A result is reproducible if someone else can follow the same method and get about the same score. That requires more than the name of the benchmark. It needs the model and version, the exact prompt, the settings such as temperature and reasoning effort, the tools and environment, the number of attempts and how answers were graded.
AI evaluations are hard to reproduce for several reasons. Model output varies from run to run, providers update models without notice, and hosted services change over time. Small differences in prompts or scoring scripts can move a score by several points.
Reproducibility is also what lets other people find mistakes. A published method with the code and data allows others to check the result, and a vague one does not.
When you run your own evaluations, record every setting, fix the versions where you can, run the test more than once and report the spread as well as the average.