Verification before completion
Iron law
NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE
Evidence must come from this turn and cover the exact claim.
Gate
Before a success claim, commit, push, or PR:
- Identify: Name the claim and the command or inspection that proves it
- Run: Execute the complete check now
- Read: Inspect full output, exit code, and failure count
- Compare: Confirm the evidence proves the claim
- Report: State the result with its evidence
A skipped step means the gate failed.
Evidence selection
- Use the narrowest check that proves the claim
- Use broader checks for shared behavior, build config, or user workflows
- Bug fix: reproduce the original symptom
- Regression test: verify red, green, red without the fix, then green restored
- Requirements: inspect each requirement; tests alone do not prove coverage
- Delegated task status: require the worker's raw command and result for
coordination
- Designated finalizer: inspect the complete diff and run final verification.
Exactly one agent runs it. The orchestrator relays its exact result and does
not rerun the checks
| Claim |
Required evidence |
| Tests pass |
Test command, exit 0, zero failures |
| Lint/types clean |
Complete command, zero errors |
| Build succeeds |
Build command, exit 0 |
| Bug fixed |
Original reproduction now passes |
| Requirements met |
Requirement-by-requirement inspection |
| Agent completed |
Finalizer diff inspection plus verification |
Include the command, exit code, and relevant result in the final report.
Blocked verification
If the check cannot run:
- State the blocker
- Name what remains unverified
- Do not claim complete, fixed, passing, or ready
Confidence, prior runs, partial checks, and generic agent summaries are not
evidence. A designated finalizer's exact command, exit code, and relevant result
are evidence for a verbatim orchestrator relay; the orchestrator does not rerun
them.
1---2name: verification-before-completion3description: Use when about to claim work is complete, fixed, or passing, before committing, pushing, or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always4---56# Verification before completion78## Iron law910```text11NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE12```1314Evidence must come from this turn and cover the exact claim.1516## Gate1718Before a success claim, commit, push, or PR:19201. **Identify:** Name the claim and the command or inspection that proves it212. **Run:** Execute the complete check now223. **Read:** Inspect full output, exit code, and failure count234. **Compare:** Confirm the evidence proves the claim245. **Report:** State the result with its evidence2526A skipped step means the gate failed.2728## Evidence selection2930- Use the narrowest check that proves the claim31- Use broader checks for shared behavior, build config, or user workflows32- Bug fix: reproduce the original symptom33- Regression test: verify red, green, red without the fix, then green restored34- Requirements: inspect each requirement; tests alone do not prove coverage35- Delegated task status: require the worker's raw command and result for36 coordination37- Designated finalizer: inspect the complete diff and run final verification.38 Exactly one agent runs it. The orchestrator relays its exact result and does39 not rerun the checks4041| Claim | Required evidence |42| ---------------- | ------------------------------------------- |43| Tests pass | Test command, exit 0, zero failures |44| Lint/types clean | Complete command, zero errors |45| Build succeeds | Build command, exit 0 |46| Bug fixed | Original reproduction now passes |47| Requirements met | Requirement-by-requirement inspection |48| Agent completed | Finalizer diff inspection plus verification |4950Include the command, exit code, and relevant result in the final report.5152## Blocked verification5354If the check cannot run:5556- State the blocker57- Name what remains unverified58- Do not claim complete, fixed, passing, or ready5960Confidence, prior runs, partial checks, and generic agent summaries are not61evidence. A designated finalizer's exact command, exit code, and relevant result62are evidence for a verbatim orchestrator relay; the orchestrator does not rerun63them.
carlos-algms