the question was where behaviours actually enter a model — pretraining, finetuning, or somewhere less flattering. one and a half papers in, the question got bigger and I got smaller.