Analyzing how system instruction (SI) changes impact the response quality of Gemini, OpenAI, and Anthropic LLMs across 3648 astronomy based MMLU tests.