Quipu catches six of six graph defects

Quipu’s gated knowledge-graph store ended a seeded multi-writer test with zero of six planted defects, while an ungated store retained all six. The paper also reports 512 of 512 property-level governance questions answered correctly under eight record-degradation conditions, while container-presence baselines overclaimed on up to 87.5% of them. The result is a research prototype, not a production guarantee, but teams letting agents write durable facts should file away the design lesson: put trust, rules, timestamps, and signed verdicts inside the data boundary; the AISI agent-evaluation analysis shows why an external dashboard is a weaker control.