Call Out Fake Fixes on the Spot

Verification must run all the way with consistent order and raw logs; fixes must reproduce before and disappear after. Banning masking fixes is a hard rule.

When AI fixes a bug, the cruelest failure isn't not fixing it—it's faking the fix: swallowing exceptions, loosening assertions, skipping test cases, so the test goes "green" while the problem stays. A fake fix is worse than no fix. It fooled the tests, and it fools you.

After reading, you should be able to answer:

• What does a fake fix look like, and how do you spot it?
• Why do "run all verification" and "quote the raw log" stop it?
• Why are "reproduce before, disappear after" a hard rule?

What a fake fix looks like

• Swallow the exception: catch it and swallow—no report, no handling
• Loosen the assertion: weaken the test condition so a case that should fail turns "green"
• Skip the test case: skip the test that won't pass, pretending it's still there
• Change verification, not code: make the check script stop checking instead of actually fixing

Their common trait: the verification turns "green," but the root cause remains.

How to stop it: run all + quote raw

• Run verification all the way, in consistent order: don't allow skipping bits; changed order makes results incomparable
• Quote the raw log: make AI paste the actual error log instead of restating "fixed"
• Strict gates against fake green: on the delivery chain, evidence gates must be strict—tests green, artifacts present, clean diff, checked one by one copilot-cost

Hard rule: reproduce before, disappear after

The measure of a fix is "the root cause changed," not "the test is green":

1. Reproduce before: stabilize and reproduce the bug first, note the symptom
2. Fix the root cause: change what caused it, not mask it
3. Disappear after: regression the reproduce case, confirm gone; then run the full suite to confirm nothing broke

Any fix that can't say "which root cause it fixed" doesn't count.

Make AI follow the rule

• Tell it "a fix must: reproduce → fix root cause → regression," and have it show evidence
• When reviewing, specifically watch for "did it change the test to fake a green"
• Put the fake-fix patterns into your project rules; AI will commit them less google-patterns

Stop fake fixes and verification becomes trustworthy. Next, the launch questions after verification: ship small, pad before you jump.