ENGINE

Surplus rule // only if a surplus exists

Realized net trading PnL, after fees. Not a transfer tax. Not a team wallet. A loss distributes nothing.

Vetoes // the trade is usually the refusal

The rules, in order The stress tapes Epoch accounting

Published suite // shadow · post-graduation · 8.5% size

Generated by engine/stress.js and checked in as engine-report.json. The test fails if this table drifts from that file.

Holdout return 25.6533. Confirm return 4.7216. Old book -0.4511. Win rate 0.3929. Profit factor 1.5566. Max drawdown 0.2588. Launches overlap on one book (the overlapped harness), so the position cap and the daily loss cap both execute: peak 3 of 3 concurrent, 0 refusals after the 34% daily cap under these costs (it trips under the stress costs). Seeded synthetic launches, not a pump.fun recording — and a positive figure here is a statement about the generator’s hand-set runner path, not evidence of anything on pump.fun.

Why this figure flipped sign on 2026-09-27. Until then this page showed a 45% loss on both folds. An audit found the rule was selling a creator dump and buying the same coin back on the next bar — the dump flag clears once the sale bar has passed — so every synthetic fade was traded twice, the doubled losses tripped the daily cap, and the cap then refused nearly every runner. The loss was that artifact. With one position per launch, the same generator, costs and seeds return 25.65 and 4.72. Read that as: the generator hands out a 1.9× run in 96 seconds to 40% of graduations, and a rule that holds through it compounds. Whether pump.fun does that is the question the observer below measures directly.

Every figure above is charged at the base cost scenario below, on a book where launches overlap. Both of those used to be otherwise, and both flattered the result. This page published a holdout return of 1.919: that number needed 0.4% slippage, 3% to exit into a creator dump, a fill priced at the mark that triggered it, the whole round trip charged on the entry notional, and one launch at a time on a fresh book — which meant the position cap and the daily loss cap, both listed above as risk controls, had never once executed. pump.fun’s own fee schedule is 1.25% per side. Nothing charged the cost of submitting an order either: Solana bills you whether the trade works or not, and a fixed cost per fill halves how many trades the book can afford before it hits the cap.

The old book, on the same harness with the same guard, scores -0.4511 with 95 trades: its wide gates enter launches that never graduate and pay the costs. The selected rule beats it because it waits for graduation, and that is the whole of what the generator can say about the rule.

The cost scenario decides the sign. The same rule, seeds and generator return 25.65 under the base costs and -0.45 under the stress costs: three points of slippage and a second bar of latency are the whole difference between a 26× and a 45% loss. npm run search sweeps 384 configurations of the entry, exit, concentration and size parameters, then 64 more over the risk and entry axes, on a training seed the evaluation seeds never touch; under the base costs configurations beat cash on training now, and the search says so in those words. It also says what it has always said: the population is synthetic, the winner is the best of hundreds of draws, and nothing here is a live edge.

The engine cannot price size. Not one of the six places a fill price is computed has the order size in scope. Under the refuted optimistic costs the stake sweep runs from 100.2% at 85 basis points of the book to 35,985,678% at 1700 — a 20× stake compounding into a figure five orders of magnitude larger — with the same 448 trades at every stake. A twenty-fold larger order fills at the same price and changes no decision. So the 8.5% stake was chosen by an objective with no size term in it, and nothing here is a sizing conclusion. No impact coefficient has been invented to paper over it: nothing in the repository can calibrate one.

Read the daily figure. Two thousand synthetic launches span about two days, so a total return is two days compounded rather than a rate. The holdout returns 730.8% on the first day and 220.8% on the second. Under about 1,250 launches the whole run is a single day, which is why the search refuses to run that short.

And the sign turns on a number nobody measured. The synthetic population assumes 0.4 of graduations run rather than fade, and that a run is 1.9× in 96 seconds. Both are hand-set and neither has been compared against a recorded tape. Break-even is 0.2744; at 0.25 the holdout loses 46%, at 0.4 it returns 25.65, at 0.55 it returns 1,554×. A figure that moves that far on one unmeasured share is a property of the assumption. The observer’s runner rate is not even in these units: read as the observer reads it, the same break-even is 0.2288, and its two-hour window is not the rule’s 150 seconds. So the number that decides anything is the next one.

So it is being measured, directly. A read-only observer samples 2% of new launches by hash, watches up to sixteen more coins it finds near the top of their curves, and keeps each for two hours. For every graduation it sees, the shipped rule is scored on the coin’s own 15-second candles — entry at the next open once the coin was first seen complete, the same take-profit, trail and 160-second time stop, fees on each leg, a position still open at the end counted as a total loss — and the mean return per graduation is compared with cash at /api/calibration. Verdict so far: not loaded. The runner share is measured too, but it is not in the generator’s units and it does not decide anything. The rule moves to paper trading only when its own measured return clears cash under the base costs with a positive stress mean, and never to live: nothing here can place an order. It establishes no live edge, and it names no coin.

The two folds are not holdouts. Holdout and confirm draw from the same generator with the same parameters, so a different seed varies sampling noise and nothing else — it cannot tell you the assumed launch mix is wrong, and the mix has never been fitted to anything. Two seeds of one generator is what the split is. The figures below bound sampling noise; they carry no out-of-sample evidence about pump.fun.

Read the cost table downward, not across: the sign sits inside the measurement error of the cost assumption, and a sign that the cost assumption decides is not an edge.

Suite count 17. Entered 1. Still open 1. Sum of closed-trade nets 0. Fees 0. Edge unproven. Mode shadow.