Loops HouseBlindBench

NEAR HACKATHON
Links
Team
1 member- JOOwner
John Kamal
Overview
BlindBench is a marketplace for private AI evals. An expert commits a secret test set to NEAR as a Merkle root, so the questions are fixed but never published. Anyone pays the set's price in NEAR to have a model scored on it, and the page shows a live NEAR Intents quote for paying from another chain. The questions go only to NEAR AI Cloud enclaves, every answer is signed by an enclave key whose Intel TDX quote is checked down to Intel's root, and the score lands onchain as a certificate naming the model, the set's root, the judge and the evidence hash. The evidence bundle holds hashes and signatures only, so anyone can recheck a certificate in the browser or with the CLI without seeing a question. The contract pays the set's author on every run, and a bounty pays the first enclave model to clear a bar.
It is private AI benchmarking with results certified against the model and a hash of the test set, with a market built around it.
Live on NEAR testnet (blindbench-hnxot5.testnet): five private sets (on-chain tracing of real 2026 exploits, legal review of original contracts with decoys, a smart contract audit of three real codebases, spec-exact coding against hidden tests, credit analysis of an original annual report), 21 certificates across five enclave models, one bounty paid out by the contract and one open. On-demand public runs are paused in this prototype: the certificates come from the team's own paid runs. On the hardest set the best enclave model scores 36.1% where a closed frontier reference scores 79.2%, so the sets separate models. The README has do X, see Y steps: an offline one-command demo, 52 CLI tests, 19 contract tests, live contract reads, certificate verification and a NEAR Intents dry quote.