Most teams still evaluate model or prompt changes by reading a handful of outputs and deciding subjectively whether it improved. That's not a rigorous standard for a decision that affects cost, latency, and correctness at scale, and it wouldn't be accepted for any other kind of comparison in a serio
