Metrics went from telling you something broke to deciding which GPU serves your inference request. In 20 years the pattern hasn’t changed; we act on it better.