Large language models (LLMs) can analyze complex health data to generate personalized responses. Evaluating these LLM responses is crucial for accuracy and safety, but current human-expert evaluation is costly and not scalable. This paper introduces a new framework for evaluating health LLMs using Adaptive Precise Boolean rubrics. These rubrics break down complex questions into granular, Yes/No criteria to improve consistency and efficiency. The framework was tested in metabolic health and demonstrated significantly higher inter-rater reliability than traditional Likert scales. Adaptive Precise Boolean rubrics also reduced evaluation time by over 50%. This method proved more sensitive to variations in response quality compared to Likert scales. Automating the rubric filtering process with a zero-shot classifier maintained similar evaluation improvements. The framework reliably detected quality drops in LLM responses when real participant data was altered. The proposed approach offers a scalable and streamlined method for LLM evaluation in specialized domains.