P. D. Proskura

Standard KL-divergence metrics for open-set uncertainty frequently correlate inversely with true decision error, but replacing them with a four-parameter direct decision risk formula outperforms complex nonlinear calibrators across vision, audio, and text benchmarks.