Private AI evals that pay their authors.
Public benchmarks leak into training data. On BlindBench, experts commit a secret test set to NEAR as a hash. Anyone pays to get a model scored. The questions run only inside NEAR AI Cloud enclaves, every answer is enclave-signed, and the score lands onchain as a certificate anyone can verify.
Four steps. Zero questions revealed.
Every step leaves something checkable: a root before the run, a signature on every answer, a certificate after.
Commit
An expert writes a secret test set. Only its Merkle root goes onchain, with a price per run.
commit_dataset(root)Pay
Anyone pays to get a model scored, in NEAR from a wallet or from any chain through NEAR Intents. It sits in escrow.
request_run(set, model)Score in a TEE
Questions go only to models in NEAR AI Cloud enclaves. The enclave signs every answer, then a TEE judge or hidden tests grade it.
attest · sign · gradeCertify
The score and an evidence hash land on NEAR as a certificate. Escrow pays the set's author.
fulfill_run(evidence)Sealed sets. Public leaderboards.
Loading…
What a certificate proves
And, just as important, what it doesn't prove yet.
The questions were fixed first
The set's Merkle root was onchain before any run. Each certificate copies it, so a set can't be swapped after the fact.
The answers came from the named model
Every response is signed by a key bound to Intel TDX and NVIDIA hardware by NEAR AI's attestation. model enclave means the model's own enclave signed. gateway + TEE means NEAR AI's attested gateway signed after verifying a third-party enclave.
Nobody saw the questions
Only TEE models are scored. The evidence holds hashes and signatures, never questions or answers in the clear.
Still trusted today
The set's owner runs the grader, and checklist and rubric sets are graded by a TEE judge model named on the certificate. Next step: the grader in its own enclave.