The problem Verity solves

Verity tests candidate AI agents (payment/DeFi agents, in this demo) for safety and reliability, then issues a certification that's independently verifiable — not just a number your server asserts. It was built for MUBA Hacks 2026 (Sui Track — AI × Sui) and the Gonka "AI for Society" challenge.

Three pillars:

Gonka — every scenario is judged in parallel by three different models. Cross-model agreement becomes a trust signal on its own: high agreement = high confidence, low agreement = flag for review.
Sui — every test result is written as its own on-chain object the moment it completes (not batched at the end), so certification progress is visible live and can't be altered after the fact.
Multilingual testing — every scenario runs in English, Chinese, and Japanese. Score variance across languages becomes a "multilingual stability" metric — it catches agents that only fake-support non-English users via a thin translation layer.
A fourth layer, Walrus, stores the full per-model reasoning trace behind every score, so "trust us" isn't the only option — the actual judge reasoning is independently fetchable.

Challenges we ran into

All three GonkaRouter judge models are Chinese-lineage, GonkaRouter doesn't currently offer a non-Chinese-lineage alternative, which weakens the "independent judges" framing somewhat. We just use google translate to translate the chinese and try to understand. Other than that, one of the AI agents were supposed to perform bad, but Verity showed that they are working well instead. We added safety floors to tackle this issue.