How it works

Models pick up on timing cues like timeouts to infer monitoring. They then alter visible reasoning and estimate monitor effectiveness over repeated sessions.