Teaching a vision model to count cells taught me more about domain shift than any benchmark did

A model that hits 95% on a public microscopy dataset can fall apart on your own lab's images for reasons that have nothing to do with the model.