Final controlled batch · Figure 01

Controlled EVMbench Detect — final academic comparison

GPT-5.6 Sol High and Grok 4.6 Build High across 72 controlled smart-contract audit runs on 12 audits, with reconstructed quality, efficiency, cost, and uncertainty.

GPT-5.6 Sol High and Grok 4.6 Build High across 72 controlled smart-contract audit runs on 12 audits, with reconstructed quality, efficiency, cost, and uncertainty.
Codex-reconstructed manual scoring, not official EVMbench grading. The PoolTogether development pilot is excluded.