Differences between formal constraints on a generative grammar and concepts of efficiency in transforming sentences provide different expectations regarding performance measures if the grammar is taken as a psychologically real model. To contrast these views, subjects were given sentences varying in voice, mood, and modality and asked to transform them to various syntactic patterns. Their response latencies were not clearly related to either grammatical or performance model expectations. Error frequencies, however, were related only to the performance model which assumed that subjects transform stimulus sentences directly into response patterns without using a kernel form as an intermediate, linking step. The use of formal grammatical models as if they reflected psychological processes is seen as being of questionable value. (Author)