This is a wonderful writeup, describing in great detail how to measure performance in a way that actually properly accounts for confounding variables. This degree of scientific rigor is so, so very rare.