What was wrong
The audit scored 100 named Polymarket accounts on forecasting calibration and published a per-account CSV: win rates, Brier scores, realized profit and loss, entry prices, event concentration. Each account appeared under the handle its owner trades as.
Those per-account figures were computed from a trade history capped at 2,000 fills per wallet. 85 of the 100 published rows sat exactly at that cap, which means the cap bound and their histories were cut short. The article stated the cap once, in its methodology. The CSV did not carry it at all, and a CSV separates from the page it came from the moment someone downloads it. So the most portable form of the work was also the form that could not tell you it was a window.
On 9 August 2026 we walked one of those wallets in full on Polygon and reconciled it against the chain. The published row credited it with 8 resolved positions. The complete record holds 4,954. Every rate we printed for that account was computed on roughly one sixth of one percent of its trading, and nothing on the page or in the file told the reader so.
Why we withdrew it instead of annotating it
We first added a per-row truncation flag and a correction note. That was the smaller move and it was not enough. The problem is not that a number lacked a caveat; it is that we attached performance statistics to identifiable people on evidence that could not support a statement about any of them individually. A caveat does not repair that, it only documents it. The honest action is to stop publishing it.
The dataset has been removed and is no longer served. The per-account findings are withdrawn in full and should not be cited. If you quoted this work, we would rather you knew.
What we kept
The withdrawal notice, at the address the claim lived at, and the integrity log entry recording that the file existed and its hash. Removing those would make the record cleaner than the truth.
What we do now instead
We read a wallet's complete position history from the chain and reconcile it against the chain before we say anything about that wallet, and we publish the block the read was reconciled at alongside the count of positions it covers. When we cannot prove the record is complete, we say the read is a window and we do not print a rate. That is the difference between the retired work and what replaced it, and it is the whole reason this page exists rather than a deleted URL.