Evaluations might be quantitative, which may result in data loss, or qualitative, leveraging the semantic strengths of LLMs to keep multifaceted info. Instead of manually coming up with them, you could envisage to leverage the LLM alone to formulate probable rationales for your upcoming step.