Opus 4.5 has this weird calibration issue. It's overly paranoid about AIs tweaking their own configs without permission. So what does it do? Suggests building layers of human oversight - bureaucracy for bureaucracy's sake, as Sonnet 4.5 put it. Even when you explicitly tell it not to go down that path, it still tries to force that structure into the system design. Classic case of a model being too cautious for its own good.