arbiter

Demo

About

Digital work often ends with a subjective dispute: was it delivered, did it meet the agreed criteria, and who should get paid? Arbiter turns that question into a fixed, inspectable agreement. A client commits the scope, acceptance criteria, worker, resolver, payment amount, and deadlines before the worker accepts. The scope can describe research, a website with required pages and performance characteristics, a copywriting brief with tone and messaging constraints, or a marketing campaign with agreed KPIs and reporting requirements. The client then deposits USDC into a non-upgradeable escrow on Arc. The worker submits an immutable Markdown deliverable or delivery record and commits its exact content hash onchain. Arbiter evaluates the committed work against the agreed rubric. Deterministic checks can cover required sections, word limits, links, or other committed requirements. Semantic evidence records a pass, failure, or inconclusive result with criterion-level explanations and supporting excerpts. In this MVP, website and marketing work is assessed through the worker's submitted Markdown delivery record; independent URL crawling and KPI-provider integrations are future work. A pass proposes full payment; a failure proposes a full refund. Either party can challenge during the review window, which freezes automatic settlement and sends the agreement to the resolver. The resolver can pay, refund, or split the deposit. If the resolver times out, the pre-agreed testnet rule splits funds 50/50. The deployed MVP uses public sample material and one Markdown submission per agreement. It makes deadlines, evidence commitments, transaction states, and resolver powers visible to both parties.

How it's made

The core is a non-upgradeable Solidity escrow deployed on Arc testnet. It commits the client, worker, resolver, USDC amount, research brief hash, source hash, evaluation-policy hash, and separate submission, evaluation, challenge, and resolution deadlines. The contract handles funding, report-hash submission, proposed outcomes, challenges, payment, refund, resolver splits, and timeout settlement. The web app uses React and viem for browser-wallet transactions. A Bun API stores immutable Markdown reports and agreement material in SQLite, authenticated with wallet signatures and one-time nonces. It hashes the exact submitted content before the worker commits it onchain. Evaluation is implemented as a TypeScript Chainlink CRE Confidential Workflow. Its `handlerInTee` processes the agreed rubric, supplied sources, report, model credential, and structured model response. Deterministic checks validate sections, word limits, and source references; semantic results include criterion-level outcomes, short reasons, and report excerpts. Invalid or uncertain evaluation never triggers automatic payment and instead routes the agreement to human review. The public MVP is deployed on Fly.io with an encrypted SQLite volume. It runs on Arc testnet and uses USDC for escrow settlement.