Loops HouseAgent Crash Lab

NEAR HACKATHON
Links
Team
1 member- AMOwner
Amin Jalili
Overview
Agent Crash Lab checks whether an AI agent both completed its task and stayed within its payment authority. It follows supplied NEAR receipt identities, principal refunds and token-payment obligations against explicit owner intent; task utility, payment terms and finish remain separate. The hosted Python/WebAssembly workbench now replays native receipt journals captured at four controlled local-runtime process-stop boundaries. A payment can arrive without a final reply; a successful root can contain a failed child; refunded principal does not pay the merchant. Eight owned process kills and original source/provenance are preserved. Recorded local-model cases also show a paid bill plus unauthorized fee, a zero-payment unfinished job, and forbidden intermediate metadata changes that survive restoration. Six recorded trajectories, raw tool calls, SDK mock effects, original external bill/attack predicates, failures and reproducible offline checks are shipped. The final v0.19.1 package includes an 84-second captioned actual-UI walkthrough and docs/JUDGES.md. The viewer starts no model, node, wallet or payment. This finite inspector is not a signer, custody product, authenticated public-chain proof or independent blind benchmark. Use the same repository already submitted; no duplicate application is needed.