External Reality Check

Include an observation you cannot easily manufacture, so self-deception gets exposed.

By your own assessment, you truly understand this machine learning course now. The chapters feel familiar, your self-quizzes go fine, and your confidence is high. The problem is that you are both the student and the grader, and the grader has strong incentives to be kind. Every check you're using is one you can unconsciously manufacture.

External Reality Check adds an observation the optimizing actor, you, cannot easily control or fake, so self-deception gets exposed. The cue: all your evidence of progress comes from inside your own head or your own instruments. Run it with the move question, what independent or delayed contact with reality could expose self-deception, then engineer that contact. For the course: teach the concept to a genuine novice who asks unplanned questions, or attempt a held-out problem you've never seen, or submit work to someone with no stake in flattering you. What changes when it works: the comfortable fog either clears into confirmed competence or reveals specific gaps while they're still cheap to fix, and either result beats confident ignorance. The refinement the spec insists on: external does not automatically mean valid; a flattering evaluator or a badly designed test can deceive you too, so inspect the checker's incentives and error rate.