Suppose you don't know the shape of the content ahead of time. This would be the case for most apps that allow users to upload their own sources.

It sounds like (at the expense of more computation and time) the reverse adaptive tree approach you described would be ideal for those scenarios.