A statistical technique using Platt scaling with a hierarchical prior to adjust model confidence while preventing over-shrinking of extreme predictions.