← ClaudeAtlas

verification-evidencelisted

Use before claiming any work state – done, fixed, passing, green, complete, mergeable – and whenever a required or cited check times out, is unexpectedly expensive, or needs a retry. Defines the identify-run-read-verify gate, timeout evidence and retry records, and the baseline/lane-final/integrated/post-merge verification topology with evidence-identity and reuse rules.
bartoszarendt/agenticloop · ★ 2 · AI & Automation · score 69
Install: claude install-skill bartoszarendt/agenticloop
# Verification evidence Evidence before claims, always. A success claim without a fresh run behind it is misreporting, not optimism. ## Proof Kinds And Sources For a typed required check, retain its proof kind independently from its satisfaction source. Kinds are `command`, `manual`, and `contract_proof`. Sources are `pr_body`, `status_check`, `manual_observation`, and `automated_observation`. A status check is an allowed substitute only when the task explicitly permits that exact source for a command check; it never satisfies manual or contract-proof work merely because an RC ID matches. A check that declares observations always requires its structured observation records; even an exact successful status check cannot substitute them. The task contract is validated before either PR-body evidence or a status check can satisfy it. A command proof must declare its exact command in a backtick code span. Invalid kinds or sources, status-check satisfaction on a non-command proof, and observations with no structured PR-body observation source are Maintainer-owned task-policy failures. Manual and contract-proof evidence records each declared observation with the canonical structured shape: one `Observation: <id>` per declared observation, with its `Level:` (`running_path` unless the task pins `name@level`), a concrete `Result:` or bounded excerpt, the current `Artifact:`, and the entry's `Source:`. Copied observation names, generic pass counts, helper-only state, parser or unit-o