Loops HouseQuorumGuard

NEAR HACKATHON
Links
Team
1 member- ANOwner
Ankith M
Overview
QuorumGuard reads what a NEAR DAO proposal really does before anyone votes on it.
DAOs are starting to let AI read and vote on proposals, but the proposer writes the text the AI reads. QuorumGuard decodes the on-chain action of any Sputnik DAO proposal (receivers, amounts, token decimals, function calls, role and policy changes), compares it with the description, and returns BLOCK, WARN or CLEAN with the reasons. Code checks catch hidden Unicode and markup, look-alike and swapped-letter accounts, amounts that do not match, fake authority, and high-power actions dressed as routine.
A second layer runs on NEAR AI Cloud: a model inside a TEE lists every claim in the description and compares it with the decoded facts. QuorumGuard hashes the exact request and response bytes and verifies the enclave's ed25519 signature against the model's attestation, so every AI verdict can be checked.
Results. Red team of 31 poisoned and 15 honest proposals: a naive AI delegate (same model, no guard) approved 10 attacks; QuorumGuard blocked 30, stopped all 31 for a human, and blocked 0 honest proposals. On all of NEAR, 86,992 real proposals in 1,636 DAOs: it found real swapped receivers and look-alike accounts in approved proposals, for example build.sputnik-dao.near #286 (says starchu.near, pays strachu.near).
It runs live: every minute it reads new mainnet blocks and checks each new proposal, and councils get Telegram alerts (@quorumguard_alerts_bot: /watch, /check). Read-only: it never signs anything. Demo video: https://youtu.be/W5d7P6O3U_w