Loops HouseSealed Bench

NEAR HACKATHON
Links
Team
1 member- JOOwner
Joseph Azulay
Overview
Public AI benchmarks leak into training data, so their scores stop meaning much. A private test set stops the leak, but then outsiders can only trust the lab's number.
Sealed Bench keeps the test set private and lets anyone spot-check the score on NEAR. The owner commits a salted Merkle root of 100 problems to a NEAR testnet contract, runs the model through NEAR AI Cloud's TEE, and commits a Merkle root of the signed responses and grades. Any account can then call the draw, and the contract's random_seed picks 20 of the 100 entries. The owner must reveal those 20 with Merkle proofs and the TEE signatures. The contract rebuilds each request, recovers the signer with ecrecover, checks it against the signers pinned at start_run, and re-grades the answer. One failed check marks the run FAILED and pays the owner's bond to the drawer. A draw of 20 catches a run with 10% bad entries about 90% of the time.
Real runs: 8 NEAR AI TEE runs (Qwen3.8-27B and GLM-5.3-flash on AIME-style and GSM8K-style sets), all AUDITED at 20/20, with every draw called by a non-owner account. The pre-registered AIME-style test found no contamination gap: Qwen -11 pts [-29, +5], GLM +11 pts [-5, +26]. A cheating mock run was caught and marked FAILED on chain. The live site re-checks drawn entries in the browser over public RPC, and npm run verify re-checks all 8 runs from a fresh clone with no NEAR account or API key.