Verification Before Completion

Evidence before claims. Always.

Core principle: Claiming work is complete without verification is dishonesty, not efficiency.

The Iron Law

code

NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE

If you haven't run the verification command in THIS session, you cannot claim it passes.

The Gate Function

Before ANY status claim, follow this sequence:

code

1. IDENTIFY → What command proves your claim?
2. RUN      → Execute it (fresh, in full)
3. READ     → Examine full output, check exit codes
4. VERIFY   → Does output confirm your claim?
              NO  → State actual status with evidence
              YES → State claim WITH evidence
5. CLAIM    → Only now make the claim

Skipping steps = dishonesty.

Common Claims and Requirements

Claim	Requires	NOT Sufficient
"Tests pass"	`pytest` output: 0 failures	Previous run, "should pass"
"Function works"	Test output + sample run	"Code looks correct"
"Task complete"	Verification command succeeds	"Started the process"
"Bug fixed"	Failing test now passes	"Changed the code"
"Performance improved"	Before/after benchmarks	"Removed the loop"
"Build succeeds"	Build command exit 0	Linter passing

Verification Commands

Tests

bash

# Single module
pytest tests/test_<module>.py -v

# Full suite
pytest tests/ -v

# Specific test
pytest tests/test_module.py::test_specific -v

Evidence format:

code

Tests: 47 passed, 0 failed

Build/Compile

bash

# Check build succeeds
npm run build  # or make, cargo build, etc.
echo "Exit code: $?"

Evidence format:

code

Build complete: exit code 0

Correctness

python

# Compare against reference
assert result == expected
print("Verification: MATCH")

Red Flags - STOP Immediately

Watch for these danger signals:

•Using "should", "probably", "seems to"
•Feeling satisfied before verification ("Done!", "Fixed!")
•About to commit without running tests
•Trusting previous test run from earlier in session
•Thinking "just this once"
•Fatigue making you want work to be finished
•ANY wording implying success without running verification

Forbidden Phrases

code

❌ "Tests should pass now"
❌ "That should fix it"
❌ "The build will succeed"
❌ "I've implemented the feature"  (without test evidence)
❌ "Bug is fixed"  (without test evidence)
❌ "Ready for review"  (without full test run)

Required Phrases

code

✅ "Tests pass (47/47, 0 failures)"
✅ "Task complete: verification output shows success"
✅ "Bug fixed - test_handles_empty now passes"
✅ "Function matches reference (verified with test data)"

Verification Patterns

For Bug Fixes (Red-Green Cycle)

code

1. Write failing test     → Run → See FAIL
2. Implement fix          → Run → See PASS
3. (Optional) Revert fix  → Run → See FAIL again
4. Restore fix            → Run → See PASS
5. Run full suite         → All green

The revert step proves the test actually tests the fix.

For New Features

code

1. Write tests first      → Run → See FAIL (feature missing)
2. Implement feature      → Run → See PASS
3. Run full suite         → No regressions

For Refactoring

code

1. Run tests before       → Record: X passed
2. Refactor code          → (no test changes)
3. Run tests after        → Same X passed
4. Verify behavior identical

Performance Claims

Before claiming "X is faster":

python

import timeit

# Before
before = timeit.timeit(old_function, number=100)

# After
after = timeit.timeit(new_function, number=100)

print(f"Before: {before:.3f}s")
print(f"After: {after:.3f}s")
print(f"Speedup: {before/after:.1f}x")

Rationalization Prevention

Excuse	Reality
"Should work now"	RUN the verification
"I'm confident"	Confidence ≠ evidence
"Just this once"	No exceptions
"Tests passed earlier"	Run them again
"Code looks correct"	Tests prove correctness
"I'm tired"	Exhaustion ≠ excuse
"Takes too long"	Run shorter verification to confirm

Example Session

Bad:

code

Assistant: I've fixed the validation issue. The function should now
handle edge cases correctly.

Good:

code

Assistant: Validation edge case fix implemented. Verification:

$ pytest tests/test_validator.py -v
tests/test_validator.py::test_handles_empty PASSED
tests/test_validator.py::test_handles_single_value PASSED
tests/test_validator.py::test_handles_null PASSED
tests/test_validator.py::test_normal_input PASSED
tests/test_validator.py::test_boundary_values PASSED

5 passed, 0 failed

All validator tests pass. Edge cases now handled correctly.

The Bottom Line

code

No shortcuts for verification.
Run the command.
Read the output.
THEN claim the result.

This is non-negotiable.