An analysis that seems to be ruled out by <11b> is the analysis of verb second phenomena in several languages. Root sentences in German for example usually are analyzed as <12> (I do not really care about the labels of this tree here).
In these cases the semantic head seems to be a 'gap'.
Therefore probably the instantiation of e.g. its subcat list will only
be known after some relationship is established with the displaced
verb. As a result <11b> is violated.
However in the final version of BUG a solution to this problem is
The 'generate' predicate in BUG will yield a
list of lexical items instead of a list of words. To this list
a small phonological frontend will be applied. This reduces a lot
of uninteresting nondeterminism during the generation process (e.g.
the choice between 'kisses', 'kiss', and 'kissed' is only made
after the generation of the other lexical items).
We can use this strategy to allow for empty semantic heads.
During the generation process there is no
distinction between gaps of semantic heads and their lexical realization
yet (just as there is no difference between 'kisses' and 'kiss' yet).
A possible phonological realization of a feature structure is the empty
string, provided some features are specified locally
(e.g. the value of a slash-like feature as in GPSG, Gazdar et al. 1985).
In this strategy
there is no difference between associating some
inflected form of a verb (e.g. "kissed") with a feature structure
containing the appropriate features (e.g. agreement and tense)
and associating the empty string with a feature structure containing
the appropriate features (e.g. slash).
This strategy furthermore reduces a lot of ambiguity in the generation
of main (verb second) and subordinate (verb final) clauses in languages
like German and Dutch.