Transparency
The 99%: measured, not claimed
Exactly how we measure the AI success rate shown on the homepage: a fixed, reproducible benchmark and the real raw results.
Last updated:
The metric
What success means
We send a fixed set of real strategy prompts through the same AI pipeline the product uses, and score every generation against objective, automatically checkable criteria. A generation counts as a success only if it
- returns a valid strategy graph (0 structural errors), and
- compiles to runnable Python code.
That is a hard, verifiable bar, not a flattering self-rating.
The prompts
What we test
These fixed prompts are used for every measurement, so you see exactly what the AI is measured on (no cherry-picking):
| Prompt ID | Tier | Text |
|---|---|---|
s1-ema-cross | Simple | Go long when the 50 EMA crosses above the 200 EMA, and exit when the 50 EMA crosses back below the 200 EMA. H1 timeframe. |
s2-rsi-reversion | Simple | Enter long when RSI drops below 30, exit when RSI crosses back above 70, with a 2% stop loss. M15 timeframe. |
s3-macd-zero | Simple | Long entry when the MACD histogram crosses above zero. Take profit at twice the risk and place the stop at the most recent swing low. H1. |
m1-rsi-div-bb | Multi-filter | Long when RSI shows a bullish divergence and the MACD histogram is above zero, with a Bollinger Band squeeze as a volatility filter. Dynamic stop at the recent swing low, exit when RSI crosses above 70. Both long and short. H4. |
m2-stoch-trend | Multi-filter | Stochastic crossing up out of oversold is the trigger, but only when price is above the 200 EMA as a trend filter. ATR-based stop loss and a 1.5 risk-reward take profit. M15. |
m3-adx-cci | Multi-filter | Use ADX above 25 as a trend-strength filter, enter long when CCI crosses above -100, with a fixed 1.5% stop and a 3% take profit. H1. |
i1-liquidity-grab | ICT / structure | ICT liquidity grab: price sweeps below equal lows, then a market structure shift to the upside triggers a long. Stop below the grab low, take profit at 3R. M15. |
i2-htf-ob-fvg | ICT / structure | Long on a bullish fair value gap retest that forms after a higher-timeframe (H4) bullish order block, with entry on M15. Stop below the fair value gap, 2.5 risk-reward. |
i3-ob-mss-fvg-short | ICT / structure | Short setup combining a bearish order block, a market structure shift to the downside, and a fair value gap for confluence on H1, with the stop above the order block. |
o1-vp-poc-delta | Orderflow | Volume profile point of control acting as support: go long on a tag of the POC confirmed by a bullish delta divergence. ATR stop and a 2R take profit. M5. |
o2-vwap-reclaim | Orderflow | VWAP reclaim long: price reclaims the session VWAP on above-average volume; exit at plus 1.5R or when price loses the VWAP again. M15. |
c1-mean-reversion-5 | Complex | A five-condition mean-reversion: a touch of the lower Bollinger Band, RSI under 35, a bullish engulfing candle, price above the 200 EMA, and rising volume — go long; mirror the conditions for shorts. Move the stop to breakeven after 1R and trail the remainder. H1. |
Latest result
Measured: 60/60 = 100%
12 prompts × 5 runs = 60 generations · 2026-06-30. Average confidence 0.865.
| Gate | Pass rate |
|---|---|
| Returns a strategy graph | 100% |
| Valid (0 structural errors) | 100% |
| Python code generated | 100% |
By tier:
| Tier | Success | Rate |
|---|---|---|
| Simple | 15/15 | 100% |
| Multi-filter | 15/15 | 100% |
| ICT / structure | 15/15 | 100% |
| Orderflow | 10/10 | 100% |
| Complex | 5/5 | 100% |
Per prompt
Every prompt, every run
success invalid error/timeout
| Prompt | Tier | Runs | Rate |
|---|---|---|---|
s1-ema-cross | Simple | 100% | |
s2-rsi-reversion | Simple | 100% | |
s3-macd-zero | Simple | 100% | |
m1-rsi-div-bb | Multi-filter | 100% | |
m2-stoch-trend | Multi-filter | 100% | |
m3-adx-cci | Multi-filter | 100% | |
i1-liquidity-grab | ICT / structure | 100% | |
i2-htf-ob-fvg | ICT / structure | 100% | |
i3-ob-mss-fvg-short | ICT / structure | 100% | |
o1-vp-poc-delta | Orderflow | 100% | |
o2-vwap-reclaim | Orderflow | 100% | |
c1-mean-reversion-5 | Complex | 100% |
Honesty
Why 99% and not 100%
We measured 60/60. We deliberately publish a conservative 99%: a 60-sample measurement of a non-deterministic model should not claim literal perfection. Statistically, with this result the true rate is at least ~95% at 95% confidence, and the strong measurement puts it at ~99 to 100%.
The figure is about "valid, compilable strategy builds", not about a strategy's profitability. The model is non-deterministic; we re-measure on every model change.
Openly verifiable
The raw data, public
This page is the methodology. The exact prompts are listed above; the complete raw results (every prompt, every single run, with pass/fail) are machine-readable and available without signing in:
→ /api/ai-benchmark (JSON)
The figure is computed solely from the objective criteria defined above (valid graph + compilable code), against the exact same AI pipeline the product uses.