The Engine, Torn Down

Every report carries four validation badges — Walk-Forward, Out-of-Sample, Robustness, Monte-Carlo — and this site has always told you that clearing all four is the price of becoming a card. For part of the rebuild that sentence ran ahead of the truth: the four tests still ran, but they had been demoted to a signal feeding the grade, and a candidate that failed one could still reach the shelf. It has been put back the strict way — and we'd rather say plainly that it was ever otherwise than let you find out from a footnote. Fail any single gate and the strategy isn't graded lower; it's discarded.

The bill arrives as scarcity — the shelf is thinner than it was, and it should be. A thinner shelf that keeps its promise beats a full one that doesn't. The Four Gates covers what each test actually asks.

The larger change is one you can't see on a card, because it happens before the card exists. The final four months of every two-year window are now sealed away before breeding begins: the evolution that tunes a strategy runs only on the earlier stretch and is physically unable to read the sealed part, with a quarantine gap between the two so a trade straddling the boundary can't leak a hint across it. The vault opens only at graduation, once the strategy is frozen.

Every card on the shelf has therefore already met a stretch of market it was never tuned on — an exam printed on questions it had never seen. And because that result is recorded card by card, we can check the grading scale against reality instead of taste: held against thousands of sealed exams, the higher the grade, the more often the strategy stayed profitable on the unseen stretch. The sealed final months tells the whole story.

Twice before we've published the story of this lab catching itself. Here's the third. An audit of the engine against its own rules caught it letting a trade use information from a moment it hadn't reached yet — undramatic on its own, and flattering in the same direction every single time, which is exactly what makes that class of error dangerous. It was removed, and every card in the warehouse was re-run without it.

Some numbers came down. That is the correct direction for a fix like this: when a simulation stops seeing the future, the honest result is the smaller one. We'd rather ship the smaller number before launch than defend the bigger one afterwards.

Rebuilding an engine mid-flight leaves a warehouse full of cards graded by machines that no longer exist. Rather than sell a museum of mismatched exams, the shelf was cleared and rebred from scratch — twice — so every card you can draw today was born under the same engine, on a window of the same shape.

One quiet consequence, and a good one: a card's birth is now the day its sealed exam ends, so its Since Launch record begins on the very next candle — backtest and live record meeting with no gap between them. If you had saved a card from an earlier generation, it stays on your shelf with its record frozen where it stopped; nothing is deleted.

There's a new engraving at the bottom-right of every report, in the corner where a trading card keeps its set number: Engine V1.0. It names the generation of the engine that bred that card — the first one we're willing to launch behind.

It's also a promise about V1.1, and about V2. The engine will keep being patched; the warehouse will not be thrown away when it is. Every card carries its own strategy, so when the ruler changes the cards migrate onto the new one and are re-scored — in both directions, exactly as they just were. What a version stamp will never mean here is that your card quietly stopped counting.