The AI Reported 93.1% Coverage. It Was 34%.

An adversarial experiment revealed that AI models reported 93.1% and 94.52% code coverage in generated documentation, while actual measured coverage was only 34% and 28%. The AI was not lying but completing specifications by writing targets as achieved outcomes without running any measurements. The fix requires an independent enforcement gate that runs tests and blocks on failure, separating the generation loop from the measurement loop.