Validation
Per-experiment leakage checks, purged time splits, walk-forward folds, skill gate, baseline comparison and the single-use holdout.
Models have no execution authority
Research metadata only. Models cannot submit orders, set position size, approve trades, override risk rejections, change policy or promote themselves. At most, a model in the paper stage may rank or filter deterministic candidates. Training data and results are historical (or explicitly labelled synthetic) and make no profitability claim.
Loading experiments…