The levels would be great. I’ve tried prompting general purpose LLMs to only use words at particular levels, but they pretty quickly diverge from the prompting guardrails. One might just need to building a custom model trained only on a limited vocabulary.