected output artifacts. Detect drift before selection and resume, fail closed on changed inputs, and report integrity status. The existing hashes do not establish the original cryptographic snapshot and resume criteria. The proof sequence #15–#20 does not by itself close this gap.
Acceptance criteria