A comparison of speakers’ treatment of two categorically unattested phonotactic structures in Cochabamba Quechua reveals a stronger grammatical prohibition on roots with pairs of ejectives, *[k’ap’u], than on roots with a plain stop followed by an ejective, *[kap’u]. While the distribution of ejectives can be stated as a single restriction on ejectives preceded by stops (ejective or plain), *[−cont, −son][cg], speakers show evidence of having learned an additional constraint that penalizes cooccurring ejectives more harshly, *[cg][cg]. An inductive learning bias in favor of constraints with the algebraic structure of *[cg][cg] is hypothesized (Marcus 2001, Berent et al. 2002, Berent et al. 2012), allowing such constraints to be preferred by learners over constraints like *[−cont, −son][+cg], which penalize sequences of unrelated feature matrices.