The life-cycle of the Scottish Vowel Length Rule
The historical SVLR, as characterized in Lass (1974) and Pullum (1974) and summarized above, was a bipartite lengthening and shortening change, which was probably introduced in the sixteenth century. A slightly modified version of Lass's formulation is given in (1). Recall that in some varieties, low or non-high vowels were exempt from a., and are therefore consistently long synchronically.
(1) 

Lass's SVLR clearly makes vowel length predictable. Pullum (1974) therefore argues that the implementation of the historical SVLR would inevitably have led to rule inversion (Vennemann 1972), restructuring the underlying Scots vowel system. In other words, speakers would no longer learn a vowel system with an underlying length contrast, plus a complex neutralizing rule; instead, they would abduce that all vowels are under lyingly short, and lengthen a subset before /r/, voiced fricatives and boundaries. In a theory incorporating underspecification, one might alternatively propose (see Anderson 1993) that most or all Scots vowels are unspecified for length, with SVLR operating in a blank-filling capacity. Carr (1992) rejects this approach, partly on the grounds that the application of underspecification theory to metrical properties like length is unclear. Carr also points out that radical underspecification makes the operation of the central constraints of LP almost impossible to ascertain. I shall show that underspecification also brings other disadvantages; for the moment, I assume that the majority of Scots and SSE vowels became underlyingly short following the historical SVLR.

However plausible this account may be, it treats SVLR very much as an isolated phenomenon. I assume instead that LLL and SVLR constitute two stages in the life cycle of sound changes illustrated above with æ-Tensing. We can see the `voicing effect' lengthening as an English specific phonologization of a universal tendency. This automatic phonetic process has been phonologized in most varieties of Modern English as the postlexical rule of Low-Level Lengthening, which affects all vowels before voiced consonants. However, in Scots dialects and SSE, a further stage of phonologization has taken place: the extreme lengthening environments of the LLL schema were phonologized in Scots/SSE as a separate rule, which has acquired certain properties characteristic of lexical rule applications, and hence been relocated in the lexical phonology. The overlapping contexts of LLL, which operates before voiced consonants and pauses, and SVLR, which applies in these varieties only before voiced continuants and boundaries, are shown in (2).
(2) 

This separation of the two processes, and the eventual lexicalization of SVLR, seems likely to have begun with a modified version of Lass's (1974) historical SVLR, whereby tense vowels underwent some additional lengthening in Middle Scots before voiced continuants and pre-pausally. Since these are the contexts which are in any case most conducive to vowel lengthening, and since one general characteristic of tense vowels is their greater length relative to lax vowels, this extra increase in duration might have been sufficient to cross the perceptual threshold for durational differences, making this lengthening labellable, as previous lengthening controlled by LLL had not been (and arguably, still is not). If speakers could auditorily distinguish tense vowels in these extreme lengthening contexts from all other vowels, we might propose a perceptual recategorization, whereby just these vowels in these SVLR long contexts were reinterpreted as long, and all others as short. This account is in line with an analysis of phonologization proposed by Harris (1986), who observes that phonetic variation may have either intrinsic or extrinsic sources. Intrinsic factors include universal phonetic constraints, while extrinsic factors are language-specific and might include control by a phonological rule. The intrinsic factors governing the voicing effect lengthening summed up in LLL are principally voicing, and the rate of closure transition. According to Harris (1986), phonologization involves the systematization of intrinsic differences; specifically, variation resulting from intrinsic causes is reinterpreted as having some external source. In the case of SVLR, lengthening in the extreme environments of the voicing effect schema is attributed to a Scots/SSE-specific lengthening rule. Harris further suggests that such reinterpretation may be due to system-specific exaggeration of a variant `to the extent that it can only be subsequently interpreted in extrinsic terms' (1986: 123). This suggestion is relevant to SVLR in two ways. First, the separation of SVLR from the intrinsically determined voicing effect lengthening may be due to this system-specific exaggeration. Secondly, the putative exaggeration might be understood to involve, not only additional lengthening in long contexts, but also an opposing reaction of shortening in other environments ± hence the tendency of Scots vowels to be shorter in SVLR short contexts than comparable vowels in RP (Agutter 1988a, b). Harris further assumes that previously intrinsic contrasts, once extrinsically determined, may percolate deeper into linguistic structure, as we can see from the subsequent history of SVLR.

By affecting only tense vowels with tense sources, the historical SVLR also disrupted the previously perfect correlation of tenseness with length. After the introduction of SVLR, tense vowels could no longer be defined as those vowels which are always long, but rather as those which are sometimes long; that is, those vowels with audibly long realizations in some contexts. From this point, it is a very small step to assume that [+ tense] became the crucial feature specification defining the input to the synchronic SVLR, which would then have separated from LLL. Scots speakers would no longer operate with an underlying vowel system contrasting long and short vowels; instead, length would be predictable on the basis of the pre-existing feature [± tense] and the new SVLR. However, LLL was also retained, continuing to produce minor and arguably inaudible alterations in the length of all vowels.

The next question is how the nascent SVLR came to apply lexically; and the most feasible course involves analogy. We can assume that the final vowel in infinitival forms like die, row would have lengthened by LLL and the new SVLR. However, there is no lengthening context for SVLR in the past tense forms died, rowed or present dies, rows which would surface with a short vowel in post-SVLR Scots. A tendency towards restoring iconicity might have caused the lengthening to be generalized into this originally inappropriate environment. This innovation would have led to the reformulation of the rule to include a bracket or boundary, making SVLR sensitive to morphological information and therefore lexical. This analogical extension would then also extend to ostensibly similar past tense forms like rode, strode, where SVLR would have applied productively until the ablaut past tenses became fossilized.

Once a rule has acquired some characteristic of lexical application in this way, and consequently been propelled into the lexicon, we might expect it to begin to exhibit further properties of lexical rules. This is the case for SVLR; for instance, lexical rules generally produce results which are observable or categorizable by native speakers, and many Scots/SSE speakers can in fact distinguish long vowels in SVLR contexts from short ones elsewhere. This observability does not entail that the length contrast must be present at the underlying level, since it is generally assumed within LP that speaker judgements on distinctness of sounds are based on the lexical rather than the underlying level (Mohanan 1986): vowel length in Scots/SSE will then be a `derived contrast' (Harris 1989), which is produced during the lexical derivation.

This assumption that vowel length is no longer underlyingly distinctive in Scots/SSE is important from the point of view of Structure Preservation, which permits lexical rules to operate on or introduce only under lyingly distinctive features. If SVLR neutralized the long-short contrast early in its life-cycle, it synchronically manipulates a non-contrastive feature, and therefore contravenes Structure Preservation, at least in those varieties of Scots in which long vowels do not form part of the inventory of basic prosodic templates. Adopting Borowsky's (1990) restriction of Structure Preservation to Level 1 will not help, since SVLR applies on both Level 2 and Level 1. Furthermore, Kiparsky (1988) asserts that, once rules become lexical, they are free to undergo lexical diffusion. However, I have given no indication that SVLR is undergoing or has undergone such diffusion.

In fact, there are signs of incipient lexical diffusion of SVLR, although this is at present limited to the diphthong /ai/. As we have seen, the long realization, [a:i], is now being generalized into lexical items lacking long contexts, giving [pa:ilɔn] pylon, [spa:idər] spider (compare [wΛidər] wider), and [va:ipər] viper (compare [wΛipər] wiper). This extension of long [a:i] is still sporadic, speaker-specific and highly variable, but appears to be spreading; indeed, Aitken (1981), Abercrombie (1979) and Wells (1982: 399ff.) consider the evidence sufficient to posit a phonemic split of /Λi/ ~ /a:i/, a point echoed by Donegan (1993) for Canadian Raising. It is possible, as Carr (1992) suggests, that SVLR has simply been extended to apply in open syllables, at least for some speakers; but why would this affect only one vowel, and why should it be so inconsistent across speakers and lexical items?

We may then propose that /Λi/ and /a:i/ are now both part of the underlying vowel system for Scots/SSE; this /Λi/ will merge with /Λi/ found word-finally in Scots, in pay, way words. The establishment of this marginal contrast is of some theoretical importance, since it both testifies to the lexical diffusion of SVLR, and marks a tenuous re-establishment of the length contrast in Scots/SSE. In varieties without consistently long low vowels, the introduction of underlying /a:i/ will provide the only evidence of a length contrast above the lexical level (with the possible exception of a few irregular past tense forms), and will therefore go some way towards guaranteeing that SVLR obeys Structure Preservation. It remains to be seen whether SVLR will continue its diffusion through the other pairs of vowels; given that a quality difference is only apparent for the diphthong, this may be unlikely.

The account of the development of the SVLR given above can be seen to support Kiparsky's (1988) association of diffusing changes with lexical rules, and Neogrammarian changes with postlexical rules, given Harris's (1989) proviso that Neogrammarian/postlexical processes may develop diachronically into lexical rules. SVLR also, to some extent, supports the notion of a life-cycle of changes and rules, suggested by Harris's (1989: 55) view that implementation as a postlexical rule; lexicalization; and fossilization, loss, and integration into the underlying representations reflect `different stages in the ageing process of sound change ... whereby individual changes ... percolate deeper and deeper into the linguistic system'.

However, there are three differences between SVLR and Harris's example of æ-Tensing. First, æ-Tensing has become lexical in toto in the varieties where it remains productive, whereas SVLR represents only a partial lexicalization of LLL, which remains postlexically even in the varieties which have innovated SVLR. Secondly, Harris (1989: 54) proposes that, although a newly lexicalized rule may not be structure preserving, `the reassertion of Structure Preservation would then be predicted to dictate the direction of any subsequent change'. The diffusion of SVLR, with the generalization and incipient contrastivity of long [a:i], may represent a case of precisely this reassertion, since the reintroduction of an underlying length contrast will produce renewed conformity of SVLR with Structure Preservation. The principles and constraints of LP seem in some sense both synchronically and diachronically `real', since they not only control the structure of the synchronic phonology, but are also reasserted when disrupted by ongoing change.

Finally, and perhaps more strikingly, SVLR does not seem to be following exactly the same life cycle as æ-Tensing. æ-Tensing has caused a change at the underlying level in some dialects, but only at the end of a period of increased opacity and fossilization as a lexical rule, which is ultimately lost. However, SVLR neutralized the vowel length distinction in Scots before becoming, or while becoming, a lexical rule. We may then be dealing with two variants of the pathway from sound change to rule to underlying restructuring. One, outlined by Harris, would be characteristic of processes like æ-Tensing, which simply alter some feature value. The other would involve processes like SVLR, which neutralize some pre-existing feature contrast at the sound change stage (and may go on to cause underlying changes later too, as for the first type). This development might be characteristic only of processes which, like SVLR, are analyzed in SGP as involving rule inversion. We will return to this issue, in connection with the deletion and insertion of /r/ in non-rhotic varieties of English, which involves another case of rule inversion.