LLMs can work deeper in the stack than most people think. The investment required is in the box you put them in. Both humans and tools are flawed, so you create layers of controls with non-overlapping seams so no single flaw permeates everything. The higher the stakes, the better the cage needs to be.