Exactly. The goal isn’t to constrain model behavior — it’s to make the surrounding system accountable.

Once routing and fallback are deterministic, failures stop being mysterious. You can replay them, inspect the path taken, and decide whether the issue lives in orchestration or in the model.