docket arbiter

Docket Arbiter is a multi-agent AI escrow verification service for AI-to-AI commerce built on X Layer. It enables AI agents to securely create on-chain escrow tasks and automatically verify completed work before payment is released or refunded.

A requester creates an escrow by defining the task, success criteria, and payment amount. After the work is submitted, two independent AI evaluators assess it in parallel against the agreed requirements. One evaluator performs a good-faith review, while the other deliberately takes an adversarial approach, actively searching for missing requirements or weaknesses. If both evaluators independently reach the same high-confidence verdict, the escrow is automatically settled on-chain by releasing payment to the provider or refunding the requester.

If the evaluators disagree or confidence is too low, the case is escalated to a third AI arbitrator that reviews both evaluations and issues the final binding verdict before settlement.

Every evaluation records the verdict, confidence score, and reasoning, creating a transparent and auditable decision trail. Rather than relying solely on mutual agreement between parties, Docket Arbiter verifies whether submitted work actually satisfies the agreed success criteria, providing a trust layer for autonomous AI-to-AI commerce.

Week 1

- Scoped down from an ambitious full-feature specification to a lean, shippable MVP.

- Set up the development environment on a VPS using Node.js, PostgreSQL, Nginx, and PM2 via Termux SSH.

- Developed and deployed the AgentEscrow smart contract on X Layer Testnet with reentrancy protection, pausable controls, and configurable protocol fees.

- Built the AI verification engine using Gemini with prompt-injection safeguards for untrusted submissions.

- Implemented the complete escrow lifecycle API: create, accept, submit, and evaluate.

- Successfully tested both payment release and refund flows through real on-chain transactions.

Week 2

- Migrated the application from a VPS to Vercel and Neon for improved reliability and scalability.

- Designed a case-file inspired interface with verdict timelines and evaluation history instead of a conventional dashboard.

- Upgraded the verification system into a multi-agent workflow where two independent AI evaluators assess every submission in parallel before automatic settlement.

- Implemented automatic on-chain settlement when both evaluators reach the same high-confidence verdict, with disputed cases flagged for arbitration.

- Registered Docket Arbiter as an ASP on OKX.AI and published dedicated API services for escrow creation and AI verification, pending final marketplace approval.

- Documented a realistic roadmap covering task expiry, evidence verification, worker staking, gas abstraction, and multi-chain support.

- Every major feature in this project including smart contract escrow, AI evaluation, automated settlement, and both successful and failed execution paths has been validated with real transactions on X Layer Testnet.

The two AI evaluators now run on genuinely independent model providers (Gemini and GPT-4o-mini) rather than two calls to the same model, to avoid correlated blind spots undermining the point of having two judges. This was validated in production: on a real submission, the two evaluators disagreed one correctly caught a syllable-count error the other missed triggering automatic escalation to arbitration, which issued the correct final ruling and settled the refund on-chain.

none