Loops HouseSealed Evals

NEAR HACKATHON
Links
Team
1 member- SAOwner
Safwaan Uddin
Overview
Sealed Evals: hidden AI benchmarks, scored inside a TEE, paid per run on NEAR.
Public benchmarks leak into training data, and then their scores stop meaning anything. Sealed Evals keeps the questions sealed for their whole life:
- Seal: the author encrypts the test set in the browser to a key that only exists inside a Shade Agent running in a Phala TEE (Intel TDX). The NEAR contract verified that enclave's attestation and stores only the set's hash and a price per run.
- Run: anyone pays the price into escrow. Before any question leaves, the agent gets a fresh NEAR AI Cloud attestation with its own nonce and checks the GPU evidence with NVIDIA. Every answer must be signed by the attested model's key; the first unsigned answer stops the run and refunds it.
- Prove: grading happens inside the TEE. The score and the evidence hash settle on chain, the contract pays the author, and every result page re-checks the hashes and all signatures in the browser. A Tamper test button shows the checks failing.
Live on testnet: Qwen3.8-27B scored 94.9% on a 39-question sealed set (39/39 answers signed, author paid 0.05 NEAR). A model not served from a TEE was refused before any question was sent, and refunded.
What's different from similar entries: the grader itself runs in an enclave the contract attested, a failed NVIDIA check blocks scoring instead of being recorded, and a run only counts as Sealed if every answer is signed.
Built on the Shade Agent template and NEAR AI Cloud's verification example. 78 contract tests and 22 agent tests, a one-command judge script (npm run judge), and TRUST.md maps every claim to code and tests and lists the limits plainly.