Tuning Controller
Backtest a profile's real P&L over historical news + prices, see how it stacks against the live champion, and promote the winner to production — audited and reversible. Every backtest is logged to the leaderboard: the moat as a growing dataset.
No need to know the dials. Say what you want — “react faster and trust CEOs more” — and Claude proposes the exact dial changes, explains why, and backtests them so the suggestion comes with a real return number. It only proposes; you decide.
Sweeps credibility / insider / decay / threshold dials, backtests each on a train slice and a held-out testslice, and ranks on test return so the winner isn't overfit. A candidate is robust only when it earns out-of-sample with a small train→test gap.
Backtests every profile against the live champion and promotes a winner only if it clears the gates: beat the champion by the edge, stay under the drawdown ceiling, and trade enough to be meaningful. Turn on auto-promote and cron it (profiles tournament --auto-promote) for a self-improving loop — every promotion stays audited and reversible.