Abstract. We propose an extension to current view-based mediator systems called model-based mediation, in which views are defined and executed at the level of conceptual models (CMs) rather than at the structural level. Structural integration and lifting of data to the conceptual level is ``pushed down'' from the mediator to wrappers which in our system export classes, associations, constraints, and query capabilities of a source. Another novel feature of our architecture is the use of domain maps, semantic nets of concepts and relationships that are used to mediate across sources from multiple worlds (i.e., whose data are related in indirect and often complex ways). As part of registering a source's CM with the mediator, the wrapper creates a ``semantic index'' of its data into the domain map. We show that these indexes not only semantically correlate the multiple worlds data and thereby support the definition of the integrated CM, but that they are also useful during query processing, for example, to select relevant sources. A first prototype of the system has been implemented for a complex Neuroscience mediation problem.