it’s possible that the grok model was trained or fine tuned somehow to help with moderation. in that case, it’s possible that things like these bullet points are somewhere up it’s context chain, or in its training data in a manner that it can relatively accurately recall