Probability calibration (Platt scaling) fits a sigmoid mapping p_cal = sigmoid(a × logit(p_raw) + b) to past (predicted, actual) pairs. The global Calibrator averages this fit across every market condition.

The problem: a model that's well-calibrated in calm markets is often overconfident in high-vol markets. The same raw 70% might really win 70% in chop but only 55% when VIX is 25+. Averaging hides this.

The fix: maintain a separate Platt scaler per regime:

• bull — SPY day chg > +0.5%, VIX < 18
• bear — SPY day chg < -0.5%, VIX < 25
• chop — |SPY| < 0.5%, VIX < 22
• high-vol — VIX ≥ 22 OR |SPY| > 1.5%
• mixed — everything else

At predict time, the Unified Predictor picks the right calibrator based on current SPY/VIX. Falls back to the global Calibrator (or raw) if the current regime has fewer than 30 resolved samples.