We don't have a real definition of "under human control." Once a model is running multi-step tasks with tool access. Saying it's the design goal is easy. Saying we can verify it is a completely different claim. I spin up agents every day. I don't check the reasoning behind them. If it's coding work, I run it, look at the output, does it match what I asked. That's the whole check. What tools it called, what it connected to, what it searched, none of that gets reviewed. And this isn't just…