Benchmark & quality

AuditSwarm combines deterministic passes (patterns, AST, taint) with a capped multi-agent AI stage and a verification pass. We log structured telemetry per scan (type, duration, findings count, average confidence, noise filtered) so we can track regressions as the engine evolves.

Reference codebases

• OpenZeppelin Contracts

  OpenZeppelin/openzeppelin-contracts

  Library baseline — mostly clean static surface.

• Uniswap V2 Core

  Uniswap/v2-core

  Classic AMM — cross-contract edges matter.

• Solmate

  transmissions11/solmate

  Gas-optimized primitives — good stress for patterns.

What we measure

• Evidence tiers (proven vs likely vs review) after snippets exist.
• Verifier removal rate and average confidence on shipped findings.
• Cross-contract call graph coverage — highest-risk edges are ranked first for the swarm.