[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-95-percent-false-positives-is-a-design-choice":3},"\nAsk a trade compliance team what their screening system does all day\nand the honest answer is: generates work. Industry analyses\nconsistently put sanctions screening false positive rates above 95%,\nand many programmes run higher. Hundreds of alerts cleared for every\ngenuine risk found, by analysts whose expertise is spent confirming,\nover and over, that this Rodriguez is not that Rodriguez.\n\nThe scale of the matching problem is real. Hundreds of thousands of\nsanctioned and restricted entities across dozens of lists and\njurisdictions, transliterated names, aliases, shell structures. But\nthe false positive rate is not an inevitable tax on that scale. It is\nthe output of a specific design choice, made so long ago it has\nbecome invisible: the tools screen strings, while the regulation\nregulates ownership.\n\n*The list is not the law. The list is an index into an ownership\ngraph, and the tools only read the index.*\n\n## What the regulation actually says\n\nOFAC's 50 percent rule is explicit: an entity owned 50% or more,\ndirectly or indirectly, by one or more sanctioned parties is itself\nsanctioned, whether or not it appears on any list. Indirect ownership\naggregates and cascades. Two sanctioned parties at 30% and 25% of a\nholding company put that holding company, and everything it majority\nowns, in scope. The US export control regime has been moving the same\nway, extending restrictions through ownership to affiliates of listed\nparties.\n\nRead those rules as an engineer and the conclusion is immediate: the\ncompliance question is a graph traversal. Start from the counterparty,\nwalk the ownership chain upward through every parent and aggregation,\nand determine whether sanctioned control accumulates past the\nthreshold anywhere above. The name on the invoice is one node at the\nbottom of the structure the regulation actually addresses.\n\n## Wrong in both directions\n\nString screening, measured against that regulation, fails on both\nsides at once:\n\n- **Too loud where the law is quiet.** Every fuzzy resemblance\n  between a counterparty and one of the hundreds of thousands of\n  listed names fires an alert, though the vast majority have no\n  ownership connection to any sanctioned party. This is the 95%: not\n  caution, but a category error generating noise at industrial\n  scale.\n- **Silent where the law speaks.** The cleanly named trading company\n  two ownership layers beneath a sanctioned parent matches nothing,\n  because it is on no list, and the rule that puts it in scope was\n  never about its name. The most serious exposure is structurally\n  invisible to the tool the programme relies on.\n\n> The false positives and the false negatives have the same root\n> cause. Fixing the noise and fixing the blind spot are one fix, not\n> two.\n\nThe enforcement climate has removed the comfort margin. Penalties for\nexport control violations now run to nine figures: BIS settled with\n[Applied Materials for $252.5 million in February 2026](https://www.bis.gov/press-release/applied-materials-pay-252-million-penalty-bis-illegally-exporting-semiconductor-manufacturing-equipment),\nthe second-largest civil penalty in the agency's history behind\nSeagate's $300 million in 2023. Enforcement agencies increasingly\nexpect programmes to know who stands behind their counterparties, not\nmerely who they are named after.\n\n## Screening the graph\n\nThe graph-first architecture inverts the pipeline. Counterparties are\nresolved to entities, entities into an ownership graph assembled from\ncorporate registries and ownership data, and the 50 percent\narithmetic runs as a traversal: aggregate sanctioned ownership,\ndirect and indirect, computed per counterparty, continuously, so a\nTuesday change in a parent's status reprices every exposure beneath\nit by Wednesday.\n\nAlerts change species. Instead of \"this name resembles a listed\nname\", the analyst receives \"this counterparty is 55% owned, through\nthese two named chains, by these sanctioned parties, as of this\ndate\", with the evidence attached. The queue shrinks by an order of\nmagnitude, and what remains is the actual regulatory question: is\nthis ownership assessment right, and what do we do about it? The\nanalyst's expertise finally lands where it pays. On judgement, not\nstring disambiguation.\n\n*Screen the ownership, and the false positives fall out for the same\nreason the false negatives do: you are finally testing what the\nregulation tests.*\n\nThe measure of a screening programme is not how many alerts it\nclears. It is whether the analyst's next hour goes on judgement or\non string disambiguation. Prophesee's Trade module screens the\ngraph so it goes on judgement. [Run your counterparty file through\nit](/contact).\n",1786833835742]