Pre-commit gates lift self-evolving agents to 72%

Verifier-as-Gatekeeper reached 72% pass@1 on Terminal-Bench 2 with a skill pool roughly 5x smaller than unconditional accumulation, according to the authors’ new preprint. Its three critics screen each learned skill for structural validity, behavioral harmlessness, and semantic consistency before admission; the study also found that removing a defective skill later recovered only a small part of the damage inherited by descendant skills. For operators building the kind of runtime policy layer described in HiddenLayer’s agent-harness architecture, the fact to file away is that learned skills need a pre-commit gate, not merely a rollback button.