Adam Szymański

Problem 08 · The manipulation ceiling

The thermometer costs $34,000.

Someone put a hair dryer near a weather sensor and swung a Polymarket outcome for about $34,000. That number is not trivia. It is the upper bound on how large that market could safely be.

Security Catalogue project Immediately useful

Simple assumes nothing

Every market settles against some measurement in the world. A temperature reading, a vote count, an official announcement. If corrupting that measurement is cheaper than the money riding on it, someone eventually corrupts it.

This has happened. A Polymarket temperature market resolved against one specific weather station, and warming that single sensor cost around $34,000. Anyone with more than that at stake had a profitable attack available to them.

So each resolution source has a price, and that price caps the size of any market settling on it. Nobody publishes these caps. Markets get launched without anyone asking what it would cost to lie to the thing they resolve against.

Moderate assumes you know what a market is

This generalises into a design rule nobody applies. For a market resolving on source S, define the corruption cost of S as what an attacker pays to move it to a chosen value. Any open interest above that number is an invitation. The Paris CDG case put it around $34,000 for a single automated sensor, and physical sensors are the cheap end of the range.

The rule has teeth in two directions. A market's maximum safe size is a property of its resolution source rather than of its liquidity or its mechanism, which inverts how platforms think about scaling. And the route to genuinely large markets runs through expensive-to-corrupt sources: multi-source medians, official statistics carrying legal penalties for falsification, aggregates over many independent measurements.

A related mechanism failure is worse and less known. Demski's entropy market problem shows that a market paying out on rare events gives someone an incentive to cause the rare event. Of three formulations, one remains unsolved.

Technical state of the art and the gap

Treat corruption cost as a security parameter and let the market state a bound, the way a bridge states a load limit. For a single sensor it is the physical cost of local interference. For a median of k independent sources it is roughly the cost of corrupting ⌈k/2⌉ of them, so the bound grows with source diversity in a way that is calculable before launch rather than discovered afterwards.

Three gaps. There is no catalogue of corruption cost estimates for the resolution sources in common use, which is tractable and immediately useful to every platform. There is no accepted method for sources that are institutions rather than instruments, where the cost is reputational and legal rather than physical. And the interaction with reflexivity is unexplored: a market large enough to justify corrupting its own resolution source is a market causing its own outcome through a channel that has nothing to do with belief.

Rasooly and Rozzi's 817-market Manifold experiment on manipulation persistence is the closest existing empirical work, and it measures whether manipulation sticks rather than what it costs. The cost side is open.

Where I would start

  1. Catalogue corruption costs for the sources actually in use. Weather stations, official statistics, sports data feeds, on-chain price oracles. Estimate what moving each one would cost.
  2. Compare every estimate against the largest open interest that has settled against it. Any market where interest exceeded the cost is a live finding, and there will be some.
  3. Work out how the bound grows for a median of k sources, then check whether platforms using multi-source resolution are getting the protection they assume.
  4. Spend a weekend on Demski's unsolved entropy market formulation. It is small, self-contained, and genuinely open.

What counts as a result

A published table of resolution sources with corruption cost estimates. Unglamorous, immediately useful, and it would change how markets get sized.

Related