isadubois·há 3 dias·discussThis looks super clean. I'm curious about the --judge command. How does it evaluate if the cheaper model's response is a "tie" or acceptable? Is it using a specific LLM-as-a-judge prompt template?