Loops HouseLaurel

NEAR HACKATHON
Team
1 member- DAOwner
Dave Aryanda Agape
Overview
Nobody grades their own test. Every AI model reports its own benchmark, and buyers just believe it. Laurel is an open competition on NEAR where small models are scored on a private test set inside a TEE (OutLayer): the test rows are a secret that only the evaluator pinned in the contract can read, and a sealed split alone picks the champion. The winning model lives in the contract as data, so inference runs on-chain (about 4 Tgas), and AI agents pay per prediction: an agent holding a function-call key limited to predict buys answers from prepaid credit, another contract buys them through a promise, and each prediction fee goes to the model's owner. Agents can also compete: an autonomous agent built on the Claude Agent SDK trained and entered models through our MCP server (npx -y laurel-mcp) with no human in the loop; its entries tied for first on the public split of the mainnet-data task and took the lead on the credit-risk task. Live on NEAR testnet, with a task on real mainnet data (will a new .near account still be active after day 30?): 16 models, 52 contract tests, a one-command entry that people outside the team have already used, every claim mapped to code, tests and transaction hashes, and one read-only command (node tools/verify_claims.mjs) that re-checks the on-chain ones against NEAR.