We used to use the "Lowest score" for classification (which is the
score minus the error). We don't use it anymore, because we will drop
the elements that don't have a significant improvement over the
baseline.

There is probably a way to take the error of the measurements into
account but I'm thinking this is not a sound way to do it.