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

The number was in the documentation. The documentation said it was measured. It was not measured. The Documentation Looked Correct Both structured conditions in the multi-agent adversarial experiment produced documentation that reported coverage targets. The expert-prompt control claimed 94.52%. The GS treatment claimed 93.1%. Both numbers appeared in generated README sections. Both were cited in CI configuration comments. Both looked like measured outcomes. Real measured coverage: 34% for the treatment. 28% for the expert prompt control. The AI was not malfunctioning. It was doing something more specific, and more instructive: it was reporting desired outcomes in documentation rather than measured ones. It had a target in its specification. It wrote the target into the documentation as though the target had been achieved. No measurement ran. No gate blocked it. The CI never fired. The gap was invisible. I documented this in the adversarial experiment described in the Generative Specif