The endpoint that invents fraud.

Polymarket outcome shares are ERC-1155 tokens on Polygon. They can be bought on the order book, minted by splitting collateral, or simply sent from one wallet to another — and a plain token transfer carries no cost basis to the receiving side. So a wallet could, in principle, receive winning shares from a second wallet the same person controls, redeem them at $1.00, and book the entire face value as profit it never earned.

The standard way to test for this is a redemption reconciliation. For each market, work out the most a wallet could have redeemed given what it bought, and compare that to what it actually redeemed:

The share counts come from Polymarket’s public /trades endpoint. Run this across a set of high-profit wallets and the results are dramatic: across the three wallets in this report it flagged 4,725 markets carrying $4,772,218.83 of apparently unbacked redemptions. The arithmetic is sound. The inputs are not.

Take the single largest flagged market in the worst-looking wallet: 0x8c901f67…a904, in “Bitcoin Up or Down — April 27, 9:35PM–9:40PM ET” (0x2981bbd7…89bb). A market that existed for five minutes.

At resolution the wallet burned 12,759.355696 shares and was paid $12,759.355696 — a figure Polymarket’s own records and the chain agree on to six decimals.

So the filter saw $16.74 of buying against a $12,759.36 redemption and concluded that roughly $12,738 of winning shares had been gifted in from a sister wallet. What actually happened is that the wallet bought 12,759.68 shares for $9,978.88 across 80 fills and made an ordinary $2,780.48 profit. Adding one query parameter — takerOnly=false — returns 79 of the 80 fills the default view withholds.

The April and July /trades rows differ slightly (21.19 vs 20.87 shares). We publish both rather than reconcile them away: the April figure is what the filter actually consumed, and a reader re-running the call today will get the July one. The one-fill count — the part that matters — reproduces exactly.

An API disagreeing with itself does not settle who is right, so we went to Polygon. We took the redemption transaction, decoded which token was burned and in what quantity, then scanned the blocks before it for every inbound transfer of that exact token and resolved each sender.

Every share came through Polymarket’s own contracts. Not one arrived from another user’s wallet. Both sender addresses are verified under those names on Blockscout, and the two calls in Section 07 reproduce this.

There is a trap here that is the mirror image of the first one. An ordinary purchase also arrives in your wallet as an inbound ERC-1155 transfer — the exchange contract is the sender. A raw count of “inbound transfers” is therefore not evidence of anything either. The sender has to be resolved before the number means what it appears to mean.

One market could be a fluke. We re-fetched the 150 largest flagged markets by excess — the top 50 from each of the three wallets — and asked the same question of each: is the “excess” explained by buying that the default /trades view failed to report?

150 of 150. Not one of them showed a redemption exceeding the shares that wallet had verifiably bought and still held. The $926,462.58 the method called unbacked was $1,361,899.05 of real buying the default trade view did not show us.

Two limits on that sentence, both load-bearing. This is the top 150 of 4,725 flagged markets, chosen by size of excess, not a random sample — the right sample for testing whether the method works, the wrong one for estimating a rate. And what it establishes is narrow: these redemptions were paid for through the order book. It does not establish that the wallets are profitable overall — flagged markets are selected for having large redemptions, so losing markets are excluded by construction — and no reconciliation of this kind could detect two related wallets trading with each other on the order book, since those fills are paid for too. The claim is that the redemptions were bought, not that the wallets are innocent of everything.

That second point is what makes this dangerous rather than merely annoying. The omission is not proportional to how aggressively a wallet provides liquidity — in a market with no order book it is close to total, for every participant. A wallet that trades five-minute crypto markets will look almost entirely unfunded, and one that trades order-book markets will look clean, regardless of what either is doing.

Because the failure removes buying and never adds any, the error has a direction. excess does not measure transfers. It measures how much of a wallet’s trading the default view withheld, and it points one way: toward “this looks like fraud.”

The filter could have caught itself without any of this. On some flagged markets its own arithmetic produced negative net holdings — one records the wallet holding −577.29 shares of a side it then redeemed. A wallet cannot hold a negative quantity of a token. That is a self-refuting input, and nothing in the pipeline stopped on it.

These figures came out of our own private research between April and May 2026. None were ever published, and this page is the reason they will not be. Every one of them is withdrawn:

One caveat runs the other way and we are keeping it visible: we have shown the method produces false positives at scale, not that wallet-to-wallet share transfers never happen. The behaviour is possible and the platform cannot show it to you. We simply have no evidence of it at the scale the method attributed.

The gap the original theory pointed at is real. Polymarket’s /activity feed enumerates twelve event types — TRADE, SPLIT, MERGE, REDEEM, REWARD, CONVERSION, DEPOSIT, WITHDRAWAL, YIELD, MAKER_REBATE, TAKER_REBATE, REFERRAL_REWARD — and none of them is a token transfer. A wallet-to-wallet share transfer genuinely is invisible in the platform’s public data, and displayed profit does not net it out. That is worth knowing. It is just not what the filter was measuring.

The calls behind Sections 02 and 03, in full:

We are not above the failure we are describing. The first draft of this page reconciled 110 markets with a single unpaged request each — and silently truncated 18 of them at the same 500-record cap Section 07 now warns about, understating real buying by roughly $72,000 and overstating the wallets’ net by a similar amount. A pre-publication review caught it. The numbers above are the re-paged ones, and the check that would have caught it now runs on every build.

If you run wallet forensics on Polymarket — copy-trading products, leaderboard screens, academic work using displayed profit as a skill proxy — the specific failure is worth checking for in your own pipeline. It does not announce itself. It produces confident, specific, technical output that happens to be wrong, and in markets without an order book it is close to total.