Looping Is Not Reliability
Why more agent revisions can discover more correct patches while delivering fewer of them—and how state-scoped evidence, read-only observation, and explicit admission boundaries turn a mutable coding trajectory into an explainable delivery chain.