Correct, which makes sense if you have a stochastic process and you are looking for the best model snapshot. That said I guess the default use case would be min/max (and not the global variant)