The Settlement Registry

The census

Nobody had read the fine print at scale. Nearly every contract on Polymarket carries a written resolution rule naming what settles it, and on 23 July 2026 we read the whole catalogue in one pass: 584,210 markets, every rule parsed by the same deterministic script. 582,754 of them, 99.75%, carry rule text. 1,456 carry none.

What the rules say is not what a reader would guess. 60.31% of listed markets settle on the official result of a scheduled contest: a tennis match, a World Cup game, a Counter-Strike map. A further 30.41% settle on a price feed, either an on-chain oracle or a trading venue’s own candle. Together that is 90.72% of everything Polymarket has ever listed. Add the 5.55% that settle on a named published statistical series, four fifths of it, 25,734 markets, daily high and low temperature at named airport weather stations, and 96.27% of the catalogue is decided by a machine reading a number off a scoreboard, a feed or a dataset.

Machines settle the catalogue. Human judgement is a rounding error.

The Settlement Registry · Vera Research

The markets a machine cannot settle

That leaves a sliver. Human judgement, a rule that names no source but rests on a consensus of credible reporting, decides 5,396 markets, 0.92% of the catalogue. Widen it to every class where a person rather than a dataset makes the call, an announcement, a recording, a newsroom’s own verdict, a market naming no source, and it is 21,709 markets, 3.72%.

These are the markets a reader thinks of first: the wars, the elections, the leadership changes, whether a company will file. Set against the scoreboards and the price feeds, by count they barely register.

Who does the deciding

The catalogue is not settled by many hands evenly. A single on-chain oracle operator, Chainlink, decides 118,638 markets across seven price feeds, 20.31% of everything listed and more than any other source. Behind it are sports authorities reading their own scoreboards, one exchange reading its own candles and one weather service reading airport thermometers: the International Tennis Federation settles 76,519 markets, Binance 50,112, the ATP Tour 39,799, FIFA 34,506, Major League Baseball 33,430 and Weather Underground 25,734. A handful of publishers decide most of what Polymarket lists.

How the reading was checked

The nine classes were built from the corpus rather than imposed on it, and all 428 distinct hosts in the table were assigned by hand. The classifier was then frozen and a fresh sample of 149 markets was drawn and adjudicated against each market’s own rule text. All 149 agreed with the label already assigned. On a sample that size the honest statement is not that the reading is perfect but that no disagreement was found, with a 95% lower bound of 97.5%. Three earlier development samples, 300 markets in all, found seven faults; every one was fixed and every one is listed in the source analysis rather than quietly patched.

One judgement call carries real weight. 12 of the 149 rest on a convention that a third-party scoreboard, HLTV or ESPNcricinfo say, counts as the official record of a scheduled contest. A reader who rejects that convention outright would score the same sample at 91.95%. Both readings are reported, and the 149 verdicts ship with the analysis so the adjudication can be redone by anyone who disagrees.

What we are not claiming

A companion test asked whether settlement type relates to how a market’s price behaves. It ran, and it is deliberately absent here. Most of the raw contrast turned out to be an artefact of how long the contracts live: inside matched lifetime buckets, three of four were flat. A census does not need it. Nothing in this note says a market is mispriced, or that any settlement source is wrong, late or contestable. The registry is a snapshot with a date on it, because the resolution text is written by the venue and the venue can change it.

The takeaway

1. We read the written resolution rule of every market Polymarket had listed on 23 July 2026, all 584,210. 582,754 carry rule text; 1,456 carry none.
2. Counted by markets the catalogue is machine-settled: 60.31% on the official result of a scheduled contest and 30.41% on a price feed or on-chain oracle, 90.72% together. Add the 5.55% on a named published statistical series, 25,734 of those 32,396 airport weather stations, and 96.27% is read off a scoreboard, a feed or a dataset.
3. Human judgement, a rule naming no source but a consensus of credible reporting, decides 5,396 markets, 0.92%. Every class a person rather than a dataset must decide is 21,709 markets together, 3.72%.
4. Settlement authority is concentrated: one on-chain oracle operator’s price feeds decide 118,638 markets on their own, 20.31%, more than any single sports federation or exchange.
5. The reading was checked on a held-out sample of 149 markets drawn after the classifier was frozen; all 149 agreed with their assigned class, a 95% lower bound of 97.5%. Reject the third-party-scoreboard convention and the same sample scores 91.95%.

A census of what the rules say, complete over the catalogue as of 23 July 2026. Counts and shares of markets only. Nothing here is a price, a return, a direction or a forecast, and nothing here says any market is mispriced or any source is wrong.

Method and data

Built from the committed bundle at assets/settlement-registry: the 584,210-row registry, results.json, and the 149 held-out verdicts. A census of the market catalogue as of 23 July 2026, classed by what each market’s own written resolution rule says settles it. Counts and shares of markets only. Every figure re-verified against results.json on 24 July 2026.

Odds shown are each market’s own price at the time Vera flagged the story, stated as fact, not advice. Nothing here is a recommendation to buy, sell, or trade any market or asset. Vera and Crypto Briefing are not registered investment advisers.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.