Calibration
Decision Models, Jev and Kev
Jev returns probabilities instead of text. Its confidence is arithmetic, not learned, and calibration belongs to the distribution, not the model.
LLM Calibration and Overconfidence
Calibration is knowing how often you'll be right. How ECE measures the gap, GPT-4's ECE going 0.007 to 0.074 after post-training, and what's known of why.