A knowledge graph of 70 accident investigations — chemical plants and ships carrying ammonia, LNG and hydrogen — built so lessons from one industry can be read against the other. Ask in English or Chinese; Claude writes the Cypher query and Neo4j answers.
Every incident is linked to variables on six causal layers taken from the bow-tie model of process safety: organizational factors weaken safety barriers; an initiating event leads to a failure mode (loss of containment), which produces a release that causes consequences.
Ship-specific concepts are mapped onto the plant vocabulary so the two domains share nodes — a cargo-hose rupture is stored as pipe_rupture, with the original hose_rupture kept on the edge.
Incident edges record what one investigation report found. Mechanism edges between variables are causal claims merged across reports; their support is the number of reports asserting them. A chain of mechanism edges is a set of claims, not one observed sequence — use the “Chains that really happened” example for sequences recorded in a single incident.
31 investigations by the U.S. Chemical Safety Board; 39 ship accidents from JTSB, ATSB, BSU, TSB, NOPSEMA, BARPI/ARIA and REMPEC records — 31 in the core scope and 8 supplementary cases (refrigerant ammonia, non-cargo hydrogen, collisions without release) kept for failure-mechanism analogy.