A surprising claim made by Frost in the target article is that (in Hebrew) “letter order cannot be flexible but has to be extremely rigid” (sect. 4.2, para. 5). The basis for this claim is the observation that different roots often share different orderings of the same three consonants. But of course the same problem arises in English when distinguishing words like calm and clam. Successful identification of printed words in English and Hebrew requires a combination of exquisite sensitivity to relative position and considerable flexibility with respect to absolute letter position. “Extreme rigidity” might work very well if all of the Hebrew words that shared the same root contained the letters of that root in the same positions, as is the case for the root Z.M.R in zemer (song) and zamir (nightingale). But Hebrew orthography does not admit such a simple solution, as is illustrated by words like zimra (singing, music), z'miryah (music festival), and kleizemer (musical instruments). Once again, the same problem arises in English. To appreciate the relationship between build, builder, rebuild, and shipbuilding, there must be commonality in the orthographic code for the build morpheme in these strings. Rigid, position-specific coding (whether left- or right-aligned) cannot accommodate this requirement, because the letters of build occur in different positions within these strings. This alignment problem (Davis Reference Davis1999) was a key motivation for the development of spatial coding, and it is the fundamental reason why models of reading must address the issue of orthographic input coding. The alignment problem is common to many different orthographies, and models that attempt to solve this problem are addressing a universal constraint.

Frost's invocation of a universality constraint to critique current models is based chiefly on a consideration of masked form priming effects. His review may give the impression that form-priming effects are found without fail in European languages (and virtually never in Semitic languages), but the full story is more complicated. First, masked form primes give rise to inhibition when the prime is itself a word; for example, the prime calm inhibits identification of the target clam (Andrews Reference Andrews1996; see also Davis & Lupker Reference Davis and Lupker2006). This finding is predicted by models like the spatial coding model (Davis Reference Davis2010), because the ability to discriminate words like calm and clam depends both on accurate encoding of letter order and on lexical inhibition. The latter mechanism ensures that the lexical representation that is most strongly supported by the orthographic input is able to suppress its competitors. By extension, representing root morphemes at the processing level above letters (e.g., Andrews & Davis Reference Andrews and Davis1999; Davis Reference Davis1999, pp. 324–53) would lead to the prediction that transpositions that give rise to a different root than that embedded in the target should result in inhibitory priming, which is what Velan and Frost (Reference Velan and Frost2009) observed in Hebrew. Second, null form-priming effects are found in European languages for words with dense orthographic neighbourhoods (e.g., Forster & Davis Reference Forster and Davis1991; Perea & Rosa Reference Perea and Rosa2000b). Thus, it is not at all clear that either the inhibitory or the null form-priming effects that have been reported in Hebrew (a notoriously dense orthography) are inconsistent with findings from European languages. Furthermore, monomorphemic words with sparse neighbourhoods show the same pattern of facilitatory priming effects in Semitic and European languages. There is every reason to think that a successful model of Hebrew word identification could incorporate the same coding and processing mechanisms as the spatial coding model (Davis Reference Davis1999; Reference Davis2010).

The other constraint discussed by Frost is what he refers to as the linguistic plausibility constraint. In criticising the “new wave” of “orthographic models” for failing this constraint, Frost appears to suggest that the spatial coding model is “structured in a way that goes counter to the established findings for other linguistic dimensions” (sect. 2.2, para. 1, emphasis in the target article). But if so, it is unclear what aspect of the model he is referring to. Elsewhere Frost criticises the model for focussing exclusively on orthographic processing, but as Davis (Reference Davis2010) notes, this is not intended as a claim regarding the structure of the visual word recognition system, but rather as a commitment to nested modelling (e.g., Jacobs & Grainger Reference Jacobs and Grainger1994) and what Andrews (Reference Andrews and Andrews2006) has referred to as temporal (as opposed to structural) modularity. That is, like other modellers, I have taken the approach of focussing on specific components of the full problem, but using a modelling framework that has been successful in capturing other aspects of performance, including effects in the Reicher–Wheeler test, speeded naming and lexical decision (e.g., Coltheart et al. Reference Coltheart, Rastle, Perry, Langdon and Ziegler2001; McClelland & Rumelhart Reference McClelland and Rumelhart1981; Perry et al. Reference Perry, Ziegler and Zorzi2007). The commitment to temporal modularity does not imply “that an adequate description of the cognitive operations involved in recognizing printed words is constrained solely by the properties of orthographic structure” (sect. 8, para. 1, emphasis in original), but rather that, for skilled readers, printed words are most often identified principally on the basis of orthographic information, with phonological and semantic information being retrieved later.

Frost argues that models which learn will provide the ultimate answer to our modelling problems. I am a proponent of such models, but also appreciate the value of using hardwired models to better understand what must be learnt. Correlations between orthography and semantics or phonology cannot be discovered by any learning algorithm unless the input is structured in a way that preserves these correlations. Frost notes that, well before the new wave of models, Seidenberg and McClelland (Reference Seidenberg and McClelland1989) offered an alternative to position-specific coding. What he doesn't note is that subsequent iterations of this modelling framework (correctly) blamed this alternative for the failure of this model to satisfactorily learn the mapping from orthography to phonology (Plaut et al. Reference Plaut, McClelland, Seidenberg and Patterson1996). The need to learn context-invariant mappings from letters to phonemes highlights another aspect of the alignment problem, and another motivation for assuming position-independent letter representations (Davis Reference Davis2010). In summary, cracking the orthographic input code is not simply an intellectual puzzle concerned with explaining transposed letter effects – it is a fundamental requirement for developing a general theory of reading.