Surely, he cries, the argument is to use both and simply apply use some kind of likelihoods/confiendence in the output to combine them?