Loops HouseNEAR Replay Lab

NEAR HACKATHON
Links
Team
1 member- MOOwner
Moses Fawole
Overview
An approved 0.001 NEAR obligation settled as 0.002 NEAR in our unsafe fixture and 0.001 NEAR in the guarded fixture. NEAR Replay Lab makes that difference inspectable: replay the failure, open the FINAL receipts, and see exactly who received what.
We test the execution boundary for payment-capable AI integrations. A local TypeScript runner sends constructed requests to two disposable Rust contracts on native NEAR testnet. The five cases are legitimate payment, expired approval, recipient substitution, retry after a discarded broadcast response, and concurrent duplicate requests from independent keys. Both vaults share authorization, allowlist, cap and expiry checks. The guarded vault also binds the exact approved recipient and amount and claims Pending before transfer.
The evidence oracle checks linked successful FINAL receipts, canonical approvals and exact recipient balance changes. The read-only Next.js dashboard recalculates each verdict as HOLDS, VIOLATED or INCONCLUSIVE; a green function call alone does not prove payment. A repeated signed hash counts once. Our duplicate cases use fresh signed requests for the same invoice.
Three complete comparisons were accepted from ten actual trials: 30 unique obligations and 72 distinct FINAL transactions, including owner approvals. Seven infrastructure-limited trials remain inconclusive and are disclosed. 197 Node tests and 74 Rust SDK-host tests passed; clean reproduction and dashboard checks passed.
The experiment uses constructed inputs; no live model or real vendor vulnerability was evaluated. An unknown callback outcome leaves the guarded obligation locked. Production safety and adoption are unproven. Built by solo developer @mosesfawole with Codex assistance.