According to tradition, in assessing the validity of a natural language argument, context must be fixed. (We wouldn't want to treat "I am happy. So, I am happy", where I assert the premise, and you the conclusion, as a counterexample to a classically valid pattern.) Yet, research in dynamic update semantics, stresses the importance of context-change for capturing intuitions about natural language validity (for instance, in capturing intuitions about argument patterns involving modal discourse). I argue that one cannot capture validity for natural language arguments by imposing a ban on context-shifting because discourses, and so arguments, carry linguistically encoded effects on contextual parameters that affect the meanings of context-sensitive items those discourses contain. However, by contrast with standard dynamic approaches, I also argue that these effects are induced not only by the sentences the discourse comprises, but also by the linguistically encoded relations between them. As a result, individuating argument patterns in natural language discourse requires understanding them not as relations between sets of sentences and sentences, the premises and conclusion, but rather as structured discourses, the structure of which determines the information expressed by the premises and conclusion. Only once we individuate argument patterns in this way can we maintain a tight connection between validity and form, while at the same time discriminating arguments with equivocal contextual shifts, from those that are not, as a matter of their very form.