What this report contains

Where the numbers come from

We have not found the edge. We have built something rarer:
the machine that could prove one.

In the trading software industry, hundreds of vendors claim an "advantage." Almost none can substantiate it — because doing so requires an apparatus almost nobody builds: pre-registration, placebo controls, null hypotheses, a register of one's own errors. That apparatus has existed in full at OrderFlowAi since July 2026. It is the product of twelve months of work and — this is the unusual part — of 54 findings we retracted ourselves rather than sold.

The five core statements

What an "edge" actually is

An edge is not a feeling, not a hit rate, and not a good month. An edge is a measurable, repeatable deviation from chance that survives a serious attempt to refute it.

The difference between describing and predicting

The most important and most frequently overlooked distinction. A model can describe a market movement perfectly and still be worthless — because it only describes the movement once it has already happened.

The five traps — and how we fell into every one

This list is not theory. Every trap cost us time; every one is recorded with its date and its cost in the retraction register. We show them because an auditor will ask about exactly this.

The first lines: a tool, not an oracle

On August 1, 2025 the first internal alpha goes live. It predicts nothing. It displays something — and specifically something most trading programs cannot render at all: the order book, the list of all open buying and selling intentions, live and at full depth.

What distinguishes this data source from a price chart

From displaying to predicting — and the first temptation

In March 2026 the display tool becomes a learning system. In four weeks the building blocks that still carry the project today come into being — and at the same time the error that will later cost a quarter of a year to correct.

The first temptation: 70–85 %

CORTEX is born

On March 28, 2026 the neural engine receives its final name. On April 17 it ships to customers as CORTEX v0.3.0. Between those two dates lie three weeks in which a model becomes a product.

What the 41 features consist of

Each feature is sampled every three seconds; the model sees the most recent 30 steps, roughly a minute and a half of history. The selection was made on domain grounds — which of them actually carried information was only measured in July 2026 (result: the state vector as a whole is overfitted, see page 17).

The euphoria — and the first numbers that disagreed

For two months everything runs at once: product development, marketing, customer acquisition, strategy work. It is the most productive and simultaneously least scientific phase of the project.

The aggregate that arrived two months late

The break: fourteen days in which nearly every hope fell

In July 2026 testing is adequately powered for the first time — with enough data, with correct null lines, with cross-validated procedures. The result is unambiguous and unpleasant. It is also the moment a software project becomes a research project.

CORTEX-ONE — one brain instead of many opinions

The break is followed by reconstruction. The diagnosis was not "the model is too weak" but: there were too many divergent truths inside the system.

From mirror to instrument

The most recent and so far most important diagnosis: CORTEX was a mirror. It was fed from the same raw quantities as the trading strategy — and therefore inevitably confirmed whatever the strategy already thought. On July 28 the two agreed on 33 out of 33 trades.

Three findings in three days

Twelve months at a glance

What each phase contributed

The retraction register — 54 findings we overturned ourselves

The most unusual institution in this project is a file containing every statement that has proven false. Not as a footnote — as a standalone, machine-monitored register with four mandatory fields per entry.

Four falsifications in detail (1 of 2)

Selected not for drama but for instructional value: each shows a different way in which an apparent advantage arises where there is none.

Four falsifications in detail (2 of 2)

Pre-registration: the rule precedes the calculation

The procedure comes from clinical research. There it is mandatory, ever since it emerged that studies with unwelcome results were systematically going unpublished. In financial software it is virtually unknown — and therefore the largest available lever.

The placebo control — the test that cost us the most

In medicine a control group receives an inert sham treatment. If the real drug does not perform markedly better, the observed improvement was not evidence. The same principle transfers to market data — and almost nobody applies it.

The decay curve — separating describing from predicting

Perhaps the most elegant test in the whole apparatus, because it answers a question that is otherwise never asked: how long is a signal valid for?

Scope — who does this number actually belong to?

The most frequent class of error in this project, appearing five times in different disguises: two individually correct numbers describing different slices of the same reality — and therefore not comparable.

Tools that test themselves — and guards proven to be watching

A checking program that always passes is worthless. You cannot tell the difference from its output — only from whether it can still turn red at all. That is why every tool here has two built-in parts: a self-test and a counter-check.

The difference between technical and domain acceptance

The four forms of "documented but not enforced"

What is established today — the balance after twelve months

This is the most candid page of the report. It separates cleanly what has been measured, what has been refuted, and what remains open.

Established negative — tested with adequate statistical power

July 29, 2026 — a lead, explicitly not a verdict

On the day before this report was written, the CORTEX-rebuilt strategy ran for the first time with a repaired market-speed sensor. The result is striking enough to report — and weak enough that we do not call it a finding.

Nine stages — the cascade has fallen, the endpoint has moved

The July research plan consisted of nine consecutive test stages with exactly one confirmatory endpoint. On September 1 stage 5 fell — and with it, as the pre-registration states verbatim, everything above it. That is not a breach of the plan but its intended outcome: the rule preceded the calculation.

Three null findings in five days — and what they prevented

Between September 1 and 5, three pictures that felt right fell against their time-of-day-matched null line. None was a setback: each prevented a rebuild that would have cost weeks and devalued data. All three were pre-registered; all three are in the retraction register.

The delta microscope — 56.7 % of volume has no certain sign

Every delta display on the market claims to know who the aggressor of a transaction was — buyer or seller. In August we measured how often two equally legitimate classification rules disagree on the same raw data. The result changes what an "accurate delta" can even mean.

Four empty reader classes — and one axis that holds

Gate 1's "NOT LEARNABLE" explicitly applied only to a linear, pointwise reader. The obvious question: do stronger model classes see more on the same data? In late August it was answered systematically — with a twofold result.

Fifteen million configurations — one mirage and one candidate

On August 29 two exhaustive searches ran across the entire recorded data inventory. The first proved that our validation funnel would let a real edge through. The second found a candidate that passes every first examination — and along the way exposed a second one that passed none.

The replay machine — would the strategy have won the other way round?

On September 9 and 10 all 105 replayable strategy trades since July 29 were replayed tick-exact on our own tape — with the real price path, stop, target and costs (0.64 ticks per contract). Four pre-registrations, one look per variant. The question behind it was uncomfortable: is the fault in the management — or in the entry itself?

The exchange as referee — 97.5 % instead of 69.9 %

Since September 10 we have had real CME order data — every single order with its side. For the first time it was possible to check whether our rule, which assigns an aggressor to every trade, is right: not against itself, but against the exchange. Every quantity derived from the delta — absorption, exhaustion, CVD, divergence, order flow — depends on it.

ONE CORTEX decides — and first we measured what slows it down

Between September 14 and 22 the signal chain was rebuilt: the direction comes from CORTEX alone, a CORTEX counter-signal can close the trade before the first target, and the strategy's own filters now only log. On September 23 the new era went out to all customers as release 0.5.30 — with a learned state that starts at zero.

Why certification is possible — and who could issue it

The question is not "is there an authority that certifies trading advantages?" — there is none. The question is: which verifiable properties must a result have for established audit bodies to be able to attest to it? To that there are clear answers.

Four real routes, each with a clear scope

The dossier — what already exists today

A certification dossier cannot be assembled retroactively; it has to run alongside. That is why the chain of evidence was built as a machine from the outset rather than as a document. This page lists what an auditor would find today.

The claim to being first — what exactly would be "the first time"?

A claim to primacy must be phrased precisely or it will not survive scrutiny. This page phrases it narrowly enough to be defensible — and states what it does not assert.

Why the narrowness of the claim is its strength

The autonomous co-founder — what happens when the AI thinks along instead of executing

This project is built by one person and one AI agent. In 2026 that alone is no longer remarkable. What is remarkable is the operating mode: the agent is configured not as a tool but as a co-founder under an obligation to take initiative.

The compounding effect — measured, not asserted

The thesis: because the project is co-developed by an AI agent, its development speed and quality grow with every model generation. That thesis is testable — and we measured it against our own version history.

The value chain — five stages, four of them independent of the edge

The decisive strategic move of July 20, 2026 was decoupling: the product sells what demonstrably works, and the research runs separately. As a result, company value does not hinge on the outcome of a single experiment.

Risks, stated honestly

A report that documents 22 of its own errors and then falls silent on risk would lack credibility. This page names what can go wrong — and what has already been done about it.

Where we stand today — in numbers, not intentions

This chapter is the compass. It answers three questions: Where do we stand? What is missing before a verdict? And: what does the strategy base its decisions on? All figures are measurements taken on September 23, 2026; where an older measurement point applies, its date is stated.

What is missing — and what the strategy decides on now

The programme allows two admissible outcomes, and both count as a result: A — a demonstrated edge. B — a defensible no. A third outcome ("we still don't know") would be the only real failure. This page states what each of the two still requires.

What happens next