Classifier languages are typically described as not allowing direct numeral modification with any noun (Chierchia 2010, i.a.). Tiwa (Tibeto-Burman; India) provides a counterexample to this generalization — the vast majority of nouns (regardless of mass/count distinctions) require numeral classifiers, but a small set of inherently plural kinship terms combine directly with numerals. Building on Krifka 1989, I provide an analysis that captures the distribution of Tiwa classifiers, numerals and nouns on which basic numerals denote degrees, classifiers allow the numeral to combine with <e,t> nouns, while certain nominal expressions take both a degree and individual argument (i.e. are <d,<e,t>>). Given the existence of these two types of nouns, I argue that the presence or absence of numeral classifiers, at least in Tiwa, is best characterized as mediating the relationship between noun and numeral, rather than being strictly 'for numerals' despite forming a constituent with them (cf. Krifka 1995, Bale and Coon 2014, Little et al. 2021). The Tiwa data also support Bale and Coon's (2014) conjecture that the distribution of numeral classifiers is not (solely) determined by mass/count distinctions (cf. Chierchia 1998).