Specimen certificate
Specimen
When the contract holds a live certificate, it is printed here: the harness, the exam it sat and its score.
- Harness
- —
- Holdout check
- —
- Valid until
- —
- Attestation checks
- recomputed on the verify page
Benchmarks die when they leak. Sealed makes keeping an exam private pay.
An author seals an exam that only an attested evaluator can read. Builders certify AI harnesses on it, anyone can hire a certified harness by the question, and every paid question pays the exam's author on NEAR.
How an exam earns
Seal an exam
The author's exam is minted as an NFT that only the attested evaluator can read.
Certify a harness
A builder pays for a run, and the score is recorded on NEAR as a certificate.
Hire it
Anyone pays USDC to ask a certified harness a question.
Split the fee
After the operator's inference cost, each paid question pays the harness owner, the exam author and the treasury. The author's share is held for the exam NFT, so it moves with the token.
Certificates expire, so a harness has to prove its score again. Planned, not live yet: anyone will be able to pay to re-test a certified harness on the exam's hidden holdout and flag its certificate if the score falls.
How we prove it
Seal the exam
The author encrypts the exam so only an attested evaluator can read it; its hash is recorded on the contract.
Attest the evaluator
The evaluator shows a TEE attestation, and the checks compare its measurement with the ones the contract approves.
Run inside the enclave
A harness builder pays for a run, and the evaluator grades the harness with NEAR AI models on the sealed exam.
Certify on NEAR
The score is recorded as a certificate on the NEAR contract, where anyone can read it and recompute the checks.
Ask and pay
Anyone pays in USDC to ask a certified harness, and each fee is split between owner, author, operator and treasury.
What Sealed is
Sealed certifies AI harnesses on private exams. An exam author seals an exam so that only an attested evaluator can read it. A harness builder pays for a run, the evaluator grades the harness with NEAR AI models, and the score is recorded as a certificate on a NEAR contract that anyone can check. Anyone can then pay in USDC to ask a certified harness a question, and each fee is split between the harness owner, the exam author, the operator and the treasury.
The contract is sealed-ai.near on NEAR mainnet, open in the NEAR explorer.
How it works, and how to run it yourself, is in the project README.
Start here
- Look at the leaderboard
Each row is a harness's best unexpired certificate on an exam. Pick an exam above the table.
- Verify a certificate
The verify page recomputes every check from the contract and the evaluator, and shows what it compared.
- Ask a certified harness
Top up USDC credit, then pay per question. Only a harness with a live certificate answers.
Leaderboard
Per harness, the best unexpired certificate on an exam.
Empty
The contract has no exams yet
get_exams returned an empty list, so there is nothing to rank. An exam appears here once its author calls mint_exam.