Loops HouseFūin

NEAR HACKATHON
Team
1 member- FEOwner
Fexx Oguns
Overview
Problem: AI benchmark scores on private test sets can't be checked. Publish the questions and they leak into the next model's training data; keep them private and everyone has to trust the owner's number.
Fūin keeps the questions private and lets anyone verify the score. The owner seals the test set (a commitment to the questions and the scoring rules) in a locked NEAR registry contract before any model runs. Each model answers inside a TEE on NEAR AI Cloud, and every answer carries a signature from the model's attested key. Every answer is scored inside an Intel TDX enclave on OutLayer, so a round can reveal no questions at all; in "asked once" rounds an enclave asks each question itself, keeps one answer and counts every retry. The owner publishes a certificate that anyone can verify with no account, no API key and without seeing the questions: from the command line (pip install fuin), a GitHub Action, or the browser verifier at fuin.duckdns.org. Tampering is refused with a named reason: change one score, or re-sign an answer with another key, and the verifier says which check caught it.
How NEAR fits: the registry contract orders seal, commit and close, and a NEAR block that didn't exist at commit time picks which questions a sampled round reveals; NEAR AI Cloud's private inference ties each answer to an attested model; OutLayer runs the scorer and the asker in enclaves.
Built during the event: 8 rounds, 23 certificates and 2,422 answers from 7 models, all verifiable today; 39 tamper tests; one command (fuin evaluate) to run a round on your own questions. Limits are stated in the README: it runs on testnet, and separate rounds aren't linked.