I’d love to see, based on the level of non-determinism perfomance on the benchmark how many times you need to run the benchmark for the change to be relevant (or statistically significant if you want).

That would be a nice paper.