Number marking on verbs in Murrinhpatha (non-Pama-Nyungan, Australia) has two striking properties. First, apparent mismatches occur: *singular* marking is used for some *dual* subjects and *dual* marking is used for some *paucal* subjects. Second, the singular-dual alternation is conditioned by the linear position of an apparently unrelated morpheme, the non-sibling subject marker *ngintha/nintha* -- a pattern which Nordlinger and Mansfield (2021) argue diverges from known morphotactic principles reported in the literature. I present an analysis of the facts as emerging from the action of Agree operating over a featurally complex representation of number (Harbour 2014), alongside an economy condition on lexical representations (Preminger 2019). Mismatches result from `partially-defective' intervention: Agree takes place not with the subject itself but with an intervening element which bears a subset of the subject's features. The analysis has implications not only for the mechanics of Agree, but also for the nature of exponence (which I propose does not target features directly, but rather the meanings denoted by features) and mobile affixation (which I propose is syntactic).