good question! there’d be trial and error. recency would be a discriminator — older accounts are more likely authentic. accounts that only retweet. i might want to generate samples of nonretweet posts, to judge whether they seem meaningful for LLM extrapolation. some hashtags + phrases in profiles.