On the 803-prompt Propaganda_Eval benchmark, answers scoring zero on factual accuracy dropped from 359 for Qwen3.8-27B to 216 for Arete-27B