Loops HouseRERUN

NEAR HACKATHON
Team
1 member- MAOwner
Maverick
Overview
The problem: many published results can't be reproduced, even from the authors' own code and data, and nobody checks before the results get cited, reported and used.
RERUN is a forecasting market on whether a study holds up. A published result becomes a claim: the exact code commit, data fingerprints, the paper's number and the allowed margin, hashed on NEAR. People forecast yes or no. When forecasting closes, the Lab, an AI agent, fetches the pinned code and data, checks every fingerprint and runs it in a sealed container with the internet cut off. The AI may only fix setup problems, and code guards stop it touching the analysis. Plain code compares the Lab's number with the paper's, the contract re-checks that comparison itself, and the receipt goes on NEAR. Anyone can challenge it, which makes the Lab retrace it again with a different AI. Then winners collect.
How NEAR fits in:
- The contract on NEAR holds the claims, the market (FPMM), the receipts and the payouts, and only accepts a result from a Lab whose secure-hardware attestation and code fingerprint it has checked. The Lab runs on Phala Cloud (Intel TDX), on testnet and mainnet.
- The Lab thinks with NEAR AI Cloud's private models, verifying their attestation before a run and every reply's signature after it.
- Live on mainnet with Circle's USDC. People swap NEAR for USDC on the page through NEAR Intents (1Click).
- Forecasts are gas-free in points mode: the wallet signs a NEP-366 delegate action and our relayer pays the gas.
- Any other contract can settle on our results through the public get_verdict view.
Results so far: 12 checks of 11 published papers, including the minimum-wage and CV-callback studies (same result) and the Reinhart-Rogoff debt study (different result, using its critics' code). On mainnet, a real-money claim has settled, and two more close on Nov 15.