A sanctions list is a set of names. Screening checks a name against the set. That is necessary and it is the whole of what most tooling does, which means the questions that actually matter to an analyst, questions about structure, time and absence, are asked by hand, one at a time, if at all. Who is the entity nothing has been said about, but whose removal would fragment the network? Which companies are owned by a designated parent and have never been listed themselves? Who held assets that later became designated, and what did they do with them the day before?
In May 2026 we loaded the full public OpenSanctions corpus, roughly 4.28 million records across twenty-nine entity types from hundreds of national and international sources, into 8DB and asked those questions of the whole structure at once.1 This article reports what came back, and it spends as much space on why the findings can be trusted as on the findings themselves, because intelligence that cannot survive adversarial review is opinion.2
Three questions the lists cannot answer
Where is the keystone? Treat the corpus as a graph and compute which nodes hold it together: not degree, which rewards the visible apex, but spectral centrality and the fragmentation caused by removal. In the Russia-related sub-network of roughly 3,900 nodes, one state property-management agency sits at maximum centrality, and removing it splits the network into 937 additional disconnected components, more disruption than the next hundred highest-degree entities combined. That agency is designated by the European Union, the United Kingdom, Canada, Switzerland, Japan and Australia. At the time of the analysis it was not on the United States list. The same method surfaced a cohort of 924 small federal state enterprises, individually unremarkable and unlisted everywhere, that together form the network's load-bearing bridges. Neither finding is a name match. Both are properties of the shape.
Who is connected to a designated parent and absent from every list? Ask the corpus for entities recorded in the Russian company register as subsidiaries of a heavily designated defence manufacturer, then check each against the raw bytes of the current United States, European Union and United Kingdom lists by exact registration number. Seventeen subsidiaries came back, seventeen were confirmed absent from every list, with zero false positives, and each check was emitted as a signed verification record with the source files, their download hashes and the analyst's identity. One was flagged as contested by the analyst and recorded as such, because disagreement is also evidence.
Who moved the assets? Cross the corpus with a fresh copy of the United States list, resolve the same principals across both by content identity despite differing identifiers, then walk ownership as a structure with validity windows rather than as rows. One pattern that fell out: an individual on no list who held all of fourteen Arctic mining and energy companies for sixteen months and transferred every one of them, on a single day, to an entity that is designated. Another: an individual on no list acquiring stakes in a military-electronics group from two designated principals in the same window, alongside co-acquirers who are themselves designated. These are the shapes of nominee ownership and evasion by fragmentation, visible only when ownership is a traversable structure with time on it.
Why the findings hold up
Everything above would be worthless if it depended on trusting us, so the analysis was built to be checked by someone who does not.
- Every record and every finding is content-addressed. A verification record's identity is the hash of its content; re-deriving it on any machine gives the same identifier, and tampering changes it.
- Every finding carries its provenance chain: the source records, the primary-source files with their hashes and download times, the confidence vector on each input, the governance state, and a signed anchor.
- The pipeline is byte-deterministic. Ingesting the same corpus snapshot on two different machines produced identical determinism signatures. A reviewer given the binary, the snapshot and the parameters reproduces every finding exactly. There is no model card and no asterisk.
- The spectral computation is mathematics, not a score. On a sixty-entity sub-network we ran the substrate's verified-bound spectral engine and an independent classical solver on the same operator; the substrate's certified upper bound enclosed the classical residual with a wide margin.3 The bound is conservative and that is stated. An analyst can rerun both.
- Negative results are recorded. The one contested subsidiary, the loose bound, the fact that the full-corpus node count includes raw edge endpoints and therefore overstates canonical entities: each is written down beside the finding it qualifies.
Why this is a data-layer story
Nothing here is a clever model. The keystone finding needed eigendecomposition over the whole corpus as a first-class operation. The subsidiary finding needed absence to be a queryable state rather than a null. The nominee finding needed ownership as an n-ary structure with time, and identity resolved across datasets by content. Those are properties of the store. Screening tools, graph databases and language-model stacks each lack at least one of them, which is why the questions had gone unasked.4 When the substrate carries structure, time, absence, confidence and provenance together, the analyst's questions become queries, and the answers come with their proofs attached.
What to bring us
- A corpus you already hold. A sanctions, ownership or registry dataset, or the hardest cases from one. We will load it and return the structural findings with their provenance.
- A question you cannot ask today. The keystone, the unlisted subsidiary, the nominee pattern, in your domain. We will show whether the store can express it and what the proof looks like.
- Your reviewer. Give us the analyst who will try to break the findings. Byte-for-byte reproduction is the first thing we will hand them.
Write to hello@8braid.com with the subject "Structural sanctions analysis: corpus review". You will hear back from an engineer.
Sources and further reading
- OpenSanctions: open data on sanctions, politically exposed persons and persons of interest
- OFAC: the 50 Percent Rule
- 8DB: Entity Resolution That Shows Its Work
- 8DB: Show Your Work: Provenance as a Query
All sources used in the analysis are public. Findings about unlisted entities describe structural properties of public data and are not allegations; designation decisions rest with the relevant authorities.
Footnotes
-
Throughput and memory figures from the campaign are withheld from this article because they are not yet packaged as reproducible benchmarks with the fixtures and environment records our own doctrine requires. ↩
-
Private individuals are deliberately not named anywhere in this article. The findings were delivered as referral cards to the appropriate authorities, with their evidence, which is where designation decisions belong. ↩
-
The verified-bound engine has not been exercised at sub-network sizes larger than the sixty entities tested, and at that size its bound is loose by a large factor. ↩
-
No head-to-head comparison against established entity-resolution or sanctions-screening products has been run; the statement about what each class of tool lacks describes capabilities, not measured performance. ↩
Continue the technical conversation
Where could this help your work?
Bring a research question, a database workload or an application you want to build. Let’s connect the ideas in this article to an evaluation that matters to your team.
Discuss this work