Loops HouseSealed

NEAR HACKATHON
Team
1 member- RIOwner
Rifat
Overview
AI benchmarks go stale fast. Once the questions are public they end up in training data, and a high score stops telling you much. So whoever writes a good test set has two bad choices: keep it private and earn nothing, or publish it and watch it get gamed.
Sealed is our answer. You write an exam in the browser and it's encrypted before it leaves your machine. After that nobody reads the questions. Not us, not the model being tested. Grading runs inside an Intel TDX enclave on Phala: it opens the exam, runs the model, scores it and signs the result. The signed score goes on NEAR as a certificate, and anyone can re-check it on the site or with make verify.
NEAR is where ownership and money live. The contract, sealed-ai.near, is on mainnet. It takes USDC and pays out on its own. Every exam and every harness (a model plus its prompt) is an NFT, and the income follows whoever holds it. A 1 USDC certification run pays 70% to the exam's owner. When someone asks a certified agent a question, 70% goes to the harness and 20% to the exam behind it. Sell the NFT and the income moves with it.
It's live, not a mock. Three models took the same sealed exam on mainnet: Qwen 90, GLM 85, DeepSeek 80 (certificates 4, 1 and 5). Every payment was real USDC. To be upfront, we wrote that exam and paid for those runs ourselves, so the next step is getting outside authors in.
Want to try it? sealed-ai.xyz gives judges 3 free questions to a certified agent, no wallet needed. The repo has the contract, the evaluator and the site, with the test suites we ran.