The problem CreatorPay solves

Pay only when signed data proves the work is done.

Today, paying for an outcome means trusting someone's word afterwards. A brand pays a creator and argues about screenshots. A buyer hires an AI agent and has no neutral way to check the job. Creators wait for invoices, and brands pay for missed targets.

Pay-on-Proof locks the money in a Cardano smart contract before the work starts. The contract pays the seller only if a signed answer from an API3 Airnode satisfies a rule agreed in advance, for example "this post has at least 300,000 likes" or "ETH/USD is above 2,000". If the proof never arrives, the buyer is refunded after the deadline.

What people can use it for

• Creators and agencies: the funds are visible on chain before the post goes live, and they are paid out automatically when the target is met. Payments can be split into tranches, for example one third after the campaign and two thirds if the post is still up a week later.
• AI agents: a buyer pays an agent only for results that can be measured outside the agent.
• Any measurable condition: prices, rainfall, likes, views. Any registered AirnodeHub source works with the same contract.

How it makes this easier and safer

• No invoice and no chasing. The keeper finds the escrow, fetches the signed data and releases the payment without anyone sending a request.
• No trusted middleman. The Cardano contract checks the API3 signature itself, so a verifier or a person does not decide.
• Tampering is refused. An edited value, a stale response, an answer to a different question or a foreign signer is rejected by the validator. Anyone can test this at https://pay-on-proof-judge.vercel.app.
• A seller cannot sneak in different terms. The buyer's client recomputes the whole policy and refuses to pay on any mismatch.
• The buyer is protected. If the target is missed, the money returns after the deadline.

Honest limits: a signature proves what the data provider said, not that it is true. This is a testnet prototype, not externally audited, and it has no customer yet.

Challenges we ran into

1. Our own attack found a flaw in our first contract.
An attack matrix showed one transaction could satisfy several escrows at once (double satisfaction). We fixed the validator so a release must spend exactly one input of this contract and pay the beneficiary in full, re-ran the same attack (now refused), and kept the old version on chain and in the repo as evidence. 78 Aiken tests now pass in CI.

2. No AirnodeHub signature check existed on Cardano.
We had to verify an Ethereum-style signature (secp256k1 over an EIP-191 hash) inside a Plutus V3 validator, and rebuild the signed text from numeric slots so an edited number breaks the signature. We built it, covered it with the four tamper cases API3 asks for, and put a public judge site online where anyone can try to forge a response.

3. The x402 facilitator does not check what is in the payment's datum.
A dishonest seller could offer a different rule, a different refund party or a longer deadline. We made the buyer's client recompute the whole policy from its own request and refuse on any mismatch, then tested 54 lying price lists (all refused, honest ones accepted).

4. Waiting for the chain.
A release needs the ledger's time to catch up with the signing time, so a like-to-payout takes about 53 s (about 47 s of it is the chain). We measured it, designed the demo around it, and added a faster keeper mode for booths.

5. Honest documentation.
Independent audits flagged claims that were too strong (for example "no key on the host" and the refund wording). We corrected the docs and tests instead of defending them.