When you find AI still can't fix a bug no matter what
stop the endless patching
The underlying design is probably wrong
Including: model design, data flow, state management, concurrency and resource control
So the point of verification isn't to catch simple isolated interaction bugs
It's to iterate toward a stable, explainable system architecture