A strategy that passed four checks and turned out to be the market

*For a reader with no code. Every number is a record in register/programme-2.jsonl, cited by its number; the same story is told with the records beside it on docs/programme.html and in docs/PROGRAMME-2-CONCLUSION.md.*

What was asked

In September 2026 the lab screened 38,976 ways of entering and exiting a trade — every mechanism its closed list could express, on four groups of US stocks and funds, on the definition partition only, the oldest 30 % of the history, at no cost to its error budget (record #9). From that screen the author chose one cell and registered it as a question (#13): after a four-day run of lower closes on a large US stock, while the market's own trend classifier says "up", buy at the next open and hold for a while. The floor was declared before anything ran: at least 0.15 units of risk per trade after costs, at least fifty trades a year. So was the price: 0.15 of the programme's error budget, charged before a bar of the test half was read (#12).

What the machine found

On the measurement partition — the next 50 % of the history, which no one had looked at — the strategy made +0.213 units of risk per trade after costs, over 6,946 trades, about 413 a year (#19). It passed every check the lab then had:

• it beat random entry with the same holding rule, at the corrected significance;
• its neighbouring settings did about as well, so it was not one lucky parameter;
• leaving any one stock out did not break it;
• it cleared the floor.

Verdict: supported. The strategy moved to the state the lab calls FORWARD (#22). By every rule in force that day, this was a finding.

What the record beside it says

The lab writes one more record after every verdict (#24). It runs the simplest possible alternative — just be long, with the same holding period, in the same "up" regime, on the same stocks, on the same test half — and puts the two side by side.

The entry signal — the four-day run of lower closes — added nothing. What the strategy captured was the market's own return during the years its classifier called "up", 2003 to 2019, at the twenty-day holding period its sweep had chosen because that was where the raw return was largest. On the definition partition, the same comparison had shown a comfortable margin of +0.239; on the measurement partition it vanished (#24).

Why the checks did not see it

Each check asked a real question and got a true answer. Better than chance? Yes — random entry with a coin-flip direction does worse than being long in a rising market. Robust across settings and stocks? Yes — so is being long. Big enough? Yes. None of them asked better than the obvious thing you could have done instead? That is a different question, and on the day the record showed it, the lab made it the fifth check (ADR-0043) and began ranking every sweep by the margin over being long rather than by raw return (ADR-0045) — for every question after this one. This one was not re-judged; the verdict stands as reached, with the record beside it.

What happened next

The same mechanism without the regime gate was asked under the new rules (#26–#33). Ranked by margin, its winner was the short holding period, where being long earns little; it beat being long by +0.075 per trade — the entry did earn something there — and missed the floor by 0.041. Verdict: null. The author stopped the programme (#36): one winner rule had said yes, for the wrong reason; the other had said yes, not enough.

What to take from it

A backtest that beats chance, holds across settings and stocks, and clears a floor can still be measuring the market, not the idea. The only defence is to ask, before the money moves, what the passive alternative at exactly the same geometry would have made — and to write the answer down beside the verdict whether it flatters the strategy or not. This lab now does, by construction. Nothing here places an order.