Astra’s reduced visible reasoning renews a safety debate shaped by OpenAI’s chief scientist and a monitorability paper from 2025.