Goblin and related magical mentions were overrewarded in training, and the behavior was reinforced over successive models. We removed the goblin-affine reward signal for future models, and filtered training data where creatures appeared in irrelevant contexts.