Loops HouseBakeoff

NEAR HACKATHON
Links
Team
1 member- SUOwner
Success Aje
Overview
Teams know their AI workflows could be cheaper, faster or more accurate, but finding an improvement they can trust is hard. Public benchmarks can leak into training data, vendors may grade their own work, and a team’s real test data can be too sensitive to share.
Bakeoff turns that problem into a sealed, paid challenge. A team sets its baseline and requirements for accuracy, cost, latency and data leakage, commits to a hidden test set, and funds a reward. Builders or AI agents use public practice tickets to develop a configuration, then submit it for evaluation. An OutLayer trusted execution environment scores submissions against the sealed test set, using NEAR AI Cloud for model calls, and returns totals and hashes rather than the test data.
NEAR makes the challenge and payment verifiable. A smart contract holds the reward, commits to the rules, evaluator and test set by their hashes, and accepts results through the OutLayer callback for its own request. It awards the earliest qualifying submission. Once opened, the challenge’s commitments are frozen and its account keys are deleted. The board lets anyone inspect the on-chain commitments, and NEAR Intents supports funding from other chains.
We demonstrated the full flow on mainnet using a fictional store and accounts we controlled on both sides. A Claude Code agent passed the support-ticket challenge on its first submission at 60% lower cost than the baseline, and the contract paid the 1 NEAR reward.