Use an LLM to describe a difference graph between versions of a compositional document.
The LLM generates a summary of user operations, for audit or for source control commit message.
A self-hosted LLM seems a better proof of concept that simply using a public service like Open AI.
There is a quantized PoC here