The model uses 22 features normalized to [0, 1] or [-1, 1] for stable gradient descent. Every brain finding gets a feature snapshot at emit-time, stored alongside the finding. When the outcome rates 30 min later, the snapshot + label become a training row.