Metrics went from telling you something broke to deciding which GPU handles your inference request. The pattern hasn’t changed in 20 years, we just got better at acting on the data.