Bullion Vault EA: a FundedNext gold robot that enforces the rulebook

The brief

A trader needed an Expert Advisor to run a FundedNext Stellar 2-Step challenge on a $6,000 account, trading gold (XAUUSD) on the H1 timeframe. The requirement was not "make as much as possible". It was "never breach a rule, and reach the target without doing anything the firm forbids".

That inverts normal EA design. Profit becomes the constraint-satisfaction result, not the objective, so the build started from the rulebook and worked backwards.

How each firm rule maps to code

The honest part: no developer can guarantee a challenge pass, and I do not claim to. What this EA does is make sure the robot does not break the rules that fail most accounts.

How the strategy works

XAUUSD, H1, one trade at a time, hard stop-loss and take-profit on every order. No grid and no martingale in the shipped configuration. Both scale losses in exactly the way a drawdown-limited account cannot afford.

Four-factor confluence scoring

Each factor votes -1, 0 or +1 on the last closed bar, and a trade needs the absolute score to reach the minimum (3 of 4 in the shipped preset):

• Trend stack: H1 SMA20 and SMA50 alignment against price.
• Momentum extreme: H1 RSI(14) against configurable thresholds.
• Dual-horizon momentum: a 20-bar and a 5-bar percentage change that must agree, so a fresh impulse cannot be faked by a stale trend.
• Price action: engulfing candles with body-size comparison, pin bars (wick over twice the body, body under 30% of range), plus 20-bar support and resistance proximity confirmed by candle direction.

After scoring, a direction filter and a FIFO anti-hedge check run, blocking a buy while any sell is open, as US brokers and most prop firms require. Signals are evaluated once per H1 bar close, never intrabar, and every decision is logged so the client can read exactly why a bar did or did not trade.

Risk engineering

The guard runs first on every tick, before any trade management or entry logic, as an ordered chain: server-day roll, latched permanent halts, static floor, daily floor, then the profit target with realised-balance verification.

Position sizing computes the real per-lot dollar risk from tick size, tick value and the actual stop distance, subtracts the risk already committed by open orders, then either clamps the requested size or refuses the trade if even the broker minimum would breach the budget. Halt latches, the day-start anchor and the trading-day count persist across terminal restarts through namespaced global variables, and are deliberately disabled inside the Strategy Tester.

A wrong initial-balance setting used to silently mis-scale every limit. Now it raises a loud configuration error, blocks new entries and keeps the loss floors armed.

Verified results

All figures are MetaTrader 4 Strategy Tester runs on a $6,000 initial balance with a deliberately pinned 40-point spread, above typical live gold spreads, so every result already pays worst-case transaction costs. Pinning matters: with "current" spread the tester bakes whatever the market shows at that instant into the data, making runs irreproducible.

The share of winning trades is intentionally low (21% to 33%). The system runs a 1:4 reward-to-risk ratio, so most trades are small losses and the winners carry the result. The 2026 stress window is the weak row, and it stays in the table: a profit factor of 1.08 is close to flat, and it recorded zero guard halts.

A real production incident, found and fixed

After delivery the client reported that on a fresh account the EA had taken zero trades for a week. No errors, no alerts. Reproduced in the tester: the packaged binary had been compiled one build before the change that made guard persistence live-account-only, so a Strategy Tester run had written halt latches that the live chart then read on attach. An instant, permanent, silent halt.

The fix rebuilt the binary with a build assertion that the shipped executable matches source, added a one-click latch reset, and added an observability layer: a support snapshot at attach, an hourly heartbeat, per-bar signal maths and explicit guard lines. The handoff note now tells the client which three log lines to send for a complete remote diagnosis.

Method notes

• Reproducible builds: a generator derives the production and an instrumented build from one source, with every anchored replacement asserting its expected occurrence count.
• A headless tester runner with model, date range, preset, report name and pinned spread as parameters.
• Custom optimisation metrics that rank candidates by challenge safety (worst daily loss, static drawdown, guard halts, days to target) rather than raw profit. A faster 2.5%-risk variant was rejected because one window dipped to $5,420, twenty dollars from a breach.