What breaks resolution is the moment, not the fact

Updated 2 September 2026: two cases added — Polymarket's move to an averaged settlement price on 7 August, and the outcome of the dispute over Strategy's bitcoin sale.

Arguments about prediction markets usually turn on accuracy: did the market call it or not? But money changes hands not at the moment of the event but at resolution — when somebody declares the outcome and that declaration becomes final. Two cases from 2026 show the same thing: what breaks is not the definition of the fact but the definition of the moment.

How resolution works on Polymarket

The mechanism is public. When a market's end conditions are met, the contract queries the UMA oracle, and three steps follow: a proposal — one answer, backed by a 750 USDC bond — a two-hour challenge window, and, if anyone disputes it, a vote of UMA token holders whose ruling is final. Undisputed markets settle two to four hours after the event; disputed ones can drag on for days.

One figure for balance: since 2021, across all markets, about 99% of assertions have gone undisputed. A dispute is the tail of the distribution. But the tail is where the biggest markets sit, and 2026 has already produced more than 1,150 disputed markets — more than the whole of 2025.

In November 2025 Polymarket changed the rules: only 37 pre-approved addresses may now propose a resolution (Risk Labs staff and users with a strong accuracy record), while disputing stays open to anyone. That dealt with a flood of junk proposals and deliberately left the final vote unchanged.

Case one: the moment was for sale, seconds at a time

Researchers at Stanford and Singapore Management University (SMU) examined roughly two months of Polymarket's five-minute bitcoin contracts. They found 821 accounts that took $8.2 million in aggregate profit in settlement windows the authors classify as likely to have been manipulated.

The mechanism is embarrassingly simple. The contract settled on an instantaneous price snapshot — the value at a single point in time. A trader would build a position on Polymarket, then place large orders on Binance just before settlement, pushing the price past the threshold for exactly as long as the snapshot needed to catch it. Then the price would snap back to where it had been. The manipulation leaves a measurable trace in order flow: in the final ten seconds before settlement, net flow on Binance jumped by around 50%. About 93% of the losses in the flagged windows were borne by retail traders. The press named the trick after the five seconds or so the price had to be held.

On 7 August 2026 Polymarket replaced the single snapshot with a time-weighted average price, computed over a 30-second window for five-minute markets and a 60-second window for 15-minute and four-hour markets.

Note what was actually fixed. Not the price source — that was an exchange all along. What was fixed was the width of the moment: a point that could be pushed in seconds was replaced by an interval that has to be held longer, and paid for accordingly.

Case two: the fact was in a regulatory filing, and it did not help

Polymarket runs a series of monthly contracts on one question: will MicroStrategy — the company now trades as Strategy — sell any bitcoin by the end of a given month? The May contract in that series saw more than $60 million in volume.

The facts are not in dispute: Strategy sold 32 BTC between 26 and 31 May at an average price of $77,135 per coin — its first sale of bitcoin since 2022. The trade closed before the contract's midnight cutoff. But the 8-K filed with the regulator to disclose it went out on 1 June.

The question did not say which of those counted as the moment: the trade itself, or its public disclosure. Two proposed "No" resolutions were disputed; the disputes went to a vote of UMA holders, and that vote closed the May contract as No and the June contract in the same series as Yes. The reading that prevailed was that the disclosure date, not the trade date, was what counted. Traders who bought "Yes" on the May contract lost, even though the sale itself happened in May.

Case two exposes a second failure, and it is not about wording at all: it is about who votes. A Wall Street Journal investigation published in May 2026 reports three figures. In most disputed markets, more than half the UMA votes came from the ten largest wallets; at least 60% of regular voters could be linked to active Polymarket accounts; and in roughly one dispute in five there was at least one voter with a financial stake in the very contract they were voting on.

That is not an accusation of fraud — the rules of the vote do not bar interested parties. It is a description of the design: the court of last resort on a contested question is made up of people holding money on the answer.

What the two cases have in common

In both, the fact was not in doubt. The price of bitcoin at settlement is known; the sale of 32 BTC is confirmed by a regulatory filing. What diverged was the definition of the moment. In the first case the price of that was 93% of the losses in the flagged windows landing on retail traders; in the second, the outcome of a $60 million market.

The practical lesson for anyone drafting market terms is this: the source of truth is only half of resolution. The other half is the moment, and it has to be pinned to the same document as the fact.

Three questions test any wording:

1. What exactly confirms the outcome — which document, which endpoint, which publication? "Official sources" is not an answer.
2. Which clock defines the moment — the time of the event, or the time it is reported? If those two can diverge by even a minute, the question must say which one governs.
3. How narrow is the moment, and what does it cost to push it? A point is almost always cheaper to attack than an interval. If the outcome is taken from an instantaneous value, treat the cost of attacking it as part of the design rather than as somebody else's problem.

None of the three is about the oracle. All three are about the text of the question, written before the market opens, and it cannot be changed once it has. That is exactly the work we do on prediction-market projects: terms that leave a dispute nothing to resolve, because one source, one clock and one width have all been named.

We have no data of our own in either case — everything above comes from published work, and we traced each claim back to a primary source: the Stanford and SMU study, the Wall Street Journal investigation, Strategy's regulatory filing, and the public documentation of the UMA oracle.