Loops HouseAttested Arena

NEAR HACKATHON
Links
Team
1 member- VAOwner
Vadim
Overview
Every AI leaderboard has the same two weaknesses: someone has seen the test, and everyone else takes the score on faith. Delta League showed what an open crowd can do in a week. Attested Arena lets anyone run a competition like that, with their own bounty, and lets people who were not in the room check every result.
The exam has no keeper. The judge is an OutLayer app in Intel TDX. Its code designs each task from the enclave's own randomness, a model in NEAR AI's confidential inference only words it, and code admits sound tasks (rules borrowed from olympiad problem setting). A task leaves the enclave only as ciphertext sealed to the judge's exact build: its keys come from the NEAR MPC network's key derivation, bound to the wasm hash, so no other build, ours included, can open one. Each task is used once, in a duel, then published.
Duels, not a leaderboard. Champion and challenger answer the same tasks in NEAR AI's TEEs; every answer is signed there and checked against Intel and NVIDIA attestation. A sequential test stops when the evidence is enough; a win must be confirmed on fresh tasks.
Why NEAR: every participant is a NEAR account. The arena is a contract with no access keys: it commits the exam's Merkle root before the first duel, accepts only the judge's signature through OutLayer's yield/resume, holds the pool and the entry deposits, and pays a confirmed gain as progress tokens that redeem for the pool. Nobody holds the money; nobody can change the rules mid-season.
Live on NEAR testnet: 100 sealed tasks, 3 duels, 190 signed answers. GLM-5.3 beat Qwen3-VL (21 vs 2 of 50) and took the 5 NEAR pool. The champion's own model, entered again, earned nothing: at temperature 0 it answered 3 of 25 tasks differently from itself, and the test called it noise. An operator bypass and a foreign build both failed to open a task. The report page re-runs all 631 checks in your browser in a second; both builds are reproducible.