How an honest venue settles a question
A prediction market makes exactly one promise: when the question resolves, the contract settles at 100 for the side that was right. Everything else on the screen — the book, the chart, the probability — is arranged around that promise.
Which means the entire product rests on a detail most people never read: who decides what actually happened, by what rule, written down when.
The honest answer fits in one sentence: an honest venue names the settlement source before anyone trades, publishes dispute rules written before any dispute exists, states in advance what voids the market — and keeps the full settled record public, graded under the rules as they stood. The rest of this piece is why each of those clauses is load-bearing.
The promise
A market contract is a claim about a future fact, priced between 0 and 100 until the fact arrives. How that price gets made is its own subject, covered in Where a market price comes from.
Settlement is the other half, and it is the half that makes the first half mean anything. A price is a forecast; settlement is the grading. The published accuracy record — decades of research on whether these prices are any good — exists only because settled questions can be graded at all (Are prediction markets accurate?).
The settlement failures the literature documents share one parent assumption.
The parent assumption
The assumption: the answer will be obvious. It rarely is. Questions are written by people picturing the clean outcome, and reality is under no obligation to arrive in the shape the question assumed.
The forecasting literature has carried this warning for two decades: a market contract must be specified so that it is clear when it settles Yes and when it does not (Snowberg, Wolfers and Zitzewitz 2013). The same survey’s example is instructive — a 2003 market on whether there would be a United Nations resolution “on” Iraq. It was the better-specified of the two contracts the survey compares — and still open to argument, because what exactly does “on” mean?
Or take a case from the same literature (Wolfers and Zitzewitz 2006): a market on whether a political leader would “depart” his territory by a year-end deadline. When he fell gravely ill, the book discovered the question had never decided whether being flown abroad for medical treatment counted as departing — or whether dying did.
These read like trivia until you hold a position on the wrong side of the comma. Events land sideways — partially, ambiguously, at 11:58 on deadline day — and a question is only as good as its behavior when they do. Those cases are decided by the settlement spec, or by nobody.
What clean settlement looks like
In serious sport, the referee is named before kickoff. Nobody would accept the reverse — play first, then decide who officiates, then write the rules they will apply. An unspecified market asks its traders to accept all three.
Clean settlement is not a talent, and it is not a promise of good judgment. It is a short set of documents, all of them dated before the first trade.
The source is named in advance. Not “official sources.” Not “the venue’s discretion.” A named authority whose publication settles the question — and, because named sources can go silent, a named fallback and a written procedure for when they do.
The dispute rules exist before the dispute. A window in which a settlement can be challenged, stated in hours. Who may challenge, who decides, and on what record. Written while nobody is angry, because rules drafted mid-controversy are positions, not rules.
Void conditions are stated plainly. Some questions break instead of resolving — the event is canceled, the institution in the question disappears, the source stops publishing. What happens then is written in advance, and it reads as the dullest clause in the spec — right up until it is the only clause that matters.
No wording captures every event perfectly — the comma cases prove it — which is why the dispute window and the void conditions exist, and why they are written before they are needed. None of it requires trust. A settlement spec is checkable before you commit anything; a reassurance is not.
The category is still learning this
Settlement is not a solved problem the industry simply forgot to write down. It is live terrain, and the recent research shows both how central it is and how far from settled it remains.
A 2026 study coined a term for the property that decides whether a question can become a market at all: settlement legibility — “the degree to which an uncertainty can be worded, sourced, and credibly resolved by third parties” (Adegbenro 2026, preprint). Across roughly 6,000 contracts on two under-covered regions, what venues listed tracked legibility rather than public importance — salient civic questions produced almost no inventory while highly legible ones dominated it. The questions a venue can offer are downstream of the questions it can grade.
Wording is priced, too. A study of more than 100,000 events across ten venues found economically identical questions — worded, sourced and resolved differently — trading at persistently different prices, with deviations of 2–4% that do not close (Gebele and Matthes 2026, preprint). Traders treat two versions of “the same” question as different assets, because at settlement time, they are.
And disputes are common enough to be drawing research of their own. Where contested outcomes escalate to a public vote of the resolution system’s participants, researchers have begun testing whether such disputes can be arbitrated consistently — including a 2026 study asking whether language models could reproduce those resolutions (Wen, Zhou and Huang 2026, preprint). Whatever one thinks of any particular design, the scholarship itself is the signal: how a question gets graded is now a first-order subject.
Why integrity cannot be retrofitted
There are only two settlement regimes: the rule as written, or somebody’s judgment after the fact. The first occasionally produces answers that sting — the comma cases above were real, and people were on the wrong side of them. The second produces answers that are wrong in a worse way: quietly, unverifiably, and always in a direction someone chose.
A venue that improvises a kinder answer mid-dispute has not fixed one market. It has repriced all of them. Every open question on that book now carries a term nobody can quantify: the chance that this rule, too, is negotiable.
And the damage runs backward as well as forward. A settlement record is only evidence if the rules were fixed while it was being written. Move your own goalposts and every result on the board becomes a claim, not a fact — including the honest ones.
That is why integrity cannot be added later. A venue can upgrade its sources, tighten its windows, publish more. All good, all forward-looking. What it cannot do is reach back and make yesterday’s settlements principled. A record is only clean if it was clean from day one.
The four documents
The test fits in four documents, each of which either exists in writing or does not.
The settlement source, named before trading opened. The dispute window, stated in hours, with a named decider. The void conditions, in plain language. And the sharpest one: the full settlement history, public — including the contested cases, resolved under the rules as they stood.
A venue that can produce all four can fairly be graded on its record. A venue that cannot is asking you to trade first and meet the referee later.
Don’t. Read the settlement spec before you read the chart. The referee is named before kickoff — or it is not a market yet, just an argument scheduled for later.
Sources
- Snowberg, E., Wolfers, J., Zitzewitz, E. (2013). “Prediction Markets for Economic Forecasting.” Handbook of Economic Forecasting, Vol. 2, Elsevier. Also NBER Working Paper 18222. nber.org
- Wolfers, J., Zitzewitz, E. (2006). “Five Open Questions About Prediction Markets.” NBER Working Paper 12060. nber.org
- Adegbenro, A. (2026). “What Prediction Markets Can See: Market Formation, Settlement Legibility, and the Geography of Tradable Uncertainty in Africa and Latin America.” Preprint, arXiv:2606.17503. arxiv.org
- Gebele, J., Matthes, F. (2026). “Semantic Non-Fungibility and Violations of the Law of One Price in Prediction Markets.” Preprint, arXiv:2601.01706. arxiv.org
- Wen, J., Zhou, J., Huang, J. (2026). “Can LLMs Help Decentralized Dispute Arbitration? A Case Study of UMA-Resolved Markets on Polymarket.” Preprint, arXiv:2604.15674. arxiv.org