03 · Graphs · typed, derived, traversed
Money, ownership and control are relationships before they are rows. We keep them as a typed graph derived from the ledger, in the FollowTheMoney model that sanctions and investigative data already speaks, and read it with algorithms instead of hunches.
The model
A company register, a leak and a news article can disagree. Each claim lands in the ledger as its own statement with a source and a date. Conflicts are kept, not resolved by whoever wrote last.
Person, Company, Ownership, Directorship, Payment. FollowTheMoney types mean sanctions lists and leak datasets join without a translation layer.
Memgraph holds a traversable view, rebuilt from statements. Change the resolution rules and replay: the whole graph re-forms, with every edge still pointing at its evidence.
Simulation · fictional structure
Thirteen entities across five jurisdictions. The register shows two shareholders, both companies. Pick a question; watch the machine walk the structure and read its reasoning beside it. Move the threshold to see who crosses it.
Who really owns Quill Freight?
Method · Propagate · effective ownership · O(k·E). Effective ownership flows upward: each holder owns its share of whatever it holds. Iterate to a fixed point, then name the natural persons over the threshold.
Effective ownership is the sum over every path of the product of its shares. Here it reaches 91.0%: the remaining 9.0% is Quill owning itself through a loop.
In production
About 11.5 million documents and 2.6 TB covering 214,488 offshore structures. ICIJ linked the metadata into a Neo4j graph of banks, clients, companies and lawyers that some 370 journalists explored through Linkurious.
Source · Linkurious case study ↗Over four million entities from 471 sources in its default collection, all in the FollowTheMoney model. The schema we build on is the one compliance data already arrives in.
Source · OpenSanctions, October 2026 ↗Roughly 500 billion facts about 5 billion entities by May 2020, up from 18 billion facts shortly after launch. Search answers that are traversals, not keyword matches.
Source · Knowledge Graph (Google) ↗A graph convolutional network over about 3 billion nodes and 18 billion edges, trained on 7.5 billion examples and deployed in production recommendations.
Source · Ying et al., KDD 2018 ↗An LLM extracts an entity graph, communities are summarised ahead of time, and answers are assembled from those summaries. Built for whole-corpus questions that plain retrieval misses.
Source · Edge et al., 2024 ↗Graphs represented as sequences of timed events, with a memory per node. The learning-side twin of an event-sourced ledger: the stream is the graph.
Source · Rossi et al., 2020 ↗