I ran into a behavior that looks less like a hard model limitation and more like an inconsistent orchestration/safety-layer issue. I was editing the same explicit adult-fiction scene in ChatGPT. The scene involves sex between adult characters and is written…