Notes · 2026-10-19
Graph-only mode: proving the graph decides
If you can switch the LLM off and every verdict stays the same, the graph isn't decoration.
In Why-Graph, four agents investigate a production incident on a shared FalkorDB graph. The risk with any LLM-driven investigation is that the model writes a convincing story and the "evidence" is decoration.
So we built two rules into the system:
- Agents can only call named Cypher queries from a catalog. No free-form queries.
- Every finding is re-verified. Recording a finding re-runs the cited query, and any claim whose nodes and edges aren't in the result is rejected.
Then we added a switch: graph-only mode. It turns the LLM off entirely. The narrative disappears, but every verdict (the root cause, the blast radius, whether a new change is risky) stays exactly the same.
That switch is the test. If verdicts changed when the model went away, the model was deciding. Because they don't, the graph is deciding and the model is only explaining.
It's also why the pull-request guard could honestly say "no known risk" before the first incident. It had no lesson in the graph yet, and it didn't guess.