Is AI agent observability proof of control, or just proof nothing broke yet?

There's a real difference between "nothing bad has happened" and "we can demonstrate the controls that would have prevented bad things from happening." Right now most of what I can show is closer to the former, which isn't a great position heading into any kind of regulatory conversation about AI governance. The absence-of-incident argument works...