One of the major constraints originally proposed for the lexical component of LP is Structure Preservation, defined by Kiparsky (1985: 87) as

the result of constraints operating over the entire lexicon. For example, if a certain feature is non-distinctive, we shall say that it may not be specified in the lexicon. This means that it may not figure in non-derived lexical items, nor be introduced by any lexical rule, and therefore may not play any role at all in the lexical phonology.

Borowsky (1990: 29) slightly redefines SP, to make it more obviously relevant to issues of syllabification, tone and other prosodic phenomena, rather than apparently limited to segmentals; her version was given in (2.20) above, repeated below as (2.41).

(2.41) Structure Preservation:

Lexical rules may not mark features which are non-distinctive, nor create structures which do not conform to the basic prosodic templates of the language.

As Kaisse and Hargus (1993: 11) note, `The basic concept of structure preservation is a simple one, though ... its proper technical instantiation may be anything but that.' Although it may seem straightforward to require that the underlying and lexical segment inventories and sets of prosodic structures should be isomorphic, with allophones and novel structures derived postlexically, this simple version of SP turns out not to be tenable. Notably, Borowsky (1989, 1990) and Harris (1987) present evidence that SP must switch off at the final lexical level. For instance, Borowsky (1989) argues that only two rhymal positions are permitted in syllables until the end of Level 1, thus necessitating CC shortening in kept, dreamt, depth, but not in Level 2 derived dreamed, deeply. Harris (1987) considers various South-Eastern Bantu languages, where half-close and half-open vowels are in regular harmonic alternation and thus non distinctive, but the relevant vowel harmony rules must be lexical. Similarly, Kaisse and Hargus (1994) review arguments that assimilation rules creating linked structures are exempt from SP. However, they show that not all such rules violate SP, and that those which do are consistently word level rules. Hence, allowing SP to switch off at the final lexical level accounts for the cases raised by Borowsky, Harris, and Kaisse and Hargus. This coheres with Borowsky's (1990: 23±4) contention that Level 2 displays `vaguely split loyalties'. If we accept the Booij and Rubach model, with Level 2 as non-cyclic, both SP and SCC will be limited to Level 1. Level 2 will then be a bridge between the maximally lexical Level 1, and the postlexicon: its lexical properties follow from its interaction with the morphology, while its phonological sympathies are postlexical. This does leave the problem, that some postlexical rules have also been claimed to be structure preserving. In this context, and to test the assumption that SP switches off at Level 2, and the possible reasons for this, we shall return to SP from time to time below. Notably, I shall present evidence which suggests that newly lexicalized rules may violate SP. Such violation of the constraint is therefore a diagnostic of a process which has recently progressed from postlexical to lexical application; and furthermore, it is possible that the reassertion of SP may dictate the future direction of change in such a case.

There remains, however, one serious challenge to SP ± and indeed, to other constraints of the lexical component like SCC. This challenge is the use of underspecification.

Although underspecification and LP are technically independent of one another, many proponents of LP argue against fully specified lexical entries. As Kiparsky (1982: 53) observes, the assumption of under specification will allow cyclic, lexical rules to apply on the first cycle to fill in feature specifications; this will not violate SCC, since such rules will introduce features rather than producing clashing feature specifications. Redundancy rules and morpheme structure rules, given this hypothesis, are simply rules of the lexical phonology applying under special circumstances.

Underspecification has a great deal of unexplored potential; as Kaisse and Hargus (1993: 15) say, `It is apparent that while no one is yet able to agree on exactly how underspecification works, it is a powerful tool.' Not all of this power will be welcome in a theory which is intended to impose and explore constraints on phonological abstractness. For instance, it appears that underspecification makes it extremely difficult to assess the operation of SP. Borowsky (1990: 30) claims that `if the segment /x/ is not a phoneme of English, there is no occurrence of it, or a partially specified form of it, anywhere in the lexicon' (at least, as Borowsky's discussion makes clear, until SP switches off at Level 2). However, if underlying representations can be maximally underspecified, it is unclear how we can tell what is a partially specified form of /x/ and what is not, especially given Borowsky's assumption (1990: 106) that different partially specified underliers may merge on the surface once all features have been filled in. To check that SP is not about to be contravened, we would then have to take all potentially eligible underlying segments, put them through all phonological and default rules, and see if [x] appears as the output and on what level it is derived.

However, SP is not the only constraint defused by the use of under specification. For example, Borowsky (1990: 49) considers the ruki rule of Sanskrit, where /s/ retroflexes after /r u k i/. Retroflexes do appear in underived environments, but the SCC is specifically designed to prevent free rides, and should stop the ruki rule from applying here. However, Borowsky makes use of the limitation of SCC to structure-changing rules, and of underspecification, to effectively neutralize the constraint: if the underlying representations of forms with surface retroflexes are left unspecified for [retroflex], a blank-filling version of ruki can apply without contravening SCC. Borowsky produces similar analyses of Velar Softening (1990: 130) and s-Voicing in English. She does concede that `abstractness moves into the system by the use of partially specified segments' (1990: 54), but seems unconcerned by this. Indeed, considering the possibility that rules might be stopped from applying in underived environments, she notes that `This is simply to miss a generalization from my point of view' (1990: 73). It is at best frustrating that we should impose constraints to force a non-abstract analysis, which can then almost infallibly be circumvented by underspecification. Borowsky has noticed this interaction, but her conclusion is that SCC should be done away with, on the grounds that it does so little work, as `the use of underspecification removes many classic cases which motivated the SCC' (1990: 28). My view is that this is not an argument against SCC, but against underspecification as part of a Lexical Phonology.

I shall therefore exclude underspecification for the time being, and will return to a detailed account of its place in LP. For the moment, discussions of phonological rules will be clearer if it is obvious that I am dealing with feature-changing rather than blank-filling or default applications. Furthermore, since I shall be arguing for a model of LP in which the constraints, particularly SCC but also SP, are of central importance on synchronic and diachronic grounds, it makes no sense to assume underspecification, which bleeds the SCC and makes the operation of SP opaque. Let us first see what these constraints can achieve unimpeded.