The issue here is what the underlying representation of invariantly flapped words like medal and mettle is. I repeat Eric’s thoughts below:
Most folks would probably not posit /mɛɾəl/ for any of the aforementioned morphemes (though this is what Lexicon Optimization would demand, if you take it seriously), but then again there is some ambiguity about whether the tap in each case is basically a /t/ or a /d/. Is it /mɛdəl/ or /mɛtəl/? Does it even matter? … Maybe to our mental representations, but to phonological theory?
I think it does.
For those of you who… ok, for all of you, Lexicon Optimization is a way of choosing formally among a set of underlying forms, all of whom converge on the same optimal output in a given constraint hierarchy. In the ‘standard’ version of Lexicon Optimization (which I believe is in P&S 93), the ‘optimal’ candidate is the one for whom the output is arrived at by the best possible evaluation. In other words, its highest violation is lower than the highest violations of all other candidate underlying forms. I’ll skip the tableau in favour of the basic consequence of this approach: in the absence of alternating forms (like in [sit] vs [siɾəd]), the input and output are identical.
This basically takes all of the fun out of Underspecification.
But maybe there is another way of addressing the lexicon that is more compatible with Underspecification – perhaps like choosing the candidate underlying form that has the fewest featural specifications. Lexicon Underoptimization, maybe. (If someone has already proposed this, I’ll gladly amend this post and catch up on my reading).
Anyway, I think that Thomason’s data basically support this approach. Here’s my reasoning:
(1) When a learner encounters a flap that alternates with /t/ (as inwrite ~ writing), she posits an underlying /t/.
(2) When a learner encounters a flap that alternates with /d/ (as inride ~ riding), she posits an underlying /d/.
(3) When a learner encounters an invariant flap (as in medal), she has no evidence to choose between /t/ or /d/. So, evidently, her strategy is choose /t/.
Obviously this is not a standard “Lexicon Optimization” strategy, but it could be a Lexicon Underoptimization strategy: the learner chooses the stripped-down, less specified, voiceless consonant as the underlying representation.