← API referenceValidation

How The Judge validates a strategy

The Judge is EX7's strategy validator. It takes the strategy you build on the canvas and runs it through a reproducible, statistics-first gauntlet. This page describes that methodology in plain language and is honest about what the result is and is not.

What The Judge actually runs

When you submit a strategy, The Judge compiles it to a backtestable specification (a graph the engine cannot represent is refused with 422 unsupported_strategy rather than faked) and runs a deterministic gauntlet:

Walk-forward windows. Your in-sample fit and your out-of-sample performance are checked across multiple contiguous windows. The result reports how many windows were profitable out of how many were run.
Monte Carlo resampling. The trade sequence is shuffled many times to test how much the result depends on order, surfacing the thinness of a small trade count rather than hiding it behind a single equity curve.
Deflated Sharpe ratio. The observed Sharpe is discounted for the multiple testing that produced the strategy, so a result that survived a sweep is not reported as if it had been tested once.

Per-instrument commissions and two-sided slippage are always deducted, so the numbers are net of realistic transaction costs, not gross best-case figures. Today the Judge scores CME micro futures (MES and MCL, on 15-minute bars); a strategy on stocks, ETFs, crypto or another future can be built and exported, and is refused before a run instead of simulated.

The score

The gauntlet ends in a few things you can read at a glance:

ScoreA number from 0 to 100, straight from the engine. Higher means the idea held up better, net of costs. Still hypothetical, not a guarantee.

Rank"#k of n": where the idea sits among your scored ideas, highest score first. It appears once two of them have a score.

ChecksEach check (walk-forward, Monte Carlo, costs, sample size and more) shows whether it held up, so you can see why the score is what it is.

Not scored yetNot run yet, or too few trades to score the idea. Run the Judge, or collect more data and widen the test before deciding.

What the score is NOT

Every result The Judge reports is hypothetical. It describes how a strategy would have behaved on historical data under a fixed set of assumptions. It is not a prediction of future returns, not an offer to sell or manage a strategy, and not investment advice. Past performance, even rigorously backtested past performance, does not predict future results. Futures trading carries substantial risk of loss.

EX7 does not promise any strategy will hold an edge. The Judge measures how defensible a strategy's backtest is under statistics; it does not certify that the edge will persist in live markets.

Screens that present a backtested or simulated result carry the CFTC Rule 4.41(b) hypothetical-performance disclaimer, which is reproduced in full in the Terms of Service.

Submitting a strategy

Submit a strategy to The Judge from the Workspace, or let your own AI do it over MCP. POST /api/judge/runs queues a run against the real engine (the route never simulates a result). GET /api/judge/health is the engine's health ping, not this POST. A rolling fair-use limit protects the shared engine, including when BYOK is connected. Poll the returned run id for the completed score and checks.