Polymarket's fine print: machines settle 96.27% of markets, judgement settles 0.92%

On 23 July 2026, Vera Research parsed the resolution rule of every market Polymarket has listed, 584,210 contracts in a single deterministic pass, and found that 96.27% are decided by a machine reading a number off a scoreboard, a price feed or a published dataset ^[1][2]. Human judgement, defined as a rule naming no source and resting on a consensus of credible reporting, decides 5,396 markets, or 0.92% ^[3]. If that reading holds, the thing an operator should be auditing in this asset class is the data pipeline, not the referee.

The composition is not what the coverage would suggest. 60.31% of listed markets settle on the official result of a scheduled contest, a tennis match or a Counter-Strike map ^[4]. Another 30.41% settle on a price feed, either an on-chain oracle or a trading venue's own candle, which puts those two classes at 90.72% of everything ever listed ^[5][6]. A further 5.55% settle on a named published statistical series, four fifths of that class, 25,734 markets, being daily high and low temperature at named airport weather stations ^[7]. 582,754 contracts, 99.75%, carry rule text at all; 1,456 carry none ^[8].

Concentration is the part with consequences. A single on-chain oracle operator, Chainlink, settles 118,638 markets across seven price feeds, 20.31% of the catalogue and more than any other source ^[9]. That is roughly two thirds of the entire price-feed class ^[10]. Behind it: the International Tennis Federation at 76,519 markets, Binance at 50,112, the ATP Tour at 39,799, FIFA at 34,506, Major League Baseball at 33,430 and Weather Underground at 25,734 ^[11]. Those seven names together account for about 64.83% of all listed markets ^[12], and Weather Underground's count is identical to the airport-temperature subset, meaning one weather service holds that entire class ^[13].

Two caveats matter for anyone tempted to over-read this. The registry reports counts and shares of markets only, never a price, a return or a direction ^[14], so nothing in it says the judgement sliver is small in money terms; the wars, elections and leadership changes are exactly the markets a reader thinks of first, and by count they barely register at 21,709 markets, 3.72%, once every person-decided class is included ^[15][16]. The authors also say explicitly that nothing in the note claims any settlement source is wrong, late or contestable ^[17].

The method is unusually legible for research of this kind. The nine classes were built from the corpus rather than imposed, all 428 distinct hosts were assigned by hand, the classifier was frozen, and a fresh sample of 149 markets was adjudicated against each market's own rule text with no disagreement found, which supports a 95% lower bound of 97.5% rather than a claim of perfection ^[18][19]. Three earlier development samples covering 300 markets found seven faults, all fixed and listed ^[20]. One convention carries weight: 12 of the 149 depend on treating a third-party scoreboard such as HLTV or ESPNcricinfo as the official record, and a reader who rejects that outright scores the same sample at 91.95% ^[21]. A companion test on whether settlement type relates to price behaviour was withheld because most of the raw contrast proved to be an artefact of contract lifetime, with three of four comparisons flat inside matched lifetime buckets ^[22].

Watch whether anyone reruns the 149 shipped verdicts and lands somewhere else ^[21]. Watch for a volume-weighted version, because the count-weighted picture and the money-weighted picture almost certainly differ ^[14]. And watch the seven-feed concentration figure: a fifth of a catalogue reading from one operator is a structural fact regardless of that operator's record ^[9].