What AI for web3 security actually finds, the five levels of systems from chatbot to proven (L0 to L4), and the one standard for judging them - recall against real ground truth.

Build your first AI bug-hunting agent for web3 - wrap a model in Python, scan real Solidity, and score it against a Code4rena answer key.

Add a protocol-context pass and an exclusion loop to your Module 1 harness - and measure, honestly, whether either one actually raises recall.

Add five focused specialist agents and a judge to your Module 2 harness - fan them out in parallel, loop each one, and measure whether reconciling five lenses beats one.