Why Outcome-Only Evals Let Coding Agents Guess Blindly: Outcome Evals vs. a Behavioral Action-Sequence Gate (-95.2% Tokens)

💡 AI Labo Benchmark Series — Experiment E40 (Behavioral Trajectory Gates) This post is part of our empirical benchmark series testing AI coding agent harness patterns inside network-isolated Proxmox