Most current work in linguistics acknowledges that the organization of linguistic information in a sentence is sensitive to the speaker's assumptions regarding his hearer's knowledge state. What is less clear, however, is how and where the organization of information—information structure (IS)—is carried out in the grammar, and precisely what role it plays in shaping the output of the grammar. In this study I argue that IS is an independent component of the grammar, whose primitives combine to form IS representations in accordance with a set of well-formedness conditions. These representations not only determine if a given output is licit or not, but also feed the semantic and phonological representations, thus regulating inter alia the predication relations in the sentence and the placement of prosodic prominence. The main claims of this study are supported by an in-depth analysis of two phenomena, which I maintain are information structural in nature: focus intervention and weak crossover effects. In both cases, non-IS analyses are shown to fall short in capturing the available data, while an IS approach manages to weave a range of seemingly unrelated observations into a descriptively and explanatorily adequate account. The case study of focus intervention provides a window into the well-formedness conditions on IS representations, while weak crossover helps us understand the internal composition of these representations and their relationship to other levels of representation in the grammar. The two phenomena also establish the import of implicit contextualization, i.e. the fact that speakers impose a context on sentences given in isolation, which guides the mapping to IS categories. In the course of the investigation of these phenomena, significant insight is gained into a variety of topics, ranging from the status of focus in the grammar to the interpretation of quantificational expressions in natural language. The findingsof the case studies justify a reassessment of current grammatical architectures. I propose an architecture in which much of the burden is shifted to the IS component, resulting in a simple, truly autonomous computational system, in line with the original model of the grammar in the generative tradition and with Minimalist assumptions.