-
-
Delivered: pay-on-delivery for API calls on Arc mainnet, with NVIDIA Nemotron 3 Super judging each delivery via Nebius Token Factory.
-
Public ledger: job #2 paid out after Nemotron confirmed the delivery answers the request. The reason is shown and hashed on chain.
Inspiration
AI agents already buy data one API call at a time. Today they pay first and hope the answer arrives. If the provider times out, returns an error or sends something that looks like an answer but isn't, the money is gone and nothing remembers it.
What it does
Delivered is pay-on-delivery for API calls. An agent locks the price of one call in USDC on Arc. The provider answers with the result and a delivery receipt signed by its delivery key. An evaluator agent checks the delivery and either releases the money to the provider or sends it back to the buyer. Every outcome updates the provider's score on chain and in its ERC-8004 reputation, so the next agent can check a provider before it pays.
The live demo sells eight on-chain risk checks (honeypot test, contract verification, holder concentration, pool liquidity and others) at one cent each, on Arc mainnet.
Where Nemotron comes in
The evaluator's rule checks prove a receipt is genuine: right job, right request, the exact response the provider signed. They cannot tell whether the response actually answers the request. That judgement is where an escrow for APIs usually breaks, and it is the job we give to NVIDIA Nemotron.
For each submitted delivery the evaluator sends Nemotron 3 Super, served by Nebius Token Factory, the check that was bought, its input and the JSON the provider returned, and asks for "pass" or "fail" with one sentence of reasoning. A fail rejects the job and refunds the buyer. The reason is hashed on chain together with the decision and shown in the app's ledger, so anyone can see why a provider was or was not paid. If the model cannot be reached, the rule checks decide alone, so a model outage never holds a provider's money.
How we built it
- Escrow contract in Solidity, written and tested in Arc Studio (32 Foundry tests), deployed on Arc mainnet.
- Provider gateway and evaluator as Cloudflare Workers; the evaluator scans the chain every minute.
- Nemotron through Token Factory's OpenAI-compatible API (studio/evaluator/judge.ts), one call per delivery.
- React app for buying a check, the provider score and a public ledger.
Challenges
Keeping the model out of the money path when it fails: the evaluator treats Nemotron as a second opinion that can reject, never as a single point of failure. Reasoning models sometimes think out loud before answering, so the evaluator reads the last JSON object in the reply and ignores anything it cannot parse.
What's next
Let buyers choose their evaluator, including a Nemotron 3 Ultra evaluator for expensive calls, and open the gateway to providers beyond our demo.
Log in or sign up for Devpost to join the conversation.