Why would you assume that models can be ranked on a single one-dimensional scale from good to bad? I mean, they can, but only retroactively and by throwing away a lot of other useful information.