Risk Gate

A language model should never be the thing that moves money. Here the strategy is plain, testable code running on closed 15-minute candles, and the same code drives both the live scanner and the backtester. When it fires, a fast AI model reviews the setup and can veto it for about a third of a cent. Only then does the request reach the gate.

Every Sunday, a larger AI model writes a review of the week's trades with lessons and suggestions for a human to read. It changes nothing on its own.

The gate holds the IG credentials, fetches every account and market value itself, and takes nothing from the requester except the request. Checks run in order and the first failure rejects; any broker error rejects too. Stops can only be tightened, never widened, and closing a position is always allowed.

if (!c.record("kill_switch", !haltActive(), haltActive() ? "HALT_TRADING exists" : undefined)) return c.result();

const lossLimit = -Math.min(LIMITS.dailyLossFraction * acct.equity, caps?.maxDailyLossGBP ?? Infinity);
if (!c.record("daily_loss_cap", todayPnl >= lossLimit, `today ${todayPnl} vs limit ${lossLimit}`))
  return c.result();

const spreadOk = spread > 0 && spread <= m.maxSpread
  && req.stop_distance >= LIMITS.minStopToSpread * spread;
if (!c.record("spread", spreadOk, `spread ${spread} (max ${m.maxSpread})`)) return c.result();

Cron runs the scan every 15 minutes and a position check every minute: flatten at session end, move stops to breakeven, record trades and send Telegram alerts. A kill switch blocks new entries from the strategy and the model, while manual closes and the end-of-day flatten still work, so a halt never traps a position overnight. Real-money trading refuses every entry until the account is explicitly armed on the server with pound caps.

The test suite covers every gate rejection, the kill switch, stop tightening, sizing, backtest costs and the approve, veto and budget paths, against a mock broker and a mock AI reviewer.