Thank you! Maybe keep generating using the Markov Chains and use a LLM to evaluate the results?