Smolensky's (1996) original proposal of a single grammar
for production and comprehension does not seem to hold
in the face of phonological alternations.
If comprehension is to be modelled as an Optimality-Theoretic
grammar, this recognition grammar must contain lexical-access
constraints whose rankings depend on the semantic context
and on frequency of occurrence.

These lexical constraints interact with faithfulness
constraints in much the same way as the
structural constraints do in the production grammar.
The lexical constraints come into existence by the violation
of an anti-lexicalization constraint in the recognition grammar,
and are subsequently automatically ranked in the appropriate
confusion-minimizing order by a gradual learning algorithm.

It is perfectly possible that the production and
recognition grammars contain the same faithfulness constraints
ranked in the same order. If this is correct, the two grammars
can be merged, because structural constraints apply vacuously
in the recognition grammar, and lexical-access constraints
apply equally vacuously in the production grammar.