In SGP, the drive to construct the simplest possible phonology (where simplicity is calculated with reference to feature counting and maximal rule application) led to the rejection of the classical phonemic level of representation or any equivalent to it, with the result that SGP lost any ready way of encoding surface contrast or the speaker intuitions which seem to relate to it, and it became impossible to restrict the distance of underlying representations from the surface. However, LP has three linguistically significant levels of phonological representation. It shares the underlying representations of individual morphemes, and the phonetic representation (the output of the morphology, phonology and syntax, which contains near-surface forms of phrases), with the SPE model. But LP also includes the lexical representation (Mohanan 1986: 10), the output of the phonological derivation at the end of the lexicon, which involves neither morphemes nor phrases, but words.

The lexical representation is not necessarily identical with the phoneme level, although it equally allows LP to refer easily to surface contrast, and equally is relevant in language acquisition, perception and production. Mohanan (1982: 12±13; 1986: ch.7) discusses a number of phenomena which seem to have the lexical representation as their locus: these include speaker judgements on whether sounds are the same or different, and speech errors which permute segments, while secret code languages like Pig Latin seem to perform a coding operation on the lexical representation, then apply the postlexical rules. Mohanan (1986: 194) also proposes that speakers enter words in the mental lexicon in their lexical representations: `underlying representations of the constituent morphemes of a word are arrived at as and when the speakers come across morphologically related words which provide evidence for the underlying forms'.

All this, however, rests on a clear distinction between lexical and postlexical rules or rule applications; and that clear distinction seems in some respects to be breaking down. We have already seen that Mohanan (1986) proposes two `levels' of postlexical rules, challenging the association of level-ordering with the lexicon; Kaisse argues similarly that `postlexical rule application is a more complex phenomenon than the simple across-the-board matter we once thought it might be' (1990: 127). Some postlexical rules may also show properties hitherto seen as lexical: for instance, Carr (1991) discusses the postlexical neutralization rule of Tyneside Weakening, whereby /t/ → [ɹ] in word-final intervocalic position, as in not a chance, put it down, delete it. Carr notes that Weakening does not affect feet like putty, fitter, which are formed in the lexicon, but only those created postlexically by cliticisation, and concludes that we require a notion of postlexical derived environment. Carr also shows that Weakening is in an Elsewhere relationship with the later, more general and across-the-board rule of Glottalization, as shown by the application of Weakening and not Glottalization in fit her. Finally, Carr argues that Weakening is Structure Preserving, although as we have seen, SP has generally been seen as a property of lexical rules, and that it is undergoing lexical diffusion; we return to lexical diffusion, which is a vital ingredient in LP accounts of sound change.

Essentially, then, Carr's paper challenges the restriction of Structure Preservation, the Elsewhere Condition, derived environment effects and lexical diffusion to the lexicon. There is a growing awareness in LP that the lexical-postlexical division may be gradient: thus Kaisse (1990: 130) observes that `the most lexical of lexical rules occur at Stratum 1, while less lexical characteristics emerge as one travels ``down'' towards the word level and the postlexical level(s)', and that, equally, we might expect those postlexical rules nearer the lexicon to share some lexical properties. A similar contention is found in Pandey (1997: 92), who identifies `a property of interfacing modules, namely, polarity, which is the presence of different properties of representation and rule application at its opposite ends'. That is to say, lexical rules may become progressively less lexical in character as we approach the postlexicon; and conversely, early postlexical rules may exhibit some properties of the lexical syndrome. This might then account both for Carr's observations on Tyneside Weakening, and the possible suspension of, for example, SP at Level 2 of the lexicon. There are obvious difficulties with Pandey's claim, which is not yet well worked out: for instance, if SP did not operate at Level 2 of the lexicon, it is hard to see how it could percolate into the high postlexicon; and Tyneside Weakening may remain problematic, since Carr does not characterize it as an early postlexical rule, but as part of a group operating between the post syntactic and across-the-board processes. More centrally, Pandey's approach of proposing a `Polarity Principle' may not be the right way of dealing with what is clearly a variable and gradient situation; and it is unclear what Pandey means by the `ends' of a component (in a two-level lexicon, both Level 1 and Level 2 may be `ends', and the whole lexicon is therefore `polar' in his terms). Nonetheless, this work is symptomatic of a realization that the dividing line between lexical and postlexical rules may not be a rigid one, and that, again, we must begin to look instead at where and why particular properties are suspended or activated.