Any signal service can publish a number. The question worth asking is what the number was allowed to do before it reached you — which ties it won, which trades it quietly dropped, and whether anyone checked it on data it had not already seen. Here are our rules. They are the reason our figures are sometimes duller than the competition's.
The code that decides whether a signal hit its target is the same code that resolves the public ledger and the same code that runs our backtests. There is no separate "reporting" version. This sounds obvious and almost nowhere is it true: the usual pattern is a live tracker and a research script that disagree quietly, and the one that flatters gets quoted.
When a single candle covers both the target and the stop, nobody can know which was touched first. We record the stop — every time, in the live tracker and in research alike. Resolving those the other way would lift every result we publish, which is exactly why we do not.
A signal that drifts barely upward and then expires is a "win" that made nobody anything. Win rate rewards that; R does not. R is profit measured in multiples of what the trade risked, so a small win cannot cancel a full stop-out, and no amount of rounding turns a loser into a neutral.
Every figure we take seriously carries a confidence interval and the sample size behind it. With a few dozen trades, an average moves further on luck than on skill — so an average alone cannot tell you which one you are looking at. When an interval includes zero we say the result is unproven, rather than quoting the midpoint and hoping.
Search enough configurations and one will look excellent by chance alone. We correct for how many were tried, and we check the winner on a period that played no part in choosing it. Plenty of ideas have passed the first test here and failed the second — and failing the second is the whole point of having it.
Setups the engine is not confident enough to put forward are marked SHADOW. They are generated, tracked and resolved on the same ledger as everything else — they are simply not presented as trades to take. Deleting them would make our record look better; keeping them is how the record stays a record.
The engine regularly re-runs its own recent history and measures what it actually produced, interval included. If that measurement no longer supports the results, it moves every signal to SHADOW automatically — it keeps working and keeps being tracked, but it stops presenting anything as actionable. Nobody has to notice first.
Coming back is deliberately not automatic. Going quiet is cheap and reversible; speaking again is a claim about your money, so a person has to make it. A system that could promote itself would eventually promote itself on a lucky week.
It retrains on fresh market data, then has to beat the model already in place on data neither of them was trained on. "Good enough" is not the bar — the existing model is free, and a change has to earn itself. Most weeks the challenger loses and nothing is swapped, which is the system working rather than stalling.
Rules about what to leave out matter more than rules about what to show, because everything below would make us look better.
It makes the scoreboard honest, which is a different and smaller claim. Markets change and any edge can stop working; what these rules guarantee is that you will be able to see it in our own numbers when it does, rather than hearing about it later.