<p>A corpus-based analysis of Polish double diminutives shows evidence of the avoidance of identical and similar consonants across a morphological boundary. New data from a large corpus are used to demonstrate that this dissimilation pressure leads to lexical gaps and is gradient in the sense that combinations of more similar consonants are avoided more strongly than combinations of less similar consonants. The type and token frequency data extracted from a corpus help establish a scale of dispreference in which phonological similarity plays the key role. A sequence of two identical retroflex affricates is avoided the most, followed by sequences of non-identical retroflexes and sibilants. The pattern is morphologically restricted in two senses: (i) it applies to heteromorphemic sequences and (ii) it targets nouns with two <i>-ek</i> suffixes (<i>root-ek-ek</i>); other constructions are not affected. It is demonstrated that (i) phonological constraints are sensitive to morphology, that is, they are indexed to a particular morphological construction, (ii) morphological computation must have access to the phonological information in fully parsed output candidates, (iii) similarity avoidance is gradiently dependent on the degree of similarity, and (iv) corpus-based analyses are instrumental in investigating gradient generalizations. The results of this study bolster the claim that universal pressures are subject to morphologization, that is, they can be attributed to a specific construction.</p>