The model trains a tiny logistic regression on resolved (base_predictions, win) tuples. These are the softmax-normalized weights it has learned. Compare them to the hand-picked starting weights below.

Starting weights (used in cold start): model 35% · ensemble 25% · bootstrap 13% · k-NN 13% · SWA 14%
Raw (logit-space) weights are above; the bias term is shown separately.

The brain has five different ways to predict the same thing:

• Model — logistic regression on 22 features, trained with Adam
• Ensemble — multi-horizon (10min/30min/4h) average
• Bootstrap — K=5 models trained on resampled data, then averaged
• k-NN — non-parametric memory of the 10 most similar past resolutions
• SWA — average of late-training weights, sits in flatter minimum

Previously we blended them with hand-picked weights. The hand-picked weights are guesses. The Meta-Stacker replaces them with weights learned from the data:

1. At predict time, collect the five base predictions as features: x = [p_model, p_ensemble, p_bootstrap, p_knn, p_swa, 1]
2. Train a logistic regression: meta_p = sigmoid(w · x)
3. On resolution, do an online SGD step on the meta-weights with the actual outcome

Cold start: Below 30 resolutions, fall back to the hand-picked weights. After that, the learned blend takes over. The Unified Predictor automatically chooses whichever is available.

Why this is good: if k-NN starts outperforming the parametric model in a specific regime, its weight grows automatically. If SWA's average becomes stale during a regime shift, its weight shrinks. The blend self-tunes.