In order to read successfully, readers need not only to identify the letters in words, but also to accurately code the positions of those letters, so that they can distinguish words like CAT and ACT. Trying to understand how readers code letter position and then use that information to identify familiar words is one of the most topical research areas in the psychology of reading. For a long time most theories claimed that readers code letter position in terms of mental representations of letters that are position-specific, e.g., CAT would be coded by C-1, A-2, and T-3. Recently, we and others have accumulated evidence highlighting problems with this approach. Three alternative theories of letter position coding have been developed to address these problems. This project tested these theories by using a masked (subliminal) priming methodology, in which we examined whether a misspelt version of a word helped to prime identification of that word. We designed experiments for which the three theories each led to distinct sets of predictions regarding the pattern of priming effects. We used a number of variants of the masked priming methodology, including one that we've recently developed ourselves that is particularly well-suited for testing subtle differences in the predictions of different theories. The outcome of our experiments strongly support an approach in which letter representations are position-invariant and position is coded dynamically in terms of patterns of activity. These results help to increase our understanding of the mental representations that form the basis of literacy.