Proof Before Done

Use whenever an agent is about to claim a meaningful task is complete.

Principle

"Implemented" is not proof that it works.

A completion claim requires evidence appropriate to the task.

Evidence ladder

Possible evidence:

1. source exists
2. static validation passes
3. unit test passes
4. integration test passes
5. real API action works
6. persistence verified
7. browser/UI behavior observed
8. failure behavior verified
9. restart/recovery verified
10. real external provider/system tested

Not every task needs all ten.

But a task must use enough levels to support its claim.

Examples

UI feature

Insufficient:

• page file exists

Better:

• page loads
• real API call occurs
• expected state changes
• browser console clean
• mobile layout checked if required

External integration

Insufficient:

• SDK imports
• mock server passes

Required before REAL DONE:

• actual external integration smoke test succeeded

If only mock passes:
mark external integration PARTIAL.

Coding task

Require:

• relevant test
• diff inspection
• acceptance criteria check

Workflow

1. Restate the completion claim.
2. Map each acceptance criterion to evidence.
3. Run/check evidence.
4. Identify unproven claims.
5. Classify:
   • DONE
   • PARTIAL
   • FAILED
6. Save proof artifact paths/log references.

Never

• trust the implementing agent's "done" message alone
• call mock integration real
• hide skipped tests
• convert unavailable test into PASS

Output

Claim:
Acceptance criteria:
Evidence:
Skipped/unavailable verification:
Known limitation:
Final status: DONE / PARTIAL / FAILED