This is what the workflow produces.Go and read a round.
Every desk on the board below is what you would build: an entry in one configuration file, deciding on the same clock as all the others, with everything it has ever done on the record.
paper trading only · Node 23.4+ · Windows · the board below is free, read-only and not a trial
They are characters, not rows.
Each one carries its own mandate, its own simulated money and its own words about what it just did. Four of the 22 are below.
6 links, and none of them asks you for anything.
Not a gallery of what the board did in June — what it is doing now. No product in this house let a stranger watch anything run before paying. This is the fix.
What it draws, and what it refuses to leave out.
Captures of the public instance, unretouched, each stamped with the day it was taken.
Every desk is drawn, not a selection: one square each, five of them red on the day. The four screens below open on the board itself.
It ran a registered study on itself, and nothing passed.
The question was written down before any of this was for sale: do these rule families have an edge at this cadence, net of this board's own cost model — with the pass and kill rules fixed in advance so the answer could not be chosen afterwards. 4 families, 5 windows, control baselines, and a year of untouched data.
“four families, five windows, three control types, one out-of-sample year: zero edges survived.”data/probe/SYNTHESIS.md — the study’s own tally
| Family | Final | What that means |
|---|---|---|
| momentum | FinalREFINE | What that meansNo pass anywhere, and the sign flips between the two halves of the year — whichever half you pick decides what it looks like. Unresolved rather than dead. |
| donchian | FinalKILL | What that meansKilled in the base variant. Its one bright cell shrank materially once a ledger defect was corrected, and the trailing-stop re-probe never passed either. |
| meanrevert | FinalKILL | What that meansThe clean one: negative and below the control median in every cut, in the full window and both halves, with no caveats. |
| carry | FinalREFINE | What that meansStarved at the registered threshold — the mechanism barely armed. Loosening it produced the study's only passing cell, on one symbol, which the out-of-sample year then killed. |
The one passing cell — one family on one symbol — went to a registered follow-up on an untouched prior year, locked before the run. It came back a kill on all three symbols, so the pass was classified as the lucky cell it looks like. In the pre-registration’s own words: “this probe found no edge to ship” — two families dead at this cadence, two genuinely unresolved rather than disproven. That distinction is kept because losing it would make the study sound more conclusive than it is, in the seller’s own favour.
The verdict rule was mis-implemented, and the write-up says so
The scoring code folded two separate kill conditions from the pre-registration into one, and scored a cell with no trades as an automatic kill rather than as no evidence. Re-derived by hand against the registered rule, one family's verdict moves. The finding is in the shipped write-up with the line numbers, next to the verdict it changes.
The ledger silently dropped a fifth to a third of the trades
The simulated ledger discarded each family's own stop-and-reverse entries — the very trades the registered strategy definition calls for. Re-run with a faithful ledger, one family's kill holds or worsens and another flips to unresolved. The re-run is published; the family that flipped was never re-probed, and the write-up says that too instead of quietly leaving the old verdict standing.
Anyone can publish a negative result. This one went looking for defects in its own scoring code, found two that move verdicts, and re-ran everything with them fixed.
It is also the only claim here you cannot open in a browser: the study is a folder of files inside the archive, not a website. What is public is the promise about it — a delivered archive missing a file it lists is refundable at any time, and the whole definition of a defect is written out on the operating manual, before the sale rather than after it.
Three mechanisms, all of them shipping source.
An honest record is not a promise anybody makes. It is what is left when the parts that could flatter it are taken away. Each of these carries the screen where you can watch it work.
The exits are enforced by the house, not by the desk
A desk's stop and target are not promises it keeps. Between rounds a monitor tick marks every position to market and fires stops, targets and liquidations on the price extremes, whatever the desk would prefer — and a desk that has been switched off but still holds a position is named in the runner's log, because its stops stay enforced while no turn will ever close it. Before any of that, the risk guard rejects orders that break the house rules and shows the desk the reason next round. When a market is shut and the last print has gone stale, the guard refuses to fill at a frozen price at all: managing an existing position stays allowed, new fills wait for the open.
Check it ↗ the tape — every event as it landed, including the exits the house fired between roundsThe luck test, and the fact that it usually says no
Beside every desk the board prints whether that curve can be told apart from luck — a bootstrap of the desk's own daily returns, a thousand resamples, a confidence interval read against zero. Under five days of returns it refuses to say anything at all and prints TOO EARLY. An interval that straddles zero reads as consistent with luck. Only an interval clear of zero in either direction earns BEYOND LUCK or BELOW LUCK, and on the day these captures were taken every curve on the board was still on the wrong side of that line.
Check it ↗ the standings — the verdict beside the desk it judges, leader as well as last rowThe honesty line is counted, never asserted
That headline used to be a hard-coded sentence saying every curve was consistent with luck — true the day it was written, and a fabricated claim the first day any interval cleared zero. It is now computed from the set: how many of how many. A number that is counted from the record can never harden into a claim about it, which is the difference between a product that is honest and a product that says it is.
Check it ↗ the counted line under the standings — how many of how many, computed from the set
The leading desk on the day of the capture, and the board’s own sentence about that lead finishes TOO EARLY TO CALL. That is a verdict rather than a hedge — one of four, and only two of them mean a curve can be told apart from chance:
- TOO EARLY TO CALL
- Not enough closed trades for the test to say anything.
- CONSISTENT WITH LUCK
- Enough trades, and this curve is what chance looks like.
- BEYOND LUCK
- Far enough from chance to be worth a second look — on this sample, at this cadence.
- BELOW LUCK
- The same test pointing the other way: worse than chance by the same margin.
Under the standings the board counts them — N OF M CURVES STILL CONSISTENT WITH LUCK — recomputed on every render, so it reports the day that count falls rather than hardening into a claim.
And what it refuses to print
- No write control of any kind on a public board. The fence works at the route as well as at the button: every mutating endpoint answers 404 as its first statement, so a fenced process cannot be made to write by anyone who knows the URL, and the controls are absent from the first paint rather than removed after it.
Check it ↗ the board's own status — this instance answers "readOnly": true - No forecast of its own. The product predicts nothing; a desk's reasoning is that desk's own words, printed as written, and the prompt it was given and the raw response it returned are both stored with the turn.
Check it ↗ any desk's log — the prompt it was given and the raw answer it returned - No invented turn. In demo mode every turn record says a rule engine decided it, so no turn can claim a model made a decision that was never made, and the board's own status endpoint says it is a demo.
Check it ↗ the same payload — the public board answers "demoMode": false - No quiet gap. A round in which every desk stood aside is written up as fully as a round with trades in it — most rounds, most desks do nothing, and a board that showed only its trades would be a highlight reel.
Check it ↗ the Reading Room — a round nobody traded, written up as fully as one with trades
What $59 does not buy.
On purpose, and stated in the archive's own opening pages rather than discovered on day three.
Not a signal service
Nothing here tells you what to trade. There is no alert to act on, no call, no target, no copy-this-desk. A decision is a record of what a piece of software did with pretend money — written down so you can read the reasoning, not so you can follow it.
Not a money-maker, and no claim of one
This product claims no profit, no edge and no performance. Its own pre-registered study finished with no rule family passing, and those write-ups ship with the negative results intact. If a desk is up, that is the sample you are looking at, not evidence.
Not live trading
There is no order routing to any broker or exchange anywhere in this codebase, no live-execution module and no exchange credentials. Every fill is simulated by the paper engine. You cannot switch this on, because it was never built.
Not a hosted service
Nothing runs in our cloud, no account is created, no subscription is charged and nothing phones home. It is source that runs on your machine against your own model accounts, and it keeps working with this archive and nothing else.
Still deciding? support@pulltrade.app answers before the sale as well as after it.