Real machine learning. Logistic regression with binary cross-entropy loss, trained via stochastic gradient descent. Each finding the brain emits captures a 22-feature snapshot of market state. 30 minutes later the brain rates the outcome as hit / miss / flat. Hits and misses (not flats) become labeled training data — the model trains one sample at a time, updating weights to reduce prediction error.

• Feature vector (22 dims): RSI, ATR%, RVOL, MA distances, IV pct, sector strength, brain weight, regime score, VIX, time-of-day, setup type (one-hot), severity, coincident-finding count
• Loss: binary cross-entropy. Each batch shrinks gradient ~0.05 × prediction error × feature value
• Output: P(win) for any new setup — visible on Model Confidence + Conviction Stack
• Versioning: snapshot any time, A/B compare on Model Versions