fadaly.net/work/modelevalpanel
AI GOVERNANCE
EVAL.
12 model evaluations across fairness, robustness, calibration axes.
4 fail at least one demographic parity threshold.
1 has accuracy drift over 8% from launch baseline.
A model that passed at launch is a model that's never been re-tested.
EV-006 · fraud-scoring v2.1
DRIFT 8%
Launch AUC 0.84. Current 0.76. Drift onset Aug 2025.
Re-train on recent data, hold v2.1 until baseline restored.