Although both external (social) and internal (linguistic) factors influence variable production, it is unclear to what extent social evaluation is sensitive to internal factors. This study asks whether T-flapping in Singapore English is evaluated differently depending on the word class hosting the variable token, comparing numbers, which show high rates of T-flapping in production, to non-numbers, which show minimal T-flapping. We conducted a series of matched-guise experiments crossing word class (number, non-number) with variant (flap, stop), eliciting native speaker ratings (n=66) designed to capture judgments of localness (naturalness, fakeness, and closeness to listener's own speech). We find a word class-by-variant interaction such that flaps are rated as sounding less local in non-numbers than numbers, whereas stops were rated equally local regardless of word class. These results add to evidence that listeners use knowledge about probabilistic constraints on variable production to inform social perception, and that social evaluation can be sensitive to internal factors. We propose that some axes of social evaluation, such as the ones used here, probe intuitions pertaining to stylistic well-formedness and knowledge of community norms of variable production, and hence are likelier to show such sensitivity.