Onchain proof on Arbitrum that an AI agent stayed within the authority it was granted.
ArbiProof is onchain proof that an AI agent stayed within the authority it was granted, on Arbitrum. Each agent gets a scoped permission registered onchain that fixes which contract it may call, the one recipient it may pay, and a per call cap. The agent declares its intent, acts only inside that permission, and writes a receipt binding its authorization, its intent, and the resulting transaction. A live verifier loads any receipt, pulls the real transaction, and checks it against the grant row by row, stamping the verdict CLEARED or FLAGGED. The verdict is recomputed from the chain, so the agent's own word is never trusted. Deployed and working on Arbitrum Sepolia with a live frontend covering both the authorized action and the blocked one.
Built ArbiProof end to end during the buildathon. Wrote and tested the contracts in Foundry, a ReceiptRegistry holding grants and receipts and a BoundedAction target that enforces the allowlist and cap onchain so any out of scope call reverts. Deployed all contracts to Arbitrum Sepolia. Built the TypeScript agent that declares intent, executes through its grant, and anchors receipts, with both an authorized run and a blocked run working. Built and shipped the verifier frontend that audits any receipt and renders a CLEARED or FLAGGED verdict, deployed live on Netlify.
Not raising. Independent buildathon project.
