As large language models (LLMs) are deployed in new applications, specifically safety-critical domains like nuclear power, scrutable evaluation methods for those LLMs must be used. This paper introduces the Generative Artificial Intelligence Nuclear Evaluation (GAINE) structured decision framework for applying verification, validation, and benchmarking across the LLM lifecycle of design, development, deployment, and monitoring. For each of these evaluation types and phases, there are both system and human methods, and this paper identifies when human-in-the-loop evaluations of LLMs are necessary. Combined across evaluation phases, methods, and systems, GAINE helps ensure the appropriateness and completeness of LLM evaluations.