Different routing thresholds should be set for different intents; there is no globally optimal threshold—intents with high fallback rates should be handled more conservatively, while intents where the small model performs stably can be handled more aggressively.