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

There is a step in the development process for large language model (LLM)-assisted tooling that most teams skip because it's tedious, time-consuming, and doesn't produce results visible to e...