The edge that was real and worth half a pip

Ask whether my engine has found a trading edge, and the honest answer from its first end-to-end hunt is: it found structure that is statistically real and economically negligible — leans of half a percent to two percent from a coin flip, undeniable at fifty thousand observations and worth a fraction of a pip before a spread is paid. I call the shape Significance Without Size: a p-value that is small because the sample is large, sitting beside an effect that cannot pay for its own execution.

This is a post-mortem rather than a study of somebody else’s idea. The hunt was built, run, adversarially checked and reported in a single day, 12 August 2026, and the report it produced is blunt about its own design errors. This note is that report, translated for a reader who does not have the engine.

What the hunt asked

The atom is one entry bar. From its close, does price touch plus T before minus S within H bars? The engine records four outcomes and never three: up first, down first, both barriers inside one bar — where the order is unknowable at bar grain and is never assumed — and no touch at all, the appetite that did not pay. A walk the tape cannot resolve is excluded and counted.

A cell is a pair, a target band and a volatility state: bands of three to eight, eight to twenty and twenty to fifty cost units, and the regime label frozen at the moment of entry, point in time. Every hourly stamp is an entry. The cell’s statistic is the probability of up-first among the decided walks, with the no-touch and ambiguous fractions printed beside it and never folded into its denominator. Over fifteen and a half years and twenty-eight pairs that gave three hundred lit cells.

Confirmation was forward and blind. Flag dates were simulated every ninety days through the history; the discovery pass at each date saw nothing after it; a mechanical rule flagged a cell when its deviation from a coin flip, scaled by the square root of its sample, reached one, with a floor of a hundred decided walks; and confirmation was measured only on the ninety days that followed. For every flag the ledger records whether the lean kept its side, and how much of it survived.

What it found

Five thousand four hundred and forty-nine flags, of which two thousand nine hundred and eighteen could be confirmed. Broad mechanical flagging fails: unique-cell persistence came out at 0.330, which is to say the typical flagged anomaly mean-reverted once it had been flagged.

A concentrated set survives. Twenty-eight of the eighty-five cells that could be ranked cleared a nominal p of 0.05, and twelve survived a Bonferroni correction across the family. Twenty-two of the twenty-eight were yen and franc crosses in the calm volatility state — one coherent phenomenon, which the report reads as funding and haven drift, rather than twenty-eight discoveries.

That is the part a marketing page would stop at. Twelve cells, Bonferroni-clean, fifteen years, walk-forward. The report did not stop there.

The central finding: put the effect size beside the p-value

The persistence statistic measures the direction of a lean, not its size. When the effect sizes are placed beside the p-values, the picture inverts:

Deviations of half a percent to two percent from a coin flip. The p-values are tiny because the samples are fifty thousand and more, not because the effect is large. Half a pip of gross expectancy does not survive a spread, let alone a commission. The report’s own sentence for this is the one I would carve above the door: a p-value without an effect size beside it is a dangerous instrument — and the engine’s edge report now refuses to print one without the other.

Four reasons the design could not have found tradable edge

The report lists them in order of damage, and the first is the one it calls the biggest single design error.

Symmetric barriers force a coin-flip baseline. With the target equal to the stop, a market that is close to a fair game gives up-first about half the time by construction. The design can therefore only ever detect tiny deviations from a half — and it has converted a question about magnitude into a question about direction, which is the one question this corpus has already answered “no” to at every resolution the engine has measured.

A touch probability is not a profit and loss. Whether price reaches one barrier before another says nothing about how far it travels, how long the capital is tied up, or what the paths that never touch were doing meanwhile. Two cells with the same up-first probability can have opposite expectancies once path and time are priced.

No cost model was armed. A cost-floor law exists in the engine — a target must be at least three times the loaded cost — and it renders unarmed, because the commission line item does not yet exist. Everything in the table above is pre-cost, and at these effect sizes cost is not a haircut. It is the verdict.

The conditioning was shallow. One state family, one horizon of twenty-four bars, one grain at a time, while the engine holds around twenty other measured conditioners that have never been crossed with the lattice.

The part that makes the null worth keeping

A null result earns its keep by having been able to come out the other way, and this one was. The machinery flagged, confirmed forward, corrected for the family, and then declined the broad claim it could easily have made from twelve Bonferroni-clean cells. Any graded claim in the engine goes through a registered-trial ledger — one pre-declared endpoint per hypothesis, a draw budget derived rather than chosen, a jointly drawn null, a negative control — and the two hypotheses this hunt drafted are waiting on that ledger rather than being announced.

The report’s diagnosis fits in one paragraph, and I will give it in the report’s own shape rather than soften it. The lattice asked which barrier gets touched first at equal distance — a direction question with a mechanically even baseline — on a market where the engine has already shown direction to be unforecastable at every resolution it has measured. It validated the answer with a statistic that is blind to magnitude, on samples large enough to make a half-percent deviation look like a one-in-ten-billion discovery, with no cost model armed. The machinery is sound: it found real structure, refused the broad claim, and reported honestly. It was pointed at the one question the corpus keeps answering no.

Where this leads

The report’s forward path is engine work, listed in order of expected value and none of it a claim: barriers made asymmetric, so the question becomes payoff geometry rather than direction; barriers scaled from the one quantity the engine can actually forecast, which is volatility; the cost floor armed, so the lattice can refuse an untradable cell by itself; the conditioners crossed instead of taken one at a time; expectancy and the full excursion distribution in place of a touch probability; the relative-value tier run through the same forward pipeline, because it is non-directional by construction; and only then a grade, through the registered-trial machinery that is built and planted-truth tested.

Two disciplines on this site exist because of exactly this shape of result. The Overfit Auditor bounds what a record means when it was selected from among many attempts — which is what eighty-five ranked cells are. The Epistemic Harness keeps the trial count honest while the search is still running, which is the number every correction above conditions on. And the Reproducible Verdict Kernel is the reason the caveat at the top of this note is temporary: when the hunt’s artifacts publish, their digests join the registers, this note gains its dataset block, and the numbers above become recomputable rather than stated. Until then the kill ledger is where the engine’s dead hypotheses are counted, and a half-pip edge is, for now, one more entry waiting to be written.