Feedback-driven refinement