According to @_avichawla, equal accuracy hides risk: pick the better-calibrated model and fix overconfidence with ECE and scaling methods.