Using semantic entropy - the disagreement across repeated rollouts - as a training signal for LLM confidence. On its own it fails to converge, but as a regularizer alongside Brier score supervision it prevents calibration collapse in the data-scarce regime, and the model learns to say 'I don't know' without ever being rewarded for it.