We are using benchmarking on our own eval sets, which makes it easier to measure the variance that I’ve found impossible to eliminate.