When the mime-type of xslt data page is application/json and output transformation is defined, than datatypes aren’t coverted correctly to JSON. Everything is converted to string instead (integer, float, boolean).
We can employ an existing field in XSL transformation called ‘Datastructure’. It’s not even beeing saved now. So we would start to save that field (simillary as it is done in Validation Rule). And it will be used during the actual transformation - the result of transformation is merged into that datastructure. But it’s used only when it’s called from the places, where a datastructure is already defined (as it is e.g. in sequential workflow transformation step)