LLM State-Tracking Lab: keeping long sagas consistent

I'm building an open lab for tracking state across long LLM-driven sagas. The reference saga is JJK: The Hollow Vow. Stack is Python + SQLite, with Sheets for display only. The problem I'm working on: over a long run, how do you stop duplicate commits, silently applied stale updates, and conflicts resolved without a human in the loop? Those are the three failure modes I'm measuring against. The repo is live and the work is in progress toward a parser that runs on standardized state blocks and a frozen demo spec, with a scored run planned. Repo: https://github.com/flamingblackman/llm-state-tracking-lab What I'd like feedback on: how would you design the state block so a stale update can't silently overwrite a newer one? If you've built something like this, what broke first?