this is impressive. genuinely better than most people appraoches with using LLM as another judge to help route. which just uses more tokens than saves