Sceptical observers, and non-generative phonologists, may see my programme as excessively idealistic, on the not unreasonable grounds that generative phonology is by its very nature far too flexible to allow adequate constraint. In other words, given phonological rules and underlying forms, an analysis can always be cobbled together which will get the right surface forms out of the proposed underliers: if the first attempt doesn't do the trick, you can alter the underliers, or the rules, until you find a set-up that works. And since LP is generative, and phonologists are no less ingenious now than in the heyday of SGP, the new model is open to precisely the same criticism as the old one. Here again, Lass and Anderson (1975: 226) ask: `But is the mere fact that a phonological solution works any guarantee that it is correct?' Of course not: it is precisely because we cannot rely purely on distribution and alternation that we need extra, `external' evidence. The analyses I shall propose will look peculiar in SGP terms; but I hope to show that they are coherent with evidence of a number of different kinds, and that they allow interesting predictions to be made. For instance, we shall see that my analysis of the English Vowel Shift specifies a principled cut-off point between what can be derived, and what cannot, giving a partial solution to the determinacy problem. A typical progression from sound changes to phonological rules will also be identified, giving a certain amount of insight into variation and change, as well as the embedding of change in the native speaker's grammar. These implications and conclusions lend support to LP, and suggest, if nothing else, that the model should be pursued and tested further. Phonetics, phonology, variation and change cannot be integrated in this way in SGP. I have not yet seen similar clusterings of evidence types in non-generative phonologies, either.

Arguments of this kind give me one reason for adopting LP, and attempting to constrain generative phonology, rather than rejecting a derivational model altogether. Nonetheless, questions will undoubtedly be raised concerning the relevance of this work, given the current move towards monostratal, declarative, and constraint-based phonologies. I cannot fully address these issues here, but the rest is intended as a partial answer; and I also have some questions of my own.