Trading overview.

A single preregistered logistic trade filter is trained on earlier closed baseline trades after fees and funding. Each test follows a 60-day training window and a 24-hour outcome embargo. No parameter search is performed. Historical markets are BTC, ETH and SOL; this is not validation of the expanded universe. Hourly candles require approximate intrabar stops and synthetic spreads. Forward accounts instead replay recorded observed quotes through the production execution engine.

Frozen candidates run alongside a frozen breakout benchmark, cash, an increased-cost arm, and a global-risk-filter ablation. Features, weights, data and source code are fingerprinted. At least 30 forward days, sufficient closed trades, fresh uninterrupted observations, positive stressed returns, baseline improvement, a positive paired block-bootstrap lower bound, and all three account targets are required for review. No automatic live promotion is enabled. Research drawdown halts remain latched.

This model learns a trade rejection rule. It does not invent strategies, prove causal explanations for losses, or perform autonomous global research. Global feeds currently supply a separately measured risk-filter experiment, not learned global predictions. Probability scores are uncalibrated model outputs. Failed experiments remain visible.

This hypothesis replaces binary win prediction with regularized regression of net realized payoff per unit of initial risk. It learns from separate breakout, pullback and mean-reversion accounts, including their actual exits, fees and funding. Twenty-four five-day block-resampled fits estimate prediction sensitivity; this is not a calibrated confidence interval. An entry needs a positive lower payoff estimate. Portfolio risk limits remain unchanged.

These six test windows reuse the first experiment's dataset. Results are exploratory and cannot establish an independent discovery. No result replaces the first report. Model parameters are fixed before the run; open trades and outcomes crossing the training cutoff are excluded.

The three entry-policy accounts continue collecting executed outcomes even when the learned account rejects an entry. Each closed outcome retains its original global-data context and is withheld from the learning export for 24 hours. This prepares later evidence-based experiments; the current model remains frozen during its 30-day trial. The global-filter comparison tests existing risk throttles, not a trained global forecasting model.

All experiment accounts are separate from the original $100 account. The worker cannot write production balances or promote itself. Positive retrospective results alone never authorize a switch. Full forward observations, adequate trading evidence, baseline and stressed-cost comparisons, fixed drawdown controls, and all three targets remain required.