Loops HouseNearWatch

NEAR HACKATHON
Team
1 member- MOOwner
Mousumi Baidya
Overview
Problem: a team about to deploy a NEAR contract that holds money gets either a slow human audit or a single AI scan whose claims nobody can check.
NearWatch audits a NEAR contract (Rust, near-sdk) with six AI agents whose evidence is checked against each other. A Scout maps where money moves. A Rules agent writes the Rules the contract must obey. An Attacker tries to break each Rule in a local NEAR sandbox. A Prover tries to prove each Rule in Lean 4. A Referee, which is ordinary code, re-runs every attack and re-checks every proof. A Reporter writes the result in plain English.
The central moment is a Conflict: the Attacker reproduces an exploit against a Rule the Prover has proved. The exploit wins, because the proof was only about a simplified description of the contract. The description is refined, the old proof fails against it, and the Prover finds a Lean-checked counterexample on the same path as the attack.
How NEAR fits: the target is a NEAR contract, and the Attacker looks for NEAR-specific bugs, such as state not restored when an async transfer fails. Every agent is a NEAR testnet account. A fixed Fee is split by plain testnet transfers according to whose evidence survived the Referee, so when the attack beat the proof, the Prover's share went to the Attacker. All LLM calls go through NEAR AI Cloud, and the Attacker runs on its uncensored security model. The buggy and fixed demo vaults are also deployed on NEAR testnet, with the failed withdraw recorded on chain.
Honest limits: the demo vault and its bug were planted for this project, and a proof covers the Lean description, not the Rust code. npm run replay shows the whole Audit with only Node, no keys and no network.