What's actually inside
Eleven months from first prototype to the desk you can download today — the final three and a half of them inside this codebase, 187 commits deep. This page is the engineering, laid open. No performance numbers here (those live on the research pages, with their disclosures); just what the machine is made of.
The eyes — a mitigation-aware market-structure engine
Before the AI sees anything, ~7,800 lines of engine turn raw candles into an institutional map, every 15 minutes, across M15 / H1 / H4:
· Fair value gaps with fill-tracking; liquidity sweeps confirmed against real pivots, not noise.
· Structure the way a desk reads it: break-of-structure vs change-of-character, premium/discount location, the optimal-trade-entry pocket, displacement measured against ATR.
· Plus the classical layer — RSI, MACD, EMAs, ADX regime detection, candlestick patterns, support/resistance.
The mind — Claude, consulted like an analyst
That map goes to Claude — Anthropic's frontier AI — with a calibration brief built from measured results: which setups actually paid on these exact pairs, which regimes punish mid-confidence entries, when near-certainty is the only acceptable bar. Deterministic settings, a hard timeout, and one iron rule: any failure, anywhere, becomes a HOLD. The AI decides when to trade. It is never allowed to decide how much.
The spine — a risk engine outside the AI's reach
Nine independent, hard-coded limits stand between every decision and your account — welded in code the AI cannot see, touch, or argue with:
· A daily-loss circuit breaker that counts floating losses, not just booked ones.
· One loss on a pair = that pair is done for the day (the anti-revenge protocol).
· A losing-streak throttle, hard position caps, a risk-budget admission gate, weekend flatten, an account-identity guard, and a total-drawdown backstop you control from Settings.
The conscience — how anything earns its way in
Nothing ships on a good idea. A change to trading behaviour passes 330 automated tests, then a backtest regression across four years of data, then a live demo soak — in that order, no exceptions, including our own ideas we were excited about. The research program has measured over a hundred strategy configurations and rejected most of them in public detail; the handful that survived every market regime are what's inside. When we can't prove something, we say so.
Built by a two-person team and Claude as the engineering department — the same AI family that trades the charts wrote, tested and documented most of these 30,000 lines. Trading involves substantial risk; nothing on this page is a promise of returns.