So far, the comprehension and production language modes have typically been studied separately in generative second language acquisition research, with the focus shifting from one to the other. This article revisits the asymmetric relationship between comprehension and production by examining the second language (L2) acquisition of the noun phrase in Bulgarian by speakers of English. Drawing on data from a forced choice comprehension task and elicited production task, I show that in some cases production of certain grammatical features may precede the full comprehension of their relevance to the morpho-syntax of the target language. An explanation is sought based on (1) Jackendoff's (2002) parallel architecture of the linguistics modules, and (2) feature differentiation based on their relevance to syntax proper and semantics. Structural features with no semantic bearing (e.g. [gender]) are argued to be overlooked by the parser in comprehension. Conversely, features with a semantic load (e.g. [number]) are argued to cause an overload in the semantic tier in production. This, in turn, prevents their overt realization. This study sets the stage for future investigations into the phenomenon exploiting more refined methodology, which will allow for teasing apart the theoretical proposal from task effects.