One nearly ubiquitous assumption in models of linguistic comprehension and of eye movement control in reading alike is of partial modularization between word-level and sentence-level processing: that the outcome of word recognition, and thus the input to sentence-level comprehension, is a categorial representation.Yet such a partial modularization throws away residual uncertainty re- garding word identity that might potentially be of value to the comprehender further downstream in the sentence.Here I describe a line of research combining computational modeling with experimental eye-tracking work to explore the consequences of removing this partial modularity assumption.