LLMs need constraint feedback environments that nudge them toward correctness. The output they generate only has some probabilistic chance of being right. Our job is to build a box around them that catches and corrects the drift. The tighter the box, the deeper in the stack you can use them.