Read news on eval harness with our app.
Read more in the app
An eval harness found what qualitative review couldn't: AI models are most confident when wrong