A coding agent that hits a gap in your request does not stop — it picks a reading and keeps going, and the guess gets built on. Asking about every gap is worse: the interruptions cost more than the bugs. This paper decides which questions are worth your attention by running them: write the two plausible answers out as rival versions of the requirement, generate three programs for each, run both sets on at least seven shared inputs, and keep the question only when the two sides give a different stable answer somewhere. Across four coding agents that scored 41.2 at matching the clarifications a human had marked necessary, against 27.3 for the best existing method.