aeondave/verification-before-completion

Verification Before Completion

Completion is a claim. Claims need fresh evidence.

Boundary: evidence-before-claims governs claim wording and evidence quality in general. This skill is the freshness gate at the completion boundary — did you actually re-run the check after the last change?

Core rule

Do not say work is complete, fixed, passing, clean, validated, or ready unless the verification command or artifact was checked after the last change in the current context.

Gate workflow

1. Name the claim: what exactly is being asserted?
2. Choose proof: command, test, diff, log, replay, artifact, or checklist that would falsify the claim.
3. Run fresh, post-edit: use the full relevant check on the current tree; any prior run before the last edit is stale.
4. Read the output: exit code, failures, warnings, skipped checks, and scope limits. Exit 0 ≠ objective met — inspect what the tool actually produced.
5. Report accurately: claim success only when evidence supports it; otherwise state the actual status and next smallest fix.

Common claim gates

| Claim | Requires | Not enough | |---|---|---| | Tests pass | fresh run after the last code edit, zero relevant failures | earlier green run before further edits | | Build succeeds | build command exit 0 on current tree | lint-only, or stale build artifact | | Bug fixed | original reproducer now fails-then-passes across the fix | code changed near symptom, or “should work” | | Requirements met | checklist mapped to each spec/request item | tests pass without covering the requirement | | Skill valid | quick_validate.py on the changed skill dir | frontmatter looks right by inspection | | Objective met via tool | tool output shows the objective state | tool exit 0 or “no error” | | Delegated / sub-agent work done | controller inspects diff/artifacts and re-runs the check | worker/sub-agent report says done | | Cleanup / remediation complete | post-action target-state re-inspected | remediation command ran without error |

Security and offensive focus

• Exploit/tooling claims need reproduced primitive, target/build context, or marked uncertainty.
• Cleanup/remediation claims need post-action target-state evidence.
• Scanner triage needs replay, source/sink confirmation, or explicit downgrade.
• Report-ready findings need evidence, limits, and smallest next verification step.

Resources

Load on demand:

• references/completion-evidence.md — detailed claim evidence table and red flags.