Role embeddings: making authorship more salient to LLMs — AI Alignment Forum