I built an answer key for eval suites: six models broken on purpose, exactly.

This week I launched evalmut — mutation testing for eval suites. Its weakness, stated in the paper: I...