Five simple controls that caught nine AI agent measurement errors in a single day

A developer using an AI agent documented nine separate measurement failures that occurred on a single day, all of which were caught before affecting anyone else. The failures stemmed from flawed search tools, reused sentinel strings, and instruments that counted the same thing in inconsistent ways. Five lightweight controls were applied to every measurement: positive controls, negative controls, cross-method counting, logical bounds checks, and idempotency verification. Each control works by being capable of disagreeing with the primary check, rather than simply confirming it. The author argues that better checks are not the solution — instead, independent controls that can actively contradict a measurement are what prevent silent failures.