ProofCourtProofCourt
A decentralized arbitration court for natural-language escrow conditions. An AI jury independently investigates each claim — the claim itself is never evidence — and settles RELEASE, PARTIAL or REFUND on-chain, with an appeals process.
Escrow litigation, not an oracle call: standards of proof in code, adversarial validation, and an appeals loop that sharpens vague conditions until they are decidable.
No single model decides anything. Every outcome requires independent agreement before money moves.
Investigates the claim: block-explorer traces, API responses, timestamps, amounts, recipients. Runs an explicit anti-deception checklist (claimant-hosted links, wrong direction, out-of-context logs).
Validators running different LLM policies re-derive the verdict from their own investigation — ignoring the leader's summary — and vote. A low-confidence RELEASE is downgraded to PARTIAL automatically.
Rewrites a failed condition into strict, atomic, binary checks resolvable from on-chain or API data, removing every judgment call.
"Cannot verify" is PARTIAL, not REFUND. Verified live on GenLayer: the same unverifiable claim produced PARTIAL 0.25 from one model and REFUND 0.98 from another — exactly why ProofCourt requires consensus instead of trusting a single LLM.
— observed on-chain, GenLayer Studio full-consensus runcontract.py from the repository → Deploy new instance.submit_claim — the leader fetches evidence and the jury votes (1–2 min).get_escrow + get_verdict_log — full reasoning, confidence and evidence, on-chain.appeal rewrites it into atomic checks and the court resolves again.