An eval harness found what qualitative review couldn't: AI models are most confident when wrong

That last finding is the one that qualitative review would never have surfaced. The model's expressed confidence didn't ...