Six months later, a regulator asks why Susan bound a $4.2M risk on March 5. This example shows how the decision and its evidence can be stored as a graph and replayed with Cypher. The people, policies, values and timestamps are fictional.
In this example, InvariantDB stores the full trail behind the final answer (BIND $4.2M): the context Susan saw, the rules that fired, the documents she opened, the reasoning she wrote, and the signature records. Each step is a node, each causal link is an edge, and each value carries a valid time and a recorded time.
AT RECORDED '2026-03-05T14:32:18Z' reconstructs what Susan saw at the moment she bound the policy, not today’s view.
GET /graphs/{name}/audit/verify re-hashes the graph’s write history from the first record and reports clean or the first broken record. The history log is not signed, so the check does not detect a rewrite of the whole log, or removal of the most recent records, by someone with write access to the server.
DERIVED_FROM edges, walked with db.derivedFrom, trace the rate quote back to its inputs: filed rates, CAT score, sprinkler credit, and loss-history factor.
This is what "build agents from decisions, not documents" looks like in practice. Replay Susan's last 800 bindings to train an underwriting assistant that learned how she actually thinks, not just what's in the SOP.