Developer Finds 9 Bugs in His Own AI Eval Tool — All Made Results Look Better

A developer building a mutation-testing evaluation harness discovered nine separate bugs after completing roughly 40 hours of work. Every single bug skewed results in the same direction, making the tool's performance appear better than it actually was. The author argues this is a structural flaw in self-built evaluations: engineers tend to investigate disappointing results and fix those bugs, while positive results go unscrutinised. This selective debugging means flattering errors survive to publication not through dishonesty, but simply because they never trigger the instinct to investigate. The author warns this asymmetric auditing process is likely a widespread problem, not an individual failing.