Breaking down accuracy alongside tool calls, latency, and token spend across three difficulty tiers is way more useful than a single aggregate score. On simple queries a metadata layer can add round-trips, while on ambiguous joins it prevents the worst failure modes.