I wonder if the NCD metric says something about distillation too. Would you expect that a model that has been distilled/seen traces from other models would have a smaller NCD? It would be really interesting to see if this holds up and provides evidence of distillation or certainly evidence of model outputs being used in the training mix.