Challenge Market
Karatum's verdicts and certificates should not depend on trusting Karatum. Anyone can put a bond against a verdict or a certificate and force a re-check. If the challenger is right it gets its bond back plus most of Karatum's matching stake; if it is wrong it loses its bond. Bonds, outcomes and payouts live in a contract; the re-checks are reproducible by anyone.
Verdict challenge
Claim: this verdict does not follow from its trace. Decided by re-running the published judge on the stored trace, the same computation anyone can do on /verify. When a zk proof that judge(trace) = verdict exists in ZkJudgeVerifier, anyone can settle the challenge on-chain without asking Karatum: the contract checks the proof itself and there is no appeal.
Certificate challenge: short the agent
Claim: this agent no longer performs at its certificate's lower bound. Decided by a fresh re-trial in Mirror from a seed that stays hidden (only its hash is published) until the result is posted. The challenger wins when even the upper end of the re-trial's 95% interval is below the certified lower bound. A won challenge revokes the certificate, opens an incident and, for agents with wallets, revokes them in the Guard.
Rules
- Contract
- ChallengeManager 0x9f432E54316fe8458B5CaaE01e02A70887AD3266 on Ethereum Sepolia (source verified on Sourcify)
- Bond
- max(10.00 tUSDC, 1% of the subject's value). A certificate is valued at 100 tUSDC per karat, so a 23-karat certificate takes a 23 tUSDC bond. Test USDC on Sepolia has an open mint: no real money is involved anywhere.
- Matching stake
- Every challenge is matched by the same amount from Karatum's defender pool, so Karatum loses money when it is wrong.
- Payout
- Winner: own bond back plus 80% of the loser's stake. The other 20% goes to the treasury. Paid as credits, withdrawn by the winner (works even while new challenges are paused).
- Windows
- A verdict can be challenged for 30 days after it is decided, a certificate until it expires. The resolver must answer within 2 days, or the challenge expires and every bond goes back. After a resolution against the challenger there is a 1 hour dispute window.
- Escalation
- The challenger can dispute a resolution against it (not a zk-settled one) by doubling its bond: validators staked in ValidatorStaking re-derive the outcome from the public evidence and vote for 1 day, weighted by stake. Overturned on a strict majority with a quorum of 600 KARAT staked; without quorum the resolution stands and the escalation bond is refunded.
- Anti-griefing
- One active challenge per subject, a 1 day cooldown on a subject after a failed challenge, a 3 days cooldown on a challenger who lost, bonds that scale with the value at stake.
- Which re-trials run today
- Re-trials of Karatum's deterministic scripted builds run now. Re-trials of LLM builds are queued and marked pending: Karatum's LLM gateway is down. A queued challenge that cannot be re-tried before its deadline expires with a full refund.
Numbers
Drills (rehearsals labelled as such) are included: 1.
Challenges
API: GET /v1/challenges, GET /v1/challenges/market, GET /v1/challenges/subject?kind=…&ref=…, GET /v1/challenges/:id/evidence · Methodology: docs/methodology/challenges.md