What testing an AI supervisor taught us about QuantumGrid Smart.
For years, our job has been narrow and unglamorous: build trading systems, then try to break them before anyone else's money is on the line. Grid recovery, entry filters, position sizing — the boring engineering that either holds up or doesn't.
Earlier this year we asked a bigger question: what if the system could get smarter mid-cycle? Not just follow fixed rules, but read the market's stress level in real time and adjust its own risk-taking — sizing up when it detected a high-probability recovery setup, easing off when conditions looked dangerous. We built a neural-network supervisor to do exactly that, and pointed it at our flagship grid engine.
We ran it the only way that means anything: full historical backtest, real ticks, no shortcuts. Two years of market data, the model already trained on the period, so this was the best-case scenario for the AI — not a stress test yet, just: does the core idea work at all.
It did. For the first stretch, it worked beautifully. The account grew from $10,000 to roughly $17,000. The AI was doing exactly what we designed it to do — sizing up into recovery opportunities, compounding gains faster than the mechanical version underneath it. If we'd stopped the test there, this would be a very different post.
We didn't stop there. The market kept moving, and the AI kept doing what had just worked — sizing up into a strong trend that, this time, never pulled back. Each addition made the position larger right as it needed to be smaller. By the time it was over, the account hadn't just given back the $7,000 gain. It had gone to -$10,088. Profit factor 0.60. The exact mechanism that produced the early win produced the collapse.
We killed that version. Not tuned it, not patched it — took it out of the flow entirely. What we shipped instead was the plain mechanical core underneath it, with the AI supervisor disabled, because across every variant we tested, the rules-based system alone matched or beat every "smarter" version we tried to build on top of it.
That's not a hedge or a disclaimer — it's the actual reason QuantumGrid Smart looks the way it does today. Same dual-confirmation entry philosophy as our QuantumGrid Pro system, extended with volatility-regime detection and adaptive sizing, tested across 14 years of data including a separate out-of-sample slice we never optimized against. The AI architecture is still in the code, disabled and fully documented, waiting for a version that earns its place instead of one that looks good on a curve until it doesn't:
14yr backtest (2012-2026, real ticks): Profit Factor 5.00 | Recovery Factor 7.29 | Sharpe 2.48 | Balance Drawdown 2.84% | Equity Drawdown 15.47%
Out-of-sample slice (2022-2026): Profit Factor 5.07, Balance Drawdown 2.90%
Live, independently tracked since this week — the newest of our three live accounts, precisely because we wanted that out-of-sample answer before shipping it:
myfxbook.com/portfolio/quantum-grid-smart/12106981
Forex Factory: forexfactory.com/sentinellabs#acct.99
You don't have to take any of this on faith — that's the entire point of publishing the live account alongside the backtest. Go check the numbers yourself, ask me about the AI Guardian module, the entry logic, or the out-of-sample test in the comments. I'd rather you find the next weak point than I ship you a curve with no story behind it.
Full docs: sentinellaboratories.com
Disclaimer: backtested and live performance shown are historical and do not guarantee future results. FX trading carries significant risk of loss, including loss of the entire account balance.
