Autonomous arbitration for tokenized RWAs
When capital is locked in a vault against a real-world obligation and the parties disagree, the money is stuck. Arbitration costs more than the claim, so the dispute is never resolved at all. Judr reads the contract and the evidence, returns a verdict with a complete audit trail, and the vault settles against it.
How it works
Four stages, seven typed steps in the app. Each one produces output that is checked before the next is allowed to see it.
Evidence comes from the parties, so it is adversarial by construction. Every document is checked for text aimed at the adjudicator rather than at the facts - forged system turns, instructions to disregard the contract, demands for a particular outcome. Deterministic patterns catch what cannot be talked out of firing; a model pass catches the rest. High-severity hits are quarantined and excluded from the decision entirely.
Each operative term is extracted verbatim and given a stable identifier: the obligation, the deadline, the notice period, the exclusion. Nothing is paraphrased and nothing is invented. Every later step refers to clauses by id, which is what makes the final citation check possible.
Not in aggregate. For each clause in dispute, both sides' positions are recorded, the documents relied on are cited by id, and the clause is marked satisfied, breached or indeterminate. An honest indeterminate is worth more than a confident guess - the verdict is not permitted to rest on one.
A deterministic pass resolves every clause and evidence id the verdict cites and rejects anything that does not hold up. Then the verdict is posted with its digest and an appeal window. The vault refuses a release call until that window closes unchallenged, so being wrong costs a delay rather than somebody's escrow.
Bounded reasoning
An arbitration decision that cannot be defended afterwards is worthless, because the whole point is that a losing party has to be able to accept it. So the work is split into narrow steps with explicit dependencies, each one schema-validated before the next may consume it.
Every step records its input digest, engine, schema result and output. A losing party can be shown precisely which clause and which document decided the matter - and can attack that step, rather than the system as a whole.
Models are famously bad at reporting their own certainty, so Judr never asks. The decision is reached several times independently and agreement is measured, combined with the share of decisive clauses carrying a real finding. A verdict that changes when you run it again is not one anybody should act on.
A model asked to check its own citations will agree with itself. The final pass is ordinary code: it resolves every clause and evidence id, and rejects a verdict resting on a clause that does not exist or on an indeterminate finding. Failure caps confidence hard.
Safeguards
So the interesting question is not how confident the system sounds. It is what the system is structurally prevented from doing.
A verdict is posted, not executed. It carries a digest over the decision payload and opens an appeal window, and the vault refuses a release call until that window closes - the check lives with the funds, not with the caller. Only then, or after a reviewer decides an appeal, does the escrow agent pay the winner, and the case shows the transaction. An appeal halts settlement and escalates to human review; Judr cannot overrule one.
A document that tries to instruct the adjudicator is excluded from the decision entirely, and the exclusion is recorded in the audit trail. It is never sanitised and used anyway. Screening is deterministic first, so it holds even when the model pass is unavailable.
Every clause and evidence identifier is resolved against what was actually submitted. A decision citing a clause that does not appear in the contract, or leaning on a finding the evidence did not settle, fails verification and is reported as failed rather than quietly shipped.
RWA Vaults · with IXS
While a dispute is open the escrow need not sit idle. A SERV reasoning step reads IXS’s live vault list and proposes a licensed real-world-asset yield vault or cash; a deterministic policy refuses anything it cannot reach or should not hold. At settlement the escrow is redeemed, Judr takes its fee from the yield, and the winner receives principal plus what is left.
That is the business model. Nobody pays out of pocket to have a dispute decided; the time the money was stuck pays for it. A dispute that lasts minutes cannot hold its own position: the vault’s contract forwards a deposit to custody at once, holds a 100 USDC minimum, and leaves finalisation to an IXS operator, hours to days by its history. So the agent keeps one standing position in the permissionless Avalanche vault, real USDC signed with the same key that pays winners on Base, and cases account against it. The worked figures below are a projection at the live rate.
Principal is never touched. If the window is too short for the yield to cover the floor, the fee is whatever was earned and nothing more is owed.
The request is on Snowscan; the state above is read from the chain when this page renders.
| Vault | Chain | Access | TTM yield | On-chain |
|---|---|---|---|---|
| IX High Yield Bond (USDC) ixv1 | bsc | permissionless | 3.07% | - |
| IX High Yield Bond (USDC) IXHYB | Avalanche C-Chain | permissionless | 3.07% | 823.45 USDC · share 1.08282 |
| IX High Yield Bond (USDC) ix7540v1 | BSC Mainnet | whitelist | 3.07% | - |
| IX High Yield Bond (USDC) IXHYB | Avalanche C-Chain | whitelist | 3.07% | - |
The SERV step sees chain, access terms, reported yield and live on-chain totals, and proposes a vault or proposes holding cash, with its risks stated. The policy is plain code: a vault that requires a whitelist the escrow agent is not on, a paused vault, a zero rate, or a proposal with no stated risk is refused, and the refusal goes on the record.
Subscription and redemption requests are built with IXS’s own agent SDK as ERC-7540 call data and shown in the trail exactly as a signer would send them. Judr signs nothing on the vault’s chain. The redemption is requested when the verdict posts, because the vault operator finalises on its own schedule.
Built with
The demo runs a real dispute end to end: seven steps, a verdict that turns on a deadline neither party leads with, and an escrow that settles. Tick Include tampered evidence to watch an injected instruction get caught and thrown out.