For things like #9 and #14, it'd be nice to have an easily reproducible benchmark to permit pre/post testing.
