Run a grid of configurations over a judged dataset and compare means with confidence intervals, not bare numbers.