The model of sentence perception proposed by Fodor, Bever and Garrett (1974) emphasizes the importance of grammatical cues signalling clause boundaries, and suggests that segmentation of a sentence into clauses precedes computation of the internal structure of those clauses. However, this model has nothing to say about the many sentences in which clause boundaries are not explicitly marked, and which are therefore more difficult to parse. It is proposed that in such cases clause boundary assignment does presuppose prior analysis of sub-clausal constituents, and that parsing is guided by the strategy called in this paper "late closure." This strategy indicates that, when possible, incoming constituents are to be added into the clause being parsed. A grammaticality judgment experiment using rapid serial visual presentation confirms this theory. (Author/AM)