MODEL OBSERVATORY / CLASSIFICATION DIAGNOSTICS

Evaluation report

600 current records analyzed with explicit calibration and slice checks.

Current performance

Accuracy0.697-0.135 vs reference
Macro F10.688-0.134 vs reference
Log loss0.876+0.269 vs reference
ECE0.169+0.075 vs reference

Distribution shift

Prediction JS divergence0.0007
Label JS divergence0.0001
Confidence PSI0.0412

Worst-first slice analysis

SliceValueCountAccuracyGapECE
regionwest2080.625-0.0720.225
devicemobile3040.671-0.0260.173
devicedesktop2960.723+0.0260.165
regionnorth1890.730+0.0330.163
regionsouth2030.739+0.0420.118

Class-level performance

ClassSupportPrecisionRecallF1
approve2690.7770.7140.744
reject1490.6080.6980.650
review1820.6700.6700.670

Selective prediction

10%0.0% risk
20%14.2% risk
30%23.3% risk
40%32.1% risk
50%36.0% risk
60%35.6% risk
70%34.8% risk
80%34.8% risk
90%33.7% risk
100%30.3% risk