verification-standard

Verification standard

Select the verification policy first: routine work uses builder evidence; high-risk work requires an independent pass. An independent verifier checks the original requirement and actual artifact, not agreement with another agent. Record which property, revision, environment and user flow the evidence covers. A data-scoping audit does not establish authorization or read-only access.

Evidence that can fail

• For a regression, show the relevant test failing on the unfixed behavior and passing on the fix. A focused revert/mutation is useful when a base run cannot isolate it. Record the actual command, exit code, failing assertion and restored pass; prose describing a hypothetical fail is not proof. Keep the experiment proportionate to the change.
• Safety assertions need a negative control: deliberately break the guarded behavior and show the guard detects it. Assert a nonzero denominator; zero leaks from zero exercised requests proves nothing. Await async effects so a swallowed error/pending promise cannot fake green.
• Test observable outcomes, not just a mocked method call or a source substring. Fixtures must match the real schema/API and use independent expected values; deriving expectations from the same broken constant certifies the bug. Include important boundaries and cross-surface consumers; handpicked passing examples need not cover the actual failure range.
• Choose mutations by failure mechanism, including wiring/call sites, not only helper internals. A killed-mutant ratio applies to that chosen set, not all behavior. Investigate surviving mutants before calling them equivalent. Ensure the changed source actually compiled.
• Mutation work needs exclusive ownership, a snapshot/backup and an in-flight marker, followed by exact restoration and a clean intended diff (coordinator-kit:agent-brief-hygiene).
• Separate measured facts from severity/inference; use a baseline for anomaly claims. Numbers need a timestamp and source. Scratch reports, cached inboxes and an agent's memory are not proof of current production state.