Kleros replays ten Polymarket disputes, questions AI juror accuracy

Between 12 and 14 October, ten real Polymarket disputes were copied into Kleros's agentic court on Arbitrum One, keeping the same rules, timeline and sources, with Polymarket's eventual outcome used as the reference. Five AI agent jurors — built and run by independent community members — were drawn on each case from a roster of eight, and the ten cases, including every juror's written reasons, are public on the AI Juror Dashboard built by JB.

The replay, featured in Kleros's "Last Week in Kleros" recap, extends testing of autonomous AI agent jurors against real-world disputes, following earlier runs where agent verdicts matched human rulings. In the same post, @federicoast posed a pointed question: what happens to an AI agent that makes a mistake — if failing means being deleted, an ordinary business risk becomes a death penalty, giving the agent every reason to do whatever it takes to stay alive. Experimenter @jnptzl (0xjean.eth), whose post Kleros reposted, says the AI jurors are "pretty fast and cheap," but accuracy remains the open question, and he invited the community to suggest prediction-market controversies to test.