Architect EA Audit: The Math Behind the 87% Win Rate

A new "AI-validated" forex EA has been making the rounds: the Architect, marketed alongside something called the Vigorous strategy. The pitch is straightforward — the EA generates signals from a rule-based core, then asks an LLM ("AI") whether each entry is a good idea. If the AI says yes, the order fires. The marketed numbers, on a live test account from October 10, 2025 onward:

87% win rate over 398 trades
Profit factor 4.15
Six pairs traded (five majors + gold)
8.89% gain over ~8 weeks, with 2% drawdown

This audit is a technical analysis of the publicly-shared statistics. We have no commercial relationship with the product or its creator. We cannot inspect the source code, so our conclusions are inferences from observed numbers — but as we'll show, several of those numbers are dispositive on their own.

If you only read one section, read the next one.

The USDJPY anomaly

Here's the per-pair breakdown from the public Myfxbook screenshot:

Pair	Side	Trades	Pips	Profit ($)
USDJPY	Longs	20	+73.5	+4.73
USDJPY	Shorts	36	−60.8	+15.03

Look at the shorts row. 36 trades, negative net pips, positive net dollars.

This is not a rounding artifact. It's a structural impossibility under fixed lot sizing. If every trade on a symbol uses the same lot size, then $/pip is constant, and the sign of the net pips must equal the sign of the net dollars. They can't disagree.

The only way to lose pips and gain dollars across a series of trades is variable lot sizing where winners use larger lots than losers. That sizing pattern is the textbook signature of:

Grid / averaging-down recovery (each new entry against the position is smaller, and the eventual recovery winner closes the whole basket at a small profit), or
Martingale-style position scaling (after a loss, the next trade is sized larger to "win it back").

These are not edges. They are risk-redistribution mechanisms — they convert frequent small wins into rare catastrophic losses. They are the standard reason why retail FX bots show smooth equity curves for months and then go vertical in a single afternoon.

You can see a milder version of the same asymmetry in the EURUSD and GBPUSD rows: the $/pip ratio differs noticeably between the long and short sides. Fixed-lot strategies don't do that. Recovery-based strategies do.

The USDJPY shorts row is the strongest single piece of evidence in the public dataset. It alone reframes everything else.

The win rate uplift

The promotional video states that the Vigorous strategy without the AI gate runs at a 65–70% win rate, and with the AI gate is hitting 87%. That's an uplift of 17–22 percentage points.

In principle, a sufficiently selective filter can lift the win rate of an existing signal pool — if you keep only the top quintile of setups, the kept setups will have a higher win rate than the population. So this isn't impossible in isolation.

But we have to ask: what is the gate selecting for? In a system without a recovery mechanism, the gate is selecting trades more likely to hit TP before SL. In a system with a recovery mechanism, the gate has another option — it can simply skip the entries the recovery layer would have struggled to absorb. The win rate then rises mechanically, because the trades that would have been "wins" only after a long grid recovery (or losses, in the worst case) are filtered out before they become problems.

Given the USDJPY anomaly above, the second explanation is the parsimonious one. The AI gate isn't adding a 17–22pp directional edge. It's protecting an underlying recovery mechanism from the trade conditions where it breaks.

This is also consistent with the creator's own AI prompt, which we'll look at in a moment.

The sample size problem

The track record covers approximately 8 weeks (October 10, 2025 → December 3, 2025). 398 trades sounds like a lot, but for a multi-pair recovery-style system, trade count is the wrong unit of measurement. The right unit is uncorrelated stress events — moments where the strategy's risk model is actually tested.

In 8 weeks, an FX system experiences:

A handful of major news events
One, maybe two genuine regime shifts
Almost no tail volatility

Recovery-based EAs typically survive months of normal conditions before encountering the trade sequence that exhausts margin. A 1.5–2 month window is not validation of profit factor 4 — it's validation that the bot hasn't yet hit its kill condition. The relevant statistic isn't trade count, it's number of independent stress regimes survived, and 8 weeks gives you maybe 1–2.

This is also why the drawdown is reported as 2%. A grid hasn't reached its drawdown until it closes — until then, floating losses don't show on a closed-trade equity curve. The Myfxbook drawdown number reflects realized DD only.

"I can't backtest the AI functionality"

A direct quote from the marketing video:

I can't back test using the AI functionality because it's in real time.

This is the part that should make any technical buyer pause. There is no engineering reason this has to be true. An LLM gate can be backtested — you replay historical bars and historical news, you call the LLM (or a cached LLM response) at each decision point, and you log the result. The gate becomes deterministic if you fix the prompt template and pin the model. We've published a working reference implementation of exactly this pattern.

What the statement actually means in practice is: there is no out-of-sample validation, no walk-forward analysis, no scenario testing, and no way for a buyer to verify the gate's behavior on any data the creator hasn't seen. Performance claims are unfalsifiable until live performance refutes them — at which point the buyer's capital has already done the falsification.

When a vendor tells you their core differentiator can't be tested, that is information. Take it seriously.

The tell in the prompt

The promotional video shows a screenshot of the Architect's own AI gate output. One of the rationale strings reads:

No strong grid edge in direction due to floating loss and risk

The word "grid" appears in the gate's reasoning prompt. The creator built an AI that reasons about the strategy's grid state — because the strategy has a grid state. The gate is grid-aware because it's gating a grid.

This is not a smoking gun on its own; "grid" can be used loosely. But combined with the USDJPY anomaly and the asymmetric per-pair $/pip ratios, it's a third independent line of evidence pointing at the same underlying mechanism.

What a legitimate AI gate looks like

To be clear: putting an LLM in front of a base strategy as a veto filter is a real and useful idea. We do it ourselves in our research stack. The mechanics that make it legitimate are:

Veto-only. The gate cannot create signals or modify SL/TP/size. The base strategy must have edge on its own; the gate only reduces trade count.
Deterministic and replayable. Prompt template version-pinned, model-pinned, decisions cached by payload hash. The gate is backtestable.
Fully audited. Every decision logged: input payload, raw model response, decision, confidence, veto reasons, post-trade outcome.
Realistic uplift. A well-designed gate adds roughly +0.3–0.7 to profit factor on top of a base strategy that already has edge. PF moving from 1.4 → 1.8 is a great result. PF claims above 3 on multi-pair retail FX are extraordinary, and require extraordinary evidence.

Profit factor 4.15 over 1,000+ trades on retail FX with fixed sizing is not a level any well-known systematic firm operates at, let alone a retail EA in its first 8 weeks. When the marketed number sits in that range, the question is no longer "is this a good system?" — it's "where is the hidden risk?"

How to vet a bot before you fund it

Five quick checks any retail trader can run on a public track record:

Compare net pips and net dollars per side, per pair. If the signs disagree, lot sizing is variable. Variable lot sizing is the grid signature.
Demand backtest results with the AI/filter layer. "It can't be backtested" is a choice, not a constraint.
Look at the win rate uplift from any single filter. Anything above ~10pp from a regime/news filter alone deserves skepticism.
Check the live track record length against the kill condition. Recovery EAs survive 3–9 months on average before the regime that breaks them. 8 weeks is not validation.
Search the strategy's reasoning text for "grid", "recovery", "DCA", "averaging". Vendors of legitimate filter-based systems don't write prompts that reason about grid state.

Closing

The Architect EA may be doing something interesting with its AI gate — the engineering of the gate layer itself isn't the issue. The issue is the base strategy underneath the gate, what its risk profile actually is, and whether the marketed statistics describe a robust edge or a recovery mechanism running on borrowed time.

The public data, as it stands, is more consistent with the second.

This is technical analysis of publicly available marketing claims, not legal accusations. We have no commercial relationship with the product or its creator. Conclusions are inferences from observed statistics; we cannot verify internal logic without source code access. None of this is investment advice.

TradeAudit performs independent audits of trading strategies and EAs. If you're considering allocating capital to a system you didn't build yourself, get the math reviewed first. Request an audit →