AI research assistant

The model proposes strategy hypotheses and explains results. It never runs a backtest, scores a strategy or changes a label: code validates its proposals, the engine tests them, and the same validation gate judges them. It sees train-window results only when proposing, and never sees the test window.