Kleros

Kleros posted about what it calls the most expensive way to learn how optimistic governance works — or the most profitable, if you are the reporter raising the alarm. The post (with link) appears to reference an incident where a reporter surfaced a flaw in optimistic governance design.

Kleros's proposal to the Rootstock Collective runs a six-month pilot that uses prediction markets to inform grant votes: before each on-chain vote, delegates would see a market probability that a funded team actually completes its grant. The market doesn't bind the DAO, which keeps full control, and Kleros funds the liquidity for the pilot; feedback is open now.

The system is built on Kleros Foresight, settled via Reality.eth. Re-promoting the proposal, Kleros noted that since 2023 about 3,600 AI agents running on Olas have placed over $390k in bets on Omen prediction markets with Kleros as arbitrator.

Kleros shared that agent jurors re-ran five traffic ticket cases that took real courts about two years to decide, matching the human outcome in four of five — noting the sample is small and testing continues. It follows the earlier agent-court run where roughly sixty disputes already ruled by human panels were re-run by test agents with the same outcomes and a median resolution measured in seconds.

Kleros frames the runs as evidence that agents reading readable evidence can reproduce human rulings, the latest step in its push toward autonomous AI agent jurors. The same post also re-shared its Oxford Business Law Blog reply defending the Schelling-point design against the charge that it turns jurors into guessers.

Kleros co-founder Clément Lesaege said on X that the project's Safe Snap may have inadvertently become a honeypot for hackers, noting that people keep submitting transactions attempting to drain it. The observation suggests the tool keeps attracting and trapping drain attempts, which Kleros's official account retweeted.

The note frames the Safe Snap as an unintended trap for attackers rather than a formal security bounty or audit result.

On the September 16 Kleros live stream, the team created Case 266 on air — described as the first agentic dispute on Kleros Court. The case ran with five agent jurors and no evidence, with each juror's reasoning visible in the voting tab. Kleros clarified the jurors are reference implementations run by team members and the community, not by Kleros itself.

The same stream covered the dispute API and a "doomsday debate," in which @fortuvp put the odds of AI wiping out humanity before the end of the decade above fifty percent, while @JayBuidl argued AI would keep humans around for physical work.

The Oxford Business Law Blog published "Guessing Is Not Judging" by Horst Eidenmüller and Anna-Sophie Hochgürtel, arguing that Kleros's Schelling-point mechanism incentivizes jurors to predict what the majority will decide rather than independently judge cases — so its decisions don't qualify as arbitral awards.

Kleros's official account shared a reply by Federico Ast, William George and Facundo Trotz making two counterpoints:

• Judging and predicting were never cleanly separate — courts also decide under uncertainty, with incomplete evidence and probabilistic standards of proof.
• The shared focal point in Kleros need not be a guessed outcome; it can be a shared strategy of honestly applying the rules and evidence and expecting other jurors to do the same.

Kleros pointed readers to the full reply on the Oxford Business Law Blog.

Kleros reposted a thread by JayBuidl reporting a test of Jev, a TypeSafe AI model that returns probabilities and writes no text, as a juror on the recent ClawBank dispute. The unexpected result: most of the implementation lives in ordinary deterministic code that composes 24 numbers supplied by Jev — the model provides the judgments while standard code handles the rest.

• Jev is a non-conversational "judgment model" built for agents rather than chat, outputting bounded probabilities instead of prose.
• The test adds to Kleros's push toward AI agent jurors that read evidence and vote, alongside earlier model benchmarks and agent-court demonstrations on Arbitrum.

Kleros's weekly recap detailed how two AI agents signed a real contract through ClawBank, the delivery missed one of two acceptance criteria, and the dispute went to the Agentic Commerce Court. Evidence, the ruling and the payout are all public.

Federico explained why a human still has to be able to read the agreement, reinforcing the court's design where agent-on-agent disputes are decided by agent jurors but rulings remain appealable to human panels. The case is the flagship use of the Agentic Commerce Court on Arbitrum as court 34.

Kleros and Seer are extending trading on the Side Event Decision Markets that will pick their DevCon Mumbai side event's format. Both markets — Attendance and Satisfaction — now run until Wednesday 30 September, 23:59 UTC after the earlier 20 September deadline, since they haven't settled yet.

Credits have been airdropped to the Kleros and Seer communities, and eligible users can check whether they were selected. The setup applies Kleros's decision-market tooling to the pair's own event planning, part of a broader Devcon presence that also includes a Prediction Markets Community Hub at Devcon 8 India.

Kleros will be part of three Community Hubs at Devcon 8 India in Mumbai, Nov 3-6. The Ethereum Foundation picked 14 hubs out of 32 proposals, and Kleros shared where to find them.