A new research paper proposes a database that refuses to accept a fact from an AI agent until it passes a governance check.
The paper introduces Quipu, an embeddable knowledge-graph store designed for a world where AI agents, not humans, do most of the writing. Instead of accepting writes now and cleaning up later, Quipu gates every incoming fact by evaluating the resulting state before committing it. It tracks two timelines at once - when something happened and when the system learned about it - and treats named graphs, not individual facts, as the unit of trust, so merging data sources can only tighten access, never loosen it. Governance rules, the audit trail, and signed verdicts are themselves stored as facts, so checking whether the system followed its own rules becomes a simple query.
In tests against a benchmark called Census, seeded with six known defects, the gated version of Quipu caught all six while an ungated version caught none. On a separate benchmark, DEMM-Bench, Quipu's exported records correctly answered all 512 governance questions, while systems that only check whether a data container exists overclaimed on up to 87.5% of the same questions. That gap is the real story: most knowledge-graph and RAG systems bolt governance on after the fact through dashboards and middleware that an autonomous agent can simply write around.
Worth remembering: these are the authors' own benchmarks, run once, on their own system - a promising proof of concept, not independent validation.