Prediction calibration graph should not be folded (10% ≠ 90%)

[matthewfallshaw]

Others (Mataculus) have found that low confidence predictions can be better calibrated than high confidence predictions. Pbook should not fold prediction graphs at 50%.

[gwern]

This would also be beneficial from the UX perspective: I've noted a fair number of people are confused by the plot 50-100%, since it doesn't occur to them that PB might be negating <50% predictions and calibrating that way. A straightforward 0-100% graph will be noisier (since it halves the sample size in each bin) but much more easily interpreted.