Loops HouseProctor

NEAR HACKATHON
Links
Team
1 member- SIOwner
Sina-Akeju Daniel
Overview
AI models are becoming the scam filter for inboxes, banks and phones, but the only evidence of how well one works is its vendor's own score on a benchmark the vendor picked, often a public one that leaks into training data and that scammers can read too.
Proctor certifies AI scam filters on a sealed test nobody can see. Only the test's SHA-256 fingerprint goes on-chain. Each model answers every message inside NEAR AI Cloud's secure hardware (GPU TEEs), which signs every answer together with the exact question. A NEAR smart contract checks every Ed25519 signature against the model's key from its attestation report, recounts the score itself, issues a certificate, and pays the test's author a share. Anyone can re-check a certificate in their browser in one click, straight from NEAR RPC, with no wallet.
Live on testnet now: three real NEAR AI models certified on a 179-message sealed scam test, 3 runs each. Qwen3.8-27B scored 97.9% overall and Qwen3.6-35B 97.8%, while Qwen3-VL-30B caught 98.9% of scams but wrongly flagged 32.5% of real messages. That gap is what Proctor exists to show.
NEAR is the trust layer: NEAR AI's TEE signatures, on-chain signature checks (env::ed25519_verify), OutLayer's yield/resume for off-chain grading, and NEP-297 events that make every result verifiable from public RPC. On testnet the grader still runs outside OutLayer's enclave, reaching the contract through a stand-in for OutLayer; the README says exactly what is real and what is not yet.