The problem here is equating the "top level" number with category in the first place. It's just an idea and the number is a location. Start by separating the two.

https://github.com/CondeNast/atjson

They're using annotations in this context more like CSS, but instead of adding the markup into the content as is done in HTML and processing it, they've physically separated the text and the markup entirely and are using location within the text to indicate where the formatting should take place.