I think the problem is not just giving hypotheses names, but giving large collections of hypotheses names. It bundles them together so that the strongest hypotheses in the group can defend the weakest ones, or the weakest ones can damage the strongest ones, even if the different hypotheses aren't actually related in a technical sense.

Dividing hypotheses into "NLP" and "not-NLP" is an attempt to carve hypothesis-space at its natural joints, and therefore needs to be justified by clear shared dependencies among those hypotheses.