Every parameter has a source.Every gap stays visible.

The horizon/rebalance tournament found the 20D target with a 10D cadence stronger than the original 5D/5D design on the committed comparison.

Prediction horizon and trading cadence were tested separately. Ten trading days retained the slower-moving 20D signal while reducing unnecessary activity.

Broader 20–75 name portfolios reduced volatility and turnover but diluted return faster than they improved risk-adjusted performance.

The locked model card specifies rank 15 to reduce turnover around the Top-10 boundary. A separate causal ablation for the exact buffer is not included in the public release snapshot.

The 60D Ledoit-Wolf maximum-diversification constructor improved the selected risk/return comparison versus inverse-volatility weighting without optimizer fallback.

A 25% signal allocation increased simulated return while leaving overall Sharpe close to the pure risk anchor. Larger blends increased volatility and drawdown.

The 18% cap is part of the locked governance contract. The public artifact does not include a standalone cap ablation, so no stronger causal claim is made.

The 1.25x ceiling is permission, not a target. The selected research candidate never exceeded 1.00x and spent most periods below full exposure.

The reported net results include an average transaction-cost field and turnover-derived costs. Market impact, taxes, borrow constraints, and realized execution remain outside the public contract.

Annual expanding-window fits generate out-of-sample scores for each test year. The repository also includes data-quality and leakage audits; residual data and universe-selection risk remains.

The public model card verifies annual expanding-window out-of-sample evaluation across 2021–2026. It does not publish a separate final untouched live holdout, and results must be read with model-selection bias in mind.