ChatGPT once made up six court cases, and two New York lawyers filed them with a real judge. They got fined $5,000. The fix for that kind of mess turned out to be almost dumb. You make a second, cheaper AI check the first one's homework before anything ships. Picture a teacher grading with the answer key. She can mark a calculus exam she couldn't have aced herself, because checking an answer is way easier than coming up with one. AI works the same way. The expensive model has to dig through…