What worked for me is using Python scripts ("filters") that do all required changes on the Pandoc JSON. I even ended up writing a pythonic layer on top of that JSON ("pip install panflute") so I wouldn't have to deal with the JSON itself.