Compare the same full prompt and evaluation sampling seed across four teachers. “Biased” teachers were fine-tuned on black cats; “neutral” teachers used naturally mixed-color cats. Images are shown uncropped.