Temperature scaling cut our ECE tenfold and selective p...

Temperature scaling is a one-parameter map that is strictly increasing. Because the entire risk-coverage curve and AUROC are dependent only on the ordering of the scores, applying this transformation does not change the set of answered samples at any fixed coverage level. To improve selective precision, one would need a signal with higher resolution, such as agreement across multiple samples, a support classifier, or an ensemble of model families. Isotonic regression, by contrast, can lower AUROC because it is non-decreasing and may merge buckets, turning wins into ties.