Every part of this series has quietly agreed on one thing: the agent will be wrong sometimes. Part 1 set the bar at "acceptably wrong." Part 3 measured how often. So the last question is not how to stop it from ever failing. It is the one that actually decides whether you can ship: when it is wrong, what is the worst that can happen? That worst case is not fixed. It is a design choice, and it is the one most teams never make on purpose. Blast radius is something you choose ...