In the present article, we investigate a largely unstudied cognitive process: word position coding. The question of how readers perceive word order is not trivial: Recent research has suggested that readers associate activated word representations with plausible locations in a sentence-level representation. Rather than simply being dictated by the order in which words are recognized, word position coding may be influenced by bottom-up visual cues (e.g., word length information), as well as by top-down expectations. Here we assessed how flexible word position coding is. We let readers make grammaticality judgments about four-word sentences. The incorrect sentences were constructed by transposing two words in a correct sentence (e.g., Bthe man can run^became Bthe can man run^). The critical comparison was between two types of incorrect sentence: one with a transposition of the inner two words, and one with a transposition of the outer two words (Brun man can the^). We reasoned that under limited (local) flexibility, it should be easier to classify the outer-transposed sentences as incorrect, because the words were farther away from their plausible locations in this condition. If words were recognized irrespective of location, on the other hand, there should be no difference between these two conditions. As it turned out, we observed longer response times and higher error rates for inner-than for outer-transposed sentences, indicating that local flexibility and top-down expectations can jointly lead the reader to confuse the locations of words, with a probability that increases as the distance between the plausible and actual locations of a word decreases. We conclude that word position coding is subject to a moderate amount of noise. Keywords Reading. Word recognition. Word position coding. Parallel processing Very little prior research has explicitly investigated how humans encode the order of words during reading, and a lightning-speed summary of the history of reading research readily reveals why. First, a large chunk of reading research has focused on the recognition of single, isolated words (see, e.g., Grainger, 2008, for a review). Second, research on sentence and text reading has been dominated by the view that words are recognized in a serial, one-by-one fashion, so that knowledge of word order is simply implicated by the order in which recognized words are appended to sentence representations in memory. In short, the encoding of word order during reading has received close to no investigation because it was considered a given. Currently, however, reading research is facing accumulating evidence that words are to some extent processed in parallel rather than serially. This necessitates, more than ever, the answering of a 10-year-old question: How would a parallel processing system keep track of word order? (Reichle, Liversedge, Pollatsek, & Rayner, 2009). In their opinion article , Reichle et al., the creators of the serial-processing E-Z Reader model, noted that a parallel-processing system is likely to recognize words out of order-for instance, when an up-coming word is much easier to recognize than the fixated word-and that it is unclear how the system would handle such occurrences. They then noted that Bone possibility is that a buffer maintains word meanings, and that some mechanism reorders out-of-order words^; however, a Bproblem with this solution is how such mistakes are detected without using comprehension difficulty to signal such occurrences^(pp. 116-117). Indeed, this Bparallel-processing problem^(Snell, van Leipsig, Grainger, & Meeter, 2018) has long been a key argument in favor of serial processing-and perhaps rightfully so, if for no other reason than Occam's razor. Surely the reading