YP
Yep, that’s very close to the motivation behind this. The main difference in my approach is that I’m trying to avoid treating the summary as a single lossy blob. I split the history into atomic facts, preserve or compress them individually, and keep mappings back to the original records. So recovery can happen selectively rather than pulling the entire log back into context.