Grammar is central to language description and a posteriori construct validation of language tests consistently identifies grammar as a significant factor in differentiating between score levels and characterizing overall proficiency. However, there is currently no model of grammatical competence robust enough to be operationalized in tests. Critical to describing grammatical competence is complexity of form and structure, yet grammatical complexity is poorly defined in linguistics and its sub-disciplines, including language assessment. This article argues that the inherent subjectiveness of complexity can be much reduced by a research methodology that combines input from corpora with expert intuition of the linguistic status of the grammatical items retrieved. The significance of the discussion is that, in addition to contributing to the on-going process of construct validity, a corpus-informed study of complexity has the potential to reconcile the tension in theoretical linguistics between grammar as being sentence-bound and grammar as discourse.