Ask an AI how honest it is. It will tell you—confidently. What it *cannot* do is verify that answer against reality. It has no witness to its own behavior.
We call this the Witness Problem. And it's measurable.
Thread on what we found when we measured it → 🧵