The Highlight Reel — What a Perfect Winning Streak Costs to Manufacture

What we granted, and what we actually tested

Granted, not measured: that the winners in a seller's reel are usually real trades, that the teaching behind them is often correct, and that the seller usually believes it. We are not alleging invented trades anywhere, about anyone. The trick we are demonstrating needs no dishonesty at all — it is selection, and selection is free.

Tested: one thing only: how large a hidden denominator it takes to manufacture a winners reel of a given shape out of a population we independently measured to lose money. Five reel sizes registered before either file was opened for values, two friction populations, two arms — 20 cells, all of them below, none dropped.

§1 · The locked pre-registration

There is no hash lock on this pre-registration, and we would rather say that than imply one. It is dated 2026-08-15 and written, by its own account, before any statistic was computed from either trade file — but 12564a3 carried the pre-registration, the script, the results file and the write-up into the repository together, so version control cannot witness that ordering for us. What it can witness is the disclosure the pre-registration itself carries: a table naming every thing that was read before it was written, and confirming that no distribution, quantile, win rate or profit figure was among them.

The falsifiable question

How large a hidden denominator is required to manufacture a given winners reel out of a strategy we measured to lose money — and can it be done at realistic screenshot counts at all?

• A yes looks like: the reel fills easily at ten and twenty screenshots and the hidden count behind it is enormous, so a compelling reel carries no information about the strategy it was cut from.
• A no looks like: a measured loser cannot fill a reel at realistic counts, or fills one only at a ratio small enough that a buyer could plausibly ask for the denominator and get it. Then the reel IS evidence — weak, but real — and this episode overstates its case.
• Ambiguous residue: a reel that fills but reads thin, which is where the magnitude cap in §5 actually lands. We report it as a limit on the demonstration rather than pretending it away.

Our stated prior: we registered no numeric prior, and that is a real weakness

This lab is a demonstration, not a hypothesis test. It has no p-value and claims none, and the pre-registration says so in its own first section. What it registered instead were the reel sizes, the ordering rule, the two dollar conventions, and a three-part falsification rule — all before either file was opened for values. That is weaker than a stated numeric prior, and we would rather say so than dress a demonstration up as a test. One of those three criteria then turned out to be unsatisfiable by construction, which is what the goalposts table below reports.

The primary configuration (locked)

Deliberately NOT tested — named in advance

1. Any real seller's actual record. A private brokerage account is not auditable from outside. That is permanent, not a budget problem, and the episode says so on camera rather than hiding it in a footnote.
2. Whether any particular reel came from a strategy that loses money. This runs one direction only: it shows the thing is manufacturable, never that it was manufactured.
3. How often it happens. No frequency claim is made, and none is supportable from this design.
4. The big-multiple reel. Arm 1's take-profit is fixed, so the population has no upside tail to cut from. This demonstration covers the perfect-streak variant and not the other one.
5. Any friction assumption other than the two registered. If someone shows the reel is unmanufacturable at a lower one, that is a new arm, not a re-cut of this one.

The goalposts — all required to proceed

3 of 4. The one that failed is criterion (b), and it is the most interesting line on this page, because it could never have passed. Gross R in this population is capped at exactly +1.000 — 0 of 244,328 trades exceed it — and net R subtracts a friction charge whose smallest value anywhere in the file is 0.003669 (computed on this page at build from the source experiment’s per-trade file, not a figure in the lab’s written record), strictly positive. So every net R is below 1.0, every subset mean is below 1.0, and a bar set at +1.0 R fires with probability 1 on any data this engine can produce. We wrote that criterion ourselves, before the run, with an authoring note calling it a real bar, not a gimme. That reasoning was backwards: the fixed take-profit is precisely what makes the bar unreachable. And the other two could not have fired either. (c) is arithmetic: the ratio is 244,328 divided by the reel size, which is 2,443 at the largest reel we registered, against a bar of 100. (a) was set against a win rate our own page had already published: about 109,000 winners against a bar of 20. Three ways to be wrong, and not one of them could ever have happened. The demonstration still stands; the test we wrote around it was not a test. We are publishing the defect instead of quietly repairing it, and a corrected criterion is deliberately not authored here, because writing one after seeing the data is the exact move pre-registration exists to prevent.

§2 · The headline arithmetic

§3 · Was it the path? The stress tests

Does the friction assumption carry the answer? No. Run the same cut on gross R — no costs at all — and the reel gets cleaner, not weaker: mean and worst are both exactly +1.000 R at every registered size, because 77,100 trades closed AT the take-profit cap (R = 1.0). The manufacture ratio does not move at all, because it is a property of the denominator and not of the costs. Both populations are in the ledger below.

Does it depend on stopping at a flattering reel size? No. All five registered sizes ship. The ratio falls as the reel gets longer — N=5 48,865.6 : 1 · N=10 24,432.8 : 1 · N=20 12,216.4 : 1 · N=50 4,886.6 : 1 · N=100 2,443.3 : 1 — and even at a hundred screenshots, a length no sales page has ever run, it is still 2,443.3 trades for every screenshot. The reel stays 100% winners at every one of them.

Is it just a big-sample artifact? Partly — and that is the finding, not an objection to it. The tenth trade sits at the 99.9959th percentile. A large population contains a beautiful streak by construction, which is precisely why a streak shown without the count it was drawn from cannot be read as skill.

§4 · The controls

The reproduction gate. Before a single reel was cut, the pre-registration required the published verdict statistics to be re-derived from the trade-level artifact using the source engine's own formula. 7 of 7 cells reproduced to the precision of the source record — n, net expectancy, net win rate, net profit factor, and the three gross equivalents. Had it failed, the pre-registration required the mismatch to become the headline and every reel below to be labelled provisional.

The second population, and it is the control the concept actually needs. Arm 2 cuts the same reel from a completely different book — 5,742 modelled 1-1-2 option trades. That population is net-positive: expectancy +0.182 credit-multiples, 78.4% win rate, profit factor 1.34. It is a ruinous-tail specimen, and this report will not describe it as anything else. Its reel is indistinguishable in shape from arm 1's — N=5 mean +1.867×, 1,148.4 : 1 · N=10 mean +1.825×, 574.2 : 1 · N=20 mean +1.744×, 287.1 : 1. A reel cut from a book that makes money and a reel cut from a book that loses money are the same shape. That is the whole point: the reel does not carry the information, in either direction.

And a limit on that control that we are stating rather than burying. Arm 2's ledger carries a friction column that is zero on 5,742 of 5,742 rows, so its net and gross arms are byte-identical. Of the 20 registered cells in the ledger below, 15 are distinct measurements and 5 are the same cells reported twice. They ship anyway, because the pre-registration said all 20 ship.

§5 · The best objection, pre-registered and answered

“Of course you found ten winners. You had 244,328 trades.” Correct — and that is the finding, stated as plainly as we can put it. The output of this lab was never the streak; it was the ratio, and the ratio was registered as the metric before anything was read. A big population contains a perfect streak the way a big lottery contains a winner. The question a buyer needs answered is not how good are these ten but ten out of how many — and that is the number nobody hands over.

“Your reel is not the reel sellers actually post.” On one axis, true, and it is a real limit on this demonstration rather than a quibble. Arm 1's strategy takes profit at a fixed level, so its best trade and its tenth-best trade are nearly the same size: 0 of 244,328 trades finish above +1.000 R. That means we can manufacture the perfect-streak reel — ten trades, ten winners — and we cannot manufacture the big-multiple reel, the one with a 3× here and a double there. Demonstrating that variant needs a population with an unbounded upside tail, and we did not run one. It is named as owed, not quietly skipped.

§6 · Their own terms

Six things on this page run against us. We would rather print them than have them found.

1. The pre-registration is dated, not hash-locked, and we are saying so. The pre-registration, the script, the results file and the write-up all landed in the same commit, 12564a3, on 2026-08-15, so version control cannot witness pre-registration-before-code for this lab. What the file does carry is a disclosure table naming exactly what was read before it was written — a schema, a row count, the engine's friction definition, and one already-published figure — and confirming that no distribution, quantile, win rate or profit figure was among them. That is a stronger claim than silence and a weaker one than a hash.

2. Arm 2's registration boundary is weaker than arm 1's, by the lab's own admission. While reading arm 2's schema the author saw its outcome value counts before writing the dollar-scale amendment. Arm 1 remains fully blind-registered; arm 2 does not, and the pre-registration says so rather than pretending otherwise.

3. Arm 2 makes money and does not carry the claim. Expectancy +0.182 credit-multiples, win rate 78.4%, profit factor 1.34. What it demonstrates is that a reel from a book with a ruinous tail is indistinguishable from a reel from a good one. It is not a reel from a measured loser, and nothing on this page presents it as one.

4. Ten of the 20 registered cells are duplicates. We publish 20 because 20 were registered; the honest denominator of distinct measurements is 15, and §4 says why.

5. Every dollar figure here is a convention, not a measurement. Arm 1's faces are R multiplied by $500 risked per trade, chosen before anything was read. Arm 2's are credit-multiples multiplied by a $1,000 average credit, which is the seller's own stated figure and not something we measured. Strip those words off either one and what is left is an invented number.

6. Arm 2's 5,742 is the count of trades we modelled. That is the scope of the figure everywhere it appears on this page, and the only scope it carries here.

§7 · The complete ledger — all 20 rows

§8 · What would change our mind

1. A seller publishes a complete, time-stamped, third-party-verifiable record — broker statements, or a verified-track-record service — and it survives a fair null. The concept then shrinks from they never show it to ask, because some will answer, which is a better episode than this one. It would also supersede this lab outright: a synthetic demonstration is a weaker instrument than a real denominator and should be retired the moment one exists.
2. A corrected criterion (b) is authored and the lab re-run, and a measured loser then fails to produce a plausible reel at realistic screenshot counts. We left this live on purpose. Writing the corrected criterion today, with the data already in front of us, is exactly the move pre-registration exists to stop.
3. A specimen surfaces whose denominator is reconstructible from public record — disclosed filings, for instance. That turns this from an explainer into an experiment and skips the synthetic detour entirely.
4. A disclosure regime changes the baseline — a platform or a regulator makes position counts standard for paid trading education. The tell stops being the norm, and this page should then be read with the date on it.
5. The source experiment is re-run and its n or its expectancy moves. Then this lab re-runs against the new artifact rather than keeping a stale denominator, and every number on this page moves with it.

§9 · What survives

The portable part is one question, and it costs nothing to ask: ten out of how many? Not are these winners real — they usually are. Not is the teaching any good — it often is. The count is the thing that turns a set of screenshots into a record, and it is the one thing that is never on the page.

And when the count does not come — which is most of the time, because for a private account it genuinely cannot — the fix is not to keep asking. It is to price the claim against what no skill at all would have produced. That is a test you can run from outside, and it needs nobody's permission.

The last thing worth carrying is the one that cost us something to print. We pre-registered a three-part falsification, it fired, and the criterion that fired it was one we had built so that it could not do anything else. On inspection the other two could not have fired either. We published that instead of repairing it quietly — because a pre-registration you edit after the run is a pre-registration in name only, and because this is the episode where we are asking other people to show their whole record.

§10 · Want to run a test like this yourself?

Everything we make is free — the reports, the ledgers, all of it. The only thing that costs us real money is the market data.

Get every report + the 7 Questions one-pager (free)