http://gump.apache.org/metadata/
The XML metadata is recursively read (into a cache) and processed using some complex Python classes. The SAX events are processed by objects that content the schema, via delegate classes. Their construction is based of a combination of the SAX events and the named (Python) attributes they match (by name) on the current Python object.
The result is a tree of Python objects that match the XML, but do allow for traditional Gump merging.