Skip to main content

Artifact promotion and release decisions — assessment

Score 0 absent/incorrect, 1 prompted/incomplete, 2 independently supported with explicit limits, 3 justified fresh transfer. Every category requires at least 2; a high average cannot hide a missing category.

Required evidence​

  • RL1 — Identity: Trace review to running digest and expected provenance identities.
  • RL2 — Negative verification: Demonstrate tampered, unsigned and wrong-identity rejection with a valid positive control using actual verifier tooling.
  • RL3 — Canary: Exercise regression, shared failure and insufficient/missing data under a recorded policy.
  • RL4 — Recovery: Prove software/data compatibility through abort, including a business-effect limitation.
  • RL5 — Delivery measurement: Retain original releasable trunk and five-metric service baseline with definitions, raw observations and bounded improvement claim.
  • RL6 — Handoff: Provide reproducible workflow, evidence retention, scoped identity and reviewable release/rollback record.

Calibration​

A matching digest is necessary for identity but does not establish trust. Zero failures in two requests is insufficient for a broad safety claim. A restored old binary that cannot read new records is not recovered. Preserve inconclusive results and actual verifier/runtime limitations.

Delayed transfer and integrity​

After at least 48 hours, have a reviewer change one assumption or select a previously unseen failure. Predict the outcome, collect fresh evidence and explain discrepancies without the calibration text. Preserve mistakes, repairs, collaboration and generated-assistance disclosures. Record unavailable tools or reviewers as pending evidence, never fabricated completion. Author reference checks are not learner runtime experiments.