Existing approaches for representing the provenance of scientific workflow runs largely ignore computation models that work over structured data, including XML. Unlike models based on transformation semantics, these computation models often employ update semantics, in which only a portion of an incoming XML stream is modified by each workflow step. Applying conventional provenance approaches to such models results in provenance information that is either too coarse (e.g., stating that one version of an XML document depends entirely on a prior version) or potentially incorrect (e.g., stating that each element of an XML document depends on every element in a prior version). We describe a generic provenance model that naturally represents workflow runs involving processes that work over nested data collections and that employ update semantics. Moreover, we extend current query approaches to support our model, enabling queries to be posed not only over data lineage relationships, but also over versions of nested data structures produced during a workflow run. We show how hybrid queries can be expressed against our model using high-level query constructs and implemented efficiently over relational provenance storage schemes.