When you must compare models/prompt approaches and need a defensible way to decide what ships.