I’m comparing DeepSeek answers for a set of technical questions and notice that small wording changes can lead to noticeably different responses. For a repeatable evaluation, how much variation should I expect when I submit the same prompt several times?

Have you found settings or testing practices that make comparisons fair without hiding the variation an ordinary user might see?