Most models of choice in language focus on broadly applicable generative knowledge, treating item-specific variation as noise. Focusing on word order preferences in binomial expressions (e.g. "bread and butter"), we find meaning in the item-specific variation: more frequent expressions have more polarized (i.e. frozen) preferences. Of many models considered, only one that takes expression frequency into account can predict the language-wide distribution of preference strengths seen in corpus data. Our results support a gradient trade-off in language processing between generative knowledge and item-specific knowledge as a function of frequency.