Defend the threshold policy: from calibrated probabilities to a business action.
1. Calibration before thresholds
PR-AUC + Brier score on holdout; reliability curve in the appendix. An uncalibrated ranker cannot set a money threshold — say this out loud first.
2. Cost-matrix threshold
False-positive cost (wasted retention offer) vs false-negative cost (lost customer LTV). Threshold = argmin expected cost, shown as a one-table sweep — not 0.5 by default. WE W3-W4 is the reference framing.
3. SHAP story in 3 slides
Top-5 drivers with direction, one slice where the model is weakest, one action per slice. Calibration -> threshold -> story is the defense spine.