Models of word production and comprehension can be split into two broad classes: localist and distributed. In localist architectures each word within the lexicon is represented by a single unit. The distributed approach, on the other hand, encodes each lexical item as a pattern of activation across a set of shared units. If we assume that the localist representations are more than a convenient shorthand for distributed representations at the neuroanatomical level, it should be possible to find patients who, after brain injury, have lost specific words from their premorbid vocabulary. Following a closed head injury, JS had severe word-finding difficulties with no measurable semantic impairment nor did he make phonological errors in naming. Cueing with an initial phoneme proved relatively ineffective. JS showed a high degree of item consistency across three administrations of two tests of naming to confrontation. This consistency could not be predicted from a linear combination of psycholinguistic variables but the distribution fitted a stochastic model in which it is assumed that a proportion of items have become consistently unavailable. Further evidence is presented which suggests that these items are not, in fact, lost but rather have a very low probability of retrieval. Given phonemic cueing of sufficient length, or delayed repetition priming from a written word, the consistently unnamed items were produced by JS. Additional data is reported which seems to support a distributed model of speech production. JS's naming accuracy for one set of pictures was found to predict his performance on a second set of items only when the names of the pictures were both semantically and phonologically related (e.g., cat-rat). There was no association for pairs of pictures if they were only semantically (e.g., cat-dog) or phonologically related (e.g., cat-cap). It is argued that JS's data are best described in terms of a graded, non-linear, distributed model of speech production.