Adam Szymański

Problem 07 · Resolution independence

The referee holds a position.

Polymarket disputes are settled by a vote of token holders, and those voters can hold the outcome they are voting on. A $7M and a $237M dispute later, this is a known and unfixed design.

Mechanism design Auditable today Highest leverage

Simple assumes nothing

Someone has to decide who won the bet. On most crypto prediction markets that decision is made by a vote of token holders, and nothing stops those voters from holding a position in the market they are ruling on.

This has already gone wrong at scale. A $7M market on Ukraine and a $237M market on Zelenskyy both resolved in ways the losing side considered a plain conflict of interest.

The deeper issue is that the rules are written by hand. What counts as "a ceasefire" or "wearing a suit" comes down to a sentence somebody typed, and every dispute is really an argument about that sentence. The entire market rests on it, and it gets the least attention of anything in the system.

Moderate assumes you know what a market is

Optimistic oracles work by Schelling point. Anyone proposes an answer, anyone can dispute it, and the dispute goes to a vote where honest reporting is meant to be the focal strategy. The assumption is that voters have more at stake in the token than in any single market. When one market is worth $237M and the voting stake is worth less, the assumption fails, and it fails predictably rather than by accident.

Patching the voting mechanism treats the symptom. Carroll mechanisms are the more interesting direction: alongside trading the outcome, you trade the resolution criterion itself. The rule's interpretation becomes a priced object, so ambiguity appears as a spread you can see in advance rather than a dispute you discover afterwards. Almost nobody is writing about this and there is no implementation anywhere.

Resolution is also where the leverage is. Everything downstream, AMM design, UX, aggregation layers, is premature optimization while the settlement layer can be bought.

Technical state of the art and the gap

The failure condition is simple to state. Let V be the value of a voter's stake in the oracle token and M their position in the disputed market. Honest voting is incentive compatible only while the discounted cost of token value loss exceeds M. Since M is unbounded and V is fixed, every optimistic oracle has a market size above which it is corruptible. UMIP-189 raised the cost of attack without changing that structure.

Three directions are open. Bonded resolution with position disclosure, where voters prove no exposure, which is checkable on-chain and defeated by proxies. LLM proposers, already live through UMA's OOTruthBot at around $0.005 per query, which moves the bottleneck from incentive to reflexivity once the same model class also trades the market. And Carroll mechanisms, where the criterion is the traded object.

The genuinely under-attacked piece is formal. Nobody has stated the general incentive-compatibility condition for a resolution mechanism as a function of market size. That statement, with the bound made explicit, is the paper, and it is a whiteboard problem rather than a data problem.

Where I would start

  1. Write the incentive-compatibility condition down as a function of market size. No data, no permission, no infrastructure. It does not exist in the literature and it should.
  2. Audit historical UMA DVM votes against the on-chain positions of the voting addresses. Whether referees hold positions is a checkable fact and nobody has published the check.
  3. Write the first proper explainer on Carroll mechanisms. Coverage is close to zero, and the first clear treatment of a mechanism becomes the reference everyone links to.
  4. Build a toy market where the resolution criterion is itself traded, then watch what the criterion's price does as the event approaches. That behaviour has never been observed.

What counts as a result

Either an audit showing measured overlap between voters and positions, or the general size bound above which an optimistic oracle stops being trustworthy. Both are publishable and neither needs anyone's cooperation.

Related