The problem: the model was trained on features collected over the past weeks. If today's features look qualitatively different (different volatility, different sector flows, different time-of-day patterns), the model's predictions are unreliable — it's extrapolating into territory it doesn't understand.

The technique: a classic Kaggle trick called adversarial validation. We collect two pools of feature vectors:

• Old pool — features captured >24h ago
• Recent pool — features captured in the last 2h

Then we label them (old=0, recent=1) and train a logistic regression to distinguish them. Test set AUC tells us how separable they are:

• AUC ≈ 0.5 — features are indistinguishable; no shift
• AUC ≈ 0.6–0.7 — slight shift; still safe
• AUC > 0.70 — clear shift; model is extrapolating

When shift is detected, the Unified Predictor multiplies position size by 0.60 (more conservative under regime change) and fires a bpleone:covariate-shift window event for any listeners.

Complements DriftPSI: DriftPSI tracks output drift (are the brain's predictions changing); this tracks input drift (is the world changing). Together they cover both sides.