How the idea was checked
Not “we think it will work,” but exactly what we checked, how we checked it, and what isn’t solved yet. The short version fits in three lines; the details are below.
What you can check yourself
You don’t have to take the main points on faith. Test the defense against cheaters in the demo lot. It runs on the mechanism’s real rules: you pick a way to cheat, move the sliders, and see two numbers—what the cheating earned and what the honest path would have cost. The sources below are named so you can look them up and read them.
You can go further and rerun the calculations yourself: all the programs we used to work out the mechanism are in one archive. All they need is Python; the instructions inside (in English and Russian) tell you what to run. For example, every number in the Lamborghini example on the home page comes from a single command.
How the work went
Who checked the sources, and how
We tried to refute every claim using the sources, not to confirm it. To be clear about who did the checking: three independent AI reviewers. Not human experts, but separate AI agents in our research pipeline, whose job was to find where each claim breaks. A claim was accepted only if it couldn’t be refuted.
We didn’t throw away what was refuted. In all, 20 claims didn’t hold up, including some nice stories we ourselves wanted to tell about the project. The main ones are taken apart in the “Idea review,” in the section “We killed our own legends.”
What the sources confirmed
In 2025, researchers proved that a mechanism for selling a single lot can have at most two of three properties. The first is simplicity: one action, no rounds of bidding. The second is “telling the truth pays best.” The third is protection against planted bids: the seller gains nothing by bidding through stand-ins or other wallets.
We deliberately chose simplicity and protection against planted bids, and gave up the second. That’s why we don’t catch cheaters with rules; we make cheating cost more than it pays. To be precise: fake sales are fully closed, while planted stakes “for” yourself remain a limited gap (see row 2 of the table below).
Komo, Kominers, Roughgarden, arXiv:2404.00475 (EC’25 conference).
Horse racing answered this long ago. A horse’s owners may bet FOR it but never AGAINST it. And the jockey, the one person who decides how the race goes, may not bet at all, in either direction, including through stand-ins. In our mechanism, the seller is the “jockey”: only the seller decides whether the sale happens.
But what we borrow from horse racing is the model, not the ban: who bears the most responsibility for the outcome, and how to define a workaround through stand-ins. We have no role-based bans, because on a blockchain you can’t tell whose wallet is whose. Instead of a ban, there’s a price: every move the seller makes “against” themselves, or “for” themselves through other people’s wallets, is worked out in the table below. Usually it costs more than the honest path, and where a gap remains, we say so.
Britain’s Jockey Club; Australian Rules of Racing, rules AR 115 and AR 237(7).
Yes, as long as you don’t lie. Some people profit when a company’s stock falls, and they criticize those companies publicly. In 2024, the US market regulator charged one of the best-known of these critics with a $20 million fraud—not for the criticism itself, but for lying about his own position and secretly funding the campaign.
So for a doubter, the line isn’t the campaigning itself (that’s legal and useful: it brings the truth about the lot to light); it’s deception.
SEC, the Andrew Left / Citron Capital case, press release 2024-89.
US courts identify a pyramid scheme by one key sign: the reward is paid for recruiting new participants, not for selling a product to a real buyer. Our supporter reward is paid only out of the money from a real sale to an outside buyer—never for bringing in new supporters.
The Koscot test: Webster v. Omnitrition (9th Circuit, 1996); FTC v. BurnLounge (2014).
What the calculation showed
A simulator reproduces every money rule of the mechanism exactly. The conditions are the worst case for us: transactions are free (so the whole cost of an attack sits in the mechanism itself, not in blockchain fees), and the cheater sees every stake and every pot, just as a real attacker would on a blockchain. We ran seven experiments in total.
What “the balance matches to the 17th decimal place” means: in every run, the total money held by all participants before and after matches to the 17th decimal place, so the arithmetic neither loses nor creates a single cent. What it does not prove is that the mechanism can’t be cheated. Resistance to cheating is shown by separate attack runs, whose results are in the table below.
Nine ways to cheat: eight calculated, one still open. How to read the table:
| Attack and status | The cheater’s result versus the honest path |
|---|---|
| 1. Fake sale Buy from yourself through a second wallet to get the deposit back. closed by the math | Costs 2.9–4.7% of the price more than giving up, even on the day that suits the cheater best. The cheater pays the platform fee, the part of what dripped that isn’t refunded, and the supporter reward, which is always paid |
| 2. Planted supporters (shills) Back yourself from other wallets to win the reward back. residual risk, capped | The gain is capped at a few percent of the price (on an NFT lot in the demo, no more than about 3.6%) because planted stakes pay the drop and the fee, just like real ones. The fix: check whether the money came from the seller’s wallets, plus reputation |
| 3. Fake sale + planted supporters Both tricks at once. closed by the math | Costs 0.8–2.1% of the price more than giving up. To get there, we reduced how much of what dripped is refunded on a sale (to 55% for expensive items and real estate) |
| 4. List-and-cancel Keep listing and pulling lots while betting against yourself, pocketing the drops of honest supporters you lured in. residual risk | With no one lured in, the cheater pays 9 times the harm they do. If real supporters join the lot, the cheater makes about 0.4–1% of the price per cycle. The fix: a public history of cancellations, plus a small fee for listing a lot (designed but not switched on yet) |
| 5. Self-buy ladder Split a fake sale into many small resales. closed by the math | Costs 9.0–14.7% of the price more than giving up (demo lot): each piece carries its share of all the costs of a fake |
| 6. Paid panic A doubter whips up fear at the right moment so the lot doesn’t sell. residual risk | A doubter’s earnings aren’t theirs until the deadline, and the pot is fed by supporters’ drops: scare everyone off, and you cut your own income. The exception: panic pays off only in the last few days before the deadline, and even then it earns pennies. Campaigning “against” is legal; you cross the line by lying about your own position or secretly funding someone else’s campaign |
| 7. Sniper Jump in as a supporter at the last moment to grab a share of the reward without risk. closed | A late stake carries almost no weight: the reward is split by how much you put in × how long you held it, and every stake has a minimum holding time |
| 8. Betting on your own failure Secretly bet against yourself and let the lot fail. residual risk | A loss of 14–15% of the price versus 0.75% for giving up (demo lot): the fat pot of a doomed lot attracts real doubters, and their money dilutes the cheater’s share. It’s closed only if enough outside doubters join the lot; in an empty market, it stays open |
| 9. Drop outsourcing A service pays the drop for supporters and splits the reward with them. open | Not calculated. Nobody in the industry has a ready-made fix |
Detailed answers with numbers are in the “Cheaters” section of the FAQ, and you can check them yourself in the demo lot.
We chose specific settings for four kinds of items, and each set passed six checks (including combined attacks). An honest seller’s costs are close to the usual middleman fees, toward the upper end: 7.9% for cars versus 5–10% for selling through a dealer; 5.3% for real estate versus an agent’s 3–6%; 10.9% for tickets versus about 10%. An NFT launch costs 14.9%, versus 10–15% for a marketplace plus marketing—right at the top of the range, and for NFTs that’s a lot.
What’s still open
The research is finished. Two questions remain open.
The drop is our entry filter: a supporter proves they believe by risking a small amount. The attack: a service covers the drop (“I’ll pay your drop, and we’ll split the reward”), and the belief check turns into an ordinary paid service. This has already happened in other crypto projects: middlemen bought up voting power that was handed out based on how much you put in × how long you held it.
We did a separate search for how the industry handles this: nobody has a ready-made fix. That isn’t a gap in our checking but a genuine negative result. It’s an open problem, and we’ll work on it by running the numbers.
The case of buying up voting power: Curve/Convex.
Our calculation shows that with five or more supporters, when the reward is split only by “how much you put in × how long you held it,” each of them is individually better off not trying and waiting for someone else to bring in the buyer. So the sale is driven by the seller and a few big supporters.
The fix is a separate payment to whoever brings in the buyer (giving them a third of the reward restores 78–94% of the ideal effort). It has been designed but isn’t switched on yet: it changes the math of cheating, so we’ll run those numbers first. More in “The honest truth about supporters” on the home page.
Beyond these two questions, the first version has some limits. For example, it works only with NFTs and tokens (for why, see “Who confirms the sale”). The full list is in the FAQ under “What have you NOT proven yet?”