verify-before-done

Verify Before Done

Iron rule

검증 증거 없이 완료 주장 절대 금지

Gate function

Before any completion/success claim:

1. Identify: What command proves this claim?
2. Run: Run that command now (previous results are not allowed)
3. Read: Check the full output, exit code, and failure count
4. Verify: Does the output confirm the claim?
   • No → report the actual state with evidence
   • Yes → claim completion with evidence
5. Only after that can you claim

Required evidence per claim

Claim	Required evidence	Insufficient evidence
"Tests pass"	Test command output: 0 failures	Previous run, "it will pass"
"Linter clean"	Linter output: 0 errors	Partial check, assumption
"Build succeeds"	Build command: exit 0	Linter passed, logs look fine
"Bug fixed"	Test of original symptom: passes	Changed code, assume it is fixed
"Requirements met"	Item-by-item checklist	Tests pass

🚨 Danger words (stop immediately when used)

If these words are about to come out of your mouth, STOP:

• "It will work" / "It seems like it will"
• "Probably" / "Mostly"
• "Good!" / "Perfect!" / "Done!" (before verification)
• "It seems like it did"
• "I checked it manually"
• "The agent said it succeeded"

Delegated work is no exception

Sub-agent reports "success" → Independent verification is required

• Check VCS diff → are the changes actually there?
• Run the tests yourself → do they really pass?
• Do not blindly trust agent reports

Preventing rationalization

Excuse	Reality
"It will work fine now"	Run the verification
"I'm sure"	Confidence ≠ evidence
"Just this once"	No exceptions
"The linter passed"	Linter ≠ compiler ≠ runtime
"I'm tired"	Being tired ≠ exemption
"Partial confirmation is enough"	Partial confirmation proves nothing

Why it matters

Completion claims without verification:

• Break trust
• Undefined functions get deployed
• Missing requirements go undiscovered
• False completion → rework → wasted time

Verification is not optional, it is mandatory.