The Paranoid Engine

A few lines into our validation layer's documentation — written to ourselves, long before any visitor would read it — sits a warning: without this module, the Factory becomes a backtest fraud factory. That sentence is the engine's constitution. It assumes every candidate strategy is lying until the evidence says otherwise, and it puts the burden of proof entirely on the strategy.

Most of the industry runs the opposite way: assume the curve is real, ship it, let the customer find out. We wrote the warning down because the easiest person to fool with a beautiful curve is the person who built the machine that drew it.

A strategy that produces too few trades to judge doesn't get "inconclusive." It gets zero — not fifty. If you didn't take the test, you didn't half-pass it. This single rule kills the most seductive kind of backtest: the one with three lucky trades and a vertical equity curve.

The same suspicion applies to records that look too comfortable. The engine writes down how every single trade actually ended — and an equity curve that never once dips, a drawdown of exactly zero, is treated as a bug or a memorized past, not a triumph, and penalized on the spot.

Everywhere else, a flawless record is the sales pitch. Here it's evidence. A strategy that never loses across years of data hasn't discovered a miracle — it has almost certainly memorized the past. So a too-perfect score is capped and flagged, not celebrated; and where a profit factor would read "infinity" because there are no losses to divide by, the card prints a dash. We'd rather show you a dash than a number that flatters.

Inside a single second, a stop-loss and a liquidation can both be true. Somebody has to decide which one counted — and most engines quietly make that call in the strategy's favor. Ours settles it the way the exchange would: the liquidation wins, and it fills at the liquidation price, not a friendlier one. The order of cruelty is fixed, written down, and always against the strategy.

How the Engine Grades covers the rest of the simulation's table manners — the real sequence inside every candle, entries that never peek at the bar they trade, every cost charged. And the Four Gates are the formal exams on top of all of it.

Here's the part no trading site publishes. Twice, the paranoia turned inward — and found something.

Once, we caught one of our own validation gates grading on easier terms than the engine it was policing — a double standard. The discovery was written into the code as a confession, and the gate was put back on the harder standard. Another time, a robustness check was fooled into calling a strategy "stable" while the settings it was shaking steered nothing; the hole was patched, and that confession stayed in the code too.

We keep those notes in writing for a simple reason: a lab that shades the truth in its smallest corner has no standing to ask your trust with the big ones. An honest machine isn't one that never errs. It's one you can watch catching itself.

One line we hold, openly: we don't publish the engine's insides. The breeding recipe, the parameters, the architecture — that's a year of one builder's life, and it stays sealed. What we publish instead is everything the engine refuses to let a strategy get away with — this page — and a record you can check without trusting a word of it: every active strategy is re-graded daily on candles it has never seen, and the result lands on every card you save, whether it flatters the strategy or not. The Hall of Fame shows each seat's reigning champion; when a strategy breaks the promise of its own backtest and is retired, its last grades stay frozen on your card instead of quietly disappearing.

The recipe is sealed. The report card is public.