Who decides what's true?
A market that pays out on real events is only as trustworthy as whoever rules on the outcome. Settlement is where prediction markets quietly live or die — from the two ways to decide the truth, to the oracle that's the most-attacked layer in crypto, to the tiered architecture it takes to settle a million markets.
It's tempting to treat the price as the interesting part of a prediction market — the probability, the calibration, the liquidity that keeps it honest. But none of that matters if the last step goes wrong. A prediction market is a promise: when this event happens, the YES holders get paid and the NO holders get nothing. Someone has to look at the messy real world, decide which of those two things occurred, and flip the switch. That step has a name — resolution — and it is the quietest, most underrated risk in the whole machine.
Here's the thing nobody tells newcomers: you can be completely right about the world and still lose. If the market asks a sloppy question, or the resolver rules against the plain meaning of events, your correct forecast pays you nothing. The price was perfect; the settlement burned you. So before you trust a number, you have to ask who turns it into a payout — and whether you trust them.
From a messy event to a clean payout
Every event contract is a binary. It resolves to \(1\) or \(0\) — there is no 73¢ at the end, only paid or not paid. The job of resolution is to take something genuinely ambiguous — an election called over a long night, a game decided in overtime, an economic figure that gets revised a week later — and collapse it into one of two states. The whole design problem is that real events don't arrive pre-labelled. Somebody, or some mechanism, has to do the labelling, and the entire credibility of the venue rests on that step being both fast and fair.
There are two broadly different answers to "who labels it," and they sit at opposite ends of a trust spectrum. One puts a regulated company in charge. The other tries to remove the company entirely and replace it with an open, adversarial process. Both work most of the time. Both have a failure mode. Understanding the trade-off is the point of this piece.
Model one — the centralized resolver
The simplest answer is: the exchange decides. On a regulated venue like Kalshi, every market ships with a rulebook before a single contract trades. The rules spell out the exact source of truth, the exact threshold, and the exact timing — not "did the candidate win" but "as reported by the Associated Press as of this date and time."1 When the moment comes, the exchange reads its own rule against the named source and settles. You get your payout fast, usually within the day, and the logic is legible because it was written down in advance.
The cost is straightforward: you are trusting the venue. The exchange wrote the rule, the exchange interprets the rule, and the exchange holds the money. In a regulated market that trust is backstopped — the venue answers to the CFTC, keeps customer funds segregated, and faces real consequences for ruling in bad faith.1 That oversight is exactly why a license is worth so much: it converts "trust us" into "trust the regulator standing behind us." But make no mistake about the shape of it — centralized resolution is fast and clean precisely because one accountable party gets the final say.
Model two — the optimistic oracle
The other answer comes from crypto, where the entire premise is to avoid a single trusted company. Polymarket settles most of its markets through UMA's optimistic oracle, and the word "optimistic" is doing real work.2 The mechanism assumes the easy path will usually be correct, and only spins up the expensive machinery when someone objects.
It runs in stages. After the event, someone — anyone — proposes the outcome and posts a bond to back their claim. That opens a dispute window: a fixed stretch of time during which anyone who thinks the proposal is wrong can challenge it by posting a bond of their own. If nobody challenges, the proposal stands and the market settles — cheap, fast, no committee. If someone does challenge, the question escalates to a vote: holders of UMA's token adjudicate, and the side that the vote agrees with takes the loser's bond.3 The bonds are the whole trick. Proposing a lie or challenging the truth costs you money, so the cheapest move is usually to be honest.
It's a genuinely elegant design. It needs no standing committee, no privileged data vendor, and no permission to participate — anyone with a bond can defend the truth. And for the overwhelming majority of markets, where the outcome is obvious, it settles on the cheap top path and never convenes a vote at all.
The two models, side by side
Neither model is "the right one." They're different answers to the same question — who do you trust, and what do you trust them with — and they make opposite bets.
Notice that the two failure modes rhyme. The centralized venue can rule against you, and you have to trust it not to. The oracle can escalate to a vote that reads an edge case the way you didn't — and a vote of bonded strangers is not obviously wiser than a regulator about, say, the precise wording of a contract. Different machinery; the same underlying problem keeps showing up. And it isn't the obvious cases.
The hard part is always ambiguity
Here is the part that surprises people: resolution almost never breaks on hard facts. When the question is "who won the game," both models settle in seconds, because there is a scoreboard and everyone agrees what it says. Disputes don't come from uncertainty about the world. They come from uncertainty about the question.
The trouble is language. A market asks something that felt perfectly clear when it was written, and then reality serves up a case the wording never anticipated. What counts as "by the end of the year" — announced, signed, or in effect? Does a thing that is technically true but obviously not in the spirit of the question count as YES? When two reputable sources report different numbers, which one is the source of truth? None of these are questions about facts. They're questions about interpretation, and that is exactly where money gets stuck.
Across 2024 and 2025, as volumes climbed from under \(\$100\text{M}\) a month to past \(\$13\text{B}\),4 the contested resolutions that drew real attention followed this pattern almost every time. Rarely did anyone dispute what happened. They disputed whether the literal text of the market matched the spirit a normal person would have read into it — vague criteria, an unforeseen technicality, a "technically true" reading that split a community down the middle.5 The lesson the operators took away is the one that matters for everyone else: the wording of the question is not paperwork. It is the product.
The deliverable of a prediction market is two things at once: a question crisp enough that a stranger could rule on it without arguing, and a resolver — venue or oracle — neutral enough that you'd accept its ruling even when it goes against you. Get either wrong and the price was never the point.
So hold two questions apart that feel like one. The first: is my forecast right? The second: if I'm right, will this market actually pay me? Those are different risks, and you can nail the first while losing to the second. Calibration and sharpness are about being right; resolution risk is about whether being right gets rewarded, and it lives entirely in the wording of the question and the credibility of whoever rules on it. The honest version of "the price is a probability" comes with an asterisk — a market can price an outcome perfectly and still burn you if the resolution is ambiguous. The number tells you what the crowd believes; the resolution tells you whether that belief ever turns into money. Both have to be right, and only one of them shows up on the screen.
That's the shape of the risk. The rest of this piece is about the machinery underneath it — because the two failure modes we just named only get harder the moment you look closer. The centralized resolver's accountability is easy enough to reason about: a regulated company answers to a regulator. The decentralized side is where the engineering lives, and it deserves its own dissection. So let's follow the optimistic oracle down past the happy-path diagram and ask the questions that actually decide whether a market settles on the truth: what is this thing, really; who can attack it; and what happens when you try to run it not for a few thousand markets but for a million.
The oracle problem
Start with where the abstraction leaks. On a regulated venue the resolver is a named company. But a huge share of today's volume settles somewhere stranger — on a blockchain, inside a smart contract that cannot see the outside world at all. A blockchain is a closed, deterministic world: every node must re-run every transaction and reach the same answer, so the chain is, by design, blind to anything outside itself. Reality — an election, a scoreline, a price — is none of those things. So how does code that lives on-chain ever learn that a candidate conceded, or that a game went to overtime? It can't, not on its own. It has to be told. The thing that tells it is the oracle, and it is the most underappreciated piece of plumbing in the entire stack.
This is the crux of it. A smart contract is just code that runs on every machine in the network and must produce an identical result everywhere — that determinism is what makes it trustworthy.6 The price of that guarantee is total isolation: the contract has no way to call out to a news site, read a sensor, or check a scoreboard. If it tried, two nodes might fetch two different answers and the network would fall out of agreement. So the chain is sealed, and anything from the outside — the very real-world data a prediction market exists to settle on — has to be carried in by something that bridges the gap. That carrier is the oracle, and the gap it spans is the oracle problem: how do you deliver outside truth to a system built to trust nothing it can't recompute?
For a prediction market the stakes are blunt. The oracle is not a nice-to-have feeding a chart — it is the component that decides who gets paid. Every YES and NO holder is waiting on one bit of information from the outside, and the oracle is what writes that bit on-chain. It is, quite literally, the part of the machine that turns a fact about the world into money. Which is why the design of the bridge matters as much as the design of the market.
Three ways to build the bridge
There is no single oracle design, because "report the truth on-chain" means different things for a basketball score than for the price of an asset. Three shapes cover most of what's deployed today, and they differ mostly in who they trust and how they handle disagreement.
The optimistic oracle
The first is the one most prediction markets lean on — and it's the one we already pulled apart above: UMA's optimistic oracle, where anyone proposes an outcome and posts a bond, a dispute window opens, and a contested call escalates to a token-holder vote, the bond making honesty the cheap move.3 What's worth adding here is where it sits on the trust axis: permissionless and adversarial, no standing committee, no privileged vendor. Hold that position — the next design is built on the opposite instinct.
The aggregated data feed
The second shape solves a different problem: continuous, machine-readable numbers — above all, asset prices — that a contract needs constantly and can't wait out a dispute window for. Here the dominant design is the aggregated data oracle, of which Chainlink is the best-known.7 Rather than trust one source, it takes readings from many independent reporters, each pulling from multiple venues, and combines them — typically around the median, so no single bad feed and no one outlier exchange can move the published number. The aggregate is what lands on-chain. This is the workhorse behind most of DeFi: lending, derivatives, and stablecoins all depend on a price oracle telling the contract what an asset is worth right now. The design philosophy is the opposite of optimism — don't wait for a challenge, drown out any single liar with redundancy.
The trust spectrum
Underneath both sits one axis, and every oracle is a point on it. At one end, a single trusted reporter: one party, or the venue itself, simply states the outcome. It's fast, cheap, and dead simple — and it's centralized, so you're back to trusting one entity, who could be wrong, coerced, or could censor a result. At the other end, a decentralized vote: no single party can be leaned on, which buys you censorship-resistance, but it's slower and it introduces a new failure mode — if the vote is weighted by tokens, whoever holds enough tokens can decide the answer. Aggregation lives in the sensible middle: many reporters, no vote, no single point of failure, but still a fixed set of feeders you have to trust to be honest and independent. There is no free lunch on this axis; every design buys one property by spending another.
The most attacked layer in crypto
Here's why the oracle is the soft underbelly. Because it's the one place a closed system reaches out to an open one, it's the natural target — and in DeFi, oracle manipulation has been one of the leading causes of exploits for years.8 The pattern is grimly consistent: an attacker doesn't break the cryptography or find a bug in the math. They corrupt the input. Push a thin market's price the way you want for a moment, get the oracle to report that distorted number, and a lending or derivatives contract — doing exactly what it was told — hands over funds against a price that was never real. The contract did nothing wrong. The truth it was fed was a lie.
A prediction market faces the same shape of risk, just aimed at the resolution vote instead of a price feed. If settlement escalates to a token-holder vote, then whoever can assemble enough voting weight can, in principle, mis-settle the market — force a NO outcome to pay YES — and walk away with the pot. This is the governance-capture or whale-vote risk, and it isn't hypothetical hand-wringing; it's the structural worry baked into any decentralized resolver.9 The defense isn't a clever line of code. It's economics. The whole system is safe only when one inequality holds:
That's the entire security model in one line. An optimistic oracle makes the left side big by forcing an attacker to out-bond honest disputers and, on escalation, to overwhelm a token vote — both expensive, by design. An aggregated feed makes corruption costly by forcing an attacker to move many independent venues at once instead of one. The market is secure precisely when breaking the oracle costs more than the prize for breaking it — and dangerous the instant a single market's payout grows large enough to outrun the cost of capturing the thing that settles it.
Why it's the truth layer
Step back and the oracle stops looking like a detail and starts looking like the point. It is the single place where a self-contained cryptographic world touches reality — the layer where "what is true" gets written down and made binding on money. Everything upstream, the matching engine and the liquidity and the slick chart, is just bookkeeping until the oracle says what happened. That's why it's worth calling it the truth layer: reliable, attack-resistant truth is not a feature bolted onto a prediction market, it's the load-bearing wall.
It also reframes a deal that read, at the time, like a simple bet on the category. When ICE — the owner of the New York Stock Exchange — agreed in late 2025 to back Polymarket with up to two billion dollars, the headline framing was a wager on prediction markets.10 But the structure of the deal was about distribution of event data — ICE became the channel for the outcome data these markets produce. That only makes sense if you believe clean, settled truth about the world is itself a valuable commodity. A pipe is worth building only when something precious flows through it. The oracle is the well; the data is the water; and a serious institution paid to own the tap. Truth, it turns out, has a price.
No oracle is trustless all the way down. Follow any design to the bottom and you hit a human or governance assumption — a reporter set you trust to stay honest, a token vote you trust not to be captured, a venue you trust to rule in good faith. Decentralization can make that assumption costlier to break and harder to hide, but it never removes it. "Trustless truth" is marketing; the real engineering is making the trust you can't avoid as expensive to betray as possible.
The bottleneck isn't liquidity
Everything so far has been about a handful of markets — whether this contract settles on the truth. Now widen the lens, because the hardest version of the problem only appears at scale, and there it stops being a question of mechanism and becomes a question of volume.
In July 2025, one of the most-traded markets on Polymarket was a question a child could answer: would Volodymyr Zelensky wear a suit before July? It drew about $242M in volume and more than $150M in live bets. Then the Ukrainian president showed up at a NATO summit in a dark jacket-and-trousers ensemble, most outlets called it a suit — and the market couldn't decide. It first resolved YES, then, after nine days and a cascade of disputes, flipped to NO.11 A quarter-billion dollars rode on a definition of "suit," and the machine built to settle it nearly broke.
Most of the energy in this space goes to liquidity — how to get enough traders into a market that its price means something. That problem is real, and it's being solved. The problem nobody puts on a pitch deck is the other half: resolution at scale. A price is only a probability if the market reliably settles on the truth, and at the volume this whole field is racing toward — a market under every story, a thousand markets per country — resolution, not liquidity, is the wall.
Recall the two live models. A centralized committee — the model Kalshi uses — is fast and accountable, but it's a trusted third party, it's bound to one jurisdiction's rules, and it has a human in the loop for every contested call. An optimistic oracle — the model UMA runs for Polymarket — is permissionless: anyone can propose an answer, anyone can dispute it, and a contested question escalates to a vote of token holders. Both settle a few thousand markets fine today. The trouble is that neither was built for a million — and the reasons why are exactly the seams we've been tracing.
Why it breaks at scale
The Zelensky market wasn't a freak event. It was a preview of four failure modes that each get worse, not better, as the number of markets explodes.
Ambiguity. This is the same enemy from earlier — uncertainty about the question, not the world — but scale changes its weight. "Did he wear a suit" sounds binary until $242M depends on the lapels. At a few thousand carefully-written markets you can keep questions crisp. At a million — every news story, every macro print, every match — ambiguity stops being the exception and becomes the median case. Most of the world does not resolve cleanly into YES and NO.
Capture. This is the whale-vote risk made concrete. A Wall Street Journal investigation found that in most disputed Polymarket markets, more than half the UMA votes came from the ten largest wallets; at least 60% of active voters could be linked to live Polymarket accounts; and roughly one in five disputes had a voter with a direct financial stake in the outcome they were ruling on.12 "Decentralized truth" turns out to be governable by capital — the richest holders decide what happened. A $60M dispute over a Bitcoin-sale market and an evidence-free "UFO" resolution forced through by whales told the same story.12
Latency and jurisdiction finish the list — and these are the two the single-market view never surfaces. Nine days to settle one contract is fine when you have a thousand markets and a slow news cycle; it's catastrophic when you have a million and the disputes pile up faster than they clear, because disputes don't parallelize. And the moment markets cross borders, "the truth" forks: a US data feed and a Vietnamese one can report the same number differently, and no single oracle is trusted everywhere.
What a truth layer that scales looks like
The fix is not a better committee or a bigger token vote. It's to stop treating all questions the same. Most markets are trivially resolvable; a few are genuinely contested; the architecture should match the work to the difficulty.
Tier one is the easy 90%. A rate decision, an index close, a match score, an official CPI print — these have an authoritative, machine-readable source. Wire the market to the feed and it settles itself, instantly, with no human and no vote. The discipline here is upstream: write questions that have a clean source, and refuse to list the ones that don't. Resolvability becomes a design constraint, not an afterthought.
Tier two is where it gets interesting, and where the timing is finally right. Large language models can now read the same wire reports, primary documents, and footage a human resolver would, at the speed and breadth a million markets demand — and, crucially, show their reasoning. An AI proposes a resolution with its sources cited; if nobody contests it within a window, it stands. This is the only mechanism that plausibly scales to the volume that "every story a market" implies.
Tier three is the genuinely contested tail — the Zelensky lapels of the world. Here you want humans, but with skin in the game: arbitrators who post a bond and lose it for a bad ruling, an appeal path that ends, in the limit, in something like real courts. The goal isn't to eliminate human judgment; it's to make sure only the 1% of questions that truly need it ever reach a human, and that the human is paid to be right.
Today every disputed market goes to the same slow, capturable vote. The fix is to auto-settle the trivial, let AI propose on the fuzzy, and reserve humans-with-stakes for the rare question that's actually hard.
Who watches the resolver?
Putting an AI at the center of the truth layer raises the obvious objection, and it's a good one: a model can be wrong, and a model can be gamed — a well-crafted fake source, a prompt-injected document, a coordinated flood of misleading reports. An AI resolver is not an oracle of truth; it's a fast, legible first proposer. That's why the appeal path matters: the AI's job is to be right on the easy-and-medium cases and to fail loudly on the hard ones, kicking them up to humans. The security doesn't come from trusting the model; it comes from the economics around it — the same cost-to-corrupt logic from before: bonds, challenges, and a final human backstop — exactly as the optimistic oracle intended, just with a far better first guess so the expensive human vote almost never fires.
And some questions should simply never be markets. "Did he wear a suit" was a bad market not because the oracle failed but because the question was irreducibly subjective — there was no fact of the matter to settle. Part of building resolution that scales is the humility to not list the unanswerable: if a question can't be tied to a source a reasonable person would accept in advance, it isn't a market, it's a debate with a pot.
The seam to watch
So the next time you see a market settle — on a regulated venue in a day, or on-chain in seconds — look past the trade to the resolver underneath it. Ask the questions that actually decide whether the number ever becomes money: who reports the outcome, what does it cost to lie, how big would a payout have to get before corrupting the resolver becomes the rational move, and — at scale — does this question even have a clean source, or is it a debate wearing a market's clothes? Those aren't trivia. They're the same question this whole piece has been circling, from the rulebook to the oracle to the tiered stack: can you trust the thing that decides what's true?
This is the part that doesn't fade with a nicer interface. Liquidity gets the headlines because it's visible — you can watch the order book fill. Resolution is invisible right up until the moment it isn't, and then it's a $242M dispute waiting to happen. A demo can render a price; a credible exchange has to be able to settle — to carry truth from the world to the ledger in a way that holds up even when real money is trying to bend it, fast and cheaply and capture-resistantly, at machine scale. Whether the resolver is a regulator, an oracle, or a tiered stack of all three, that's the bar — and clearing it, not the matching engine or the chart, is what a serious venue is actually selling. That truth layer is the real ceiling, and it's the unglamorous thing we spend our nights on at Seeker. The price is the easy half. Deciding what's true — at scale, on a truth your users actually trust — is the rest.
- Kalshi's contracts are governed by published rulebooks that specify, in advance, the authoritative settlement source and the exact resolution criteria for each market; as a CFTC-regulated designated contract market it adjudicates under those terms and operates under federal oversight with segregated customer funds. See Kalshi's market rules and the CFTC's event-contract framework.
- Polymarket settles the bulk of its markets via UMA's Optimistic Oracle — a decentralized mechanism in which an asserted outcome is accepted by default unless disputed. See Polymarket's documentation on resolution and the UMA integration.
- UMA Optimistic Oracle: an outcome is proposed with a bond, a liveness/dispute window opens, and any disputed assertion escalates to UMA's Data Verification Mechanism, where token-holders vote and the incorrect party forfeits its bond. See UMA's protocol documentation (docs.uma.xyz).
- On the category's trajectory and composition: combined monthly volume climbed from under $100M in early 2024 to past $13B by the end of 2025 (Pew; The Block), and sports contracts have made up roughly 80% of volume since launching in mid-2024 (The Block).
- The pattern of resolution disputes through 2024–2025 is described here at the level of recurring failure modes — ambiguous wording, unanticipated edge cases, and "technically true" readings — rather than litigating the specifics of any single contested market.
- The "oracle problem": because a blockchain must reach deterministic consensus, smart contracts are isolated from off-chain state and cannot natively fetch external data — an oracle is the bridge that delivers it. See the standard treatments in the Ethereum documentation on oracles and Chainlink's writing on the oracle problem.
- Aggregated / decentralized price oracles: many independent node operators source from multiple venues and the readings are aggregated (commonly around the median) before being written on-chain, so no single feed or exchange can move the reported value. Chainlink Price Feeds are the most widely used example; see the Chainlink documentation.
- Oracle manipulation has repeatedly ranked among the leading root causes of DeFi exploits — typically by distorting a thin or single-source price feed so a contract acts on a value that was never real. See post-mortem aggregations from security auditors and incident trackers (e.g. Chainalysis and rekt.news write-ups of price-oracle attacks).
- Governance / whale-vote capture: where resolution escalates to a token-weighted vote, an actor able to amass sufficient voting weight could in principle force an incorrect settlement. The mitigating discipline is economic — keeping the cost to corrupt the vote above the value at stake in any single market. This is a recognized, structural risk of decentralized oracles rather than a claim about any specific incident.
- In October 2025, Intercontinental Exchange (ICE, the owner of the NYSE) agreed to invest up to $2B in Polymarket, the deal was structured around the global distribution of Polymarket's event data. That a major exchange operator paid to distribute settled outcome data is itself evidence that reliable real-world truth is a valuable asset. See contemporaneous reporting (ICE; FinTech Weekly; The Defiant).
- The Polymarket market "Will Zelensky wear a suit before July?" drew ~$242M in volume and $150M+ in bets; after he appeared in a dark suit-like outfit at the June 2025 NATO summit, the market resolved YES, then flipped to NO after roughly nine days of disputes — turning on whether the outfit counted as a "suit." CoinDesk (Jul 7, 2025); The Defiant; CryptoSlate; blocmates.
- A Wall Street Journal investigation (May 2025) found that in most disputed Polymarket markets more than half of UMA votes came from the ten largest wallets, ≥60% of active UMA voters could be linked to live Polymarket accounts, and ~1 in 5 disputes had a voter with a financial stake in the outcome ruled on; related controversies include a ~$60M dispute over a Strategy (MicroStrategy) Bitcoin-sale market and an evidence-free "UFO" resolution pushed by large holders. The Defiant; Crypto Briefing; Cryptopolitan; Bitget News.