Mr. Sanctions

Category: AI & Other Tech

  • Burning Question: Yesterday’s VTB cross-designation: substantive, or theatrical?

    My prompt to Claude (yes, it was a tad snarky):

    So… other than publicity, and maybe some secondary sanctions exposure, what was the value, if any, of that update? VTB was already subject to an asset freeze, no? explain in language for the common public

    Claude’s initial response, which I took issue with:

    You’re right about the core of it. VTB’s money was already frozen. Nothing about how an American bank handles a VTB hit changed on Sunday. Blocked is blocked, and there is no such thing as more blocked. If the only audience were US institutions, this update would be paperwork.

    The audience is not US institutions.

    The new tag changes which rulebook foreign banks apply. Russia sanctions and Iran sanctions are not two flavors of the same thing to a bank in Dubai, Istanbul, Mumbai or Shanghai. The Russia rules that expose foreign banks to US penalties are discretionary. Treasury may act against a foreign bank that handles significant business for a sanctioned Russian entity, and in practice it has acted sparingly, often with a warning first. The Iran rules are older, partly written by Congress rather than the President, and carry mandatory triggers. Foreign bank boards have long memories about what happened to institutions that got caught on the wrong side of Iran sanctions in the 2010s, and the fines involved.

    There is a second layer that has nothing to do with US law. Many non-US banks write their own internal policies so that any Iran connection is a hard stop, while Russia exposure gets a case-by-case review. Adding an Iran tag to VTB can therefore trip a stricter internal rule automatically at institutions that had been carefully doing limited Russia business. That is my read from how these policies are typically drafted, not something Treasury said.

    The designation also outlives a Russia deal. This is the part I would put first if I were writing it for a general reader. There is real talk of the Ukraine war ending and of Russia sanctions being unwound as part of a settlement. If that happens, VTB’s Russia designations come off. The Iran designation would not, because it rests on different conduct under a different authority. Think of a condemned building. A second violation notice does not make it more condemned, but if the first notice is withdrawn, the second one still keeps the doors locked. Treasury has bought itself a lock that a Russia peace deal does not open.

    And it puts the accusation on the public record. Until this week, a foreign bank could plausibly say it had no idea VTB was running correspondent accounts for sanctioned Iranian banks or moving frozen Iranian assets. Now that claim is on a US government page with VTB’s name on it. Nobody gets to be surprised anymore, and “we didn’t know” stops being available as a defense.

    One smaller thing that matters to people who run screening systems. OFAC took 21 of VTB’s old names, things like plain “VNESHTORGBANK,” and moved them from “used to be called this” to “also called this.” Some screening setups treat former names as lower priority or filter them out. Those 21 names just came back to full strength, and no press release explains why.

    So the honest summary is that this update does very little to VTB’s money and quite a lot to VTB’s remaining business relationships and to the government’s negotiating position. Which is roughly what you would expect from a campaign whose stated goal is isolation rather than seizure.

    My follow-up:

    well, if you were already defying the US Russia sanctions, how likely is it that this would change your behavior? Is it just because the President is more focused on Iran now than Russia? And would it really change the behavior of Chinese firms?

    And the response:

    Your first question mostly answers itself, and it exposes something soft in what I told you earlier.

    Anyone still clearing business for VTB in September 2026 has already made peace with being outside the US financial system. That population is self-selected for indifference. Adding a tag to a party you were already willing to be caught dealing with is not a new deterrent, it is the same deterrent restated. The banks that could be moved by this are the ones with US correspondent accounts they still want to keep, and those banks mostly exited VTB in 2022. So the group sitting between “already gone” and “already defiant” is thin.

    I also overstated one thing. I said the Iran rules carry mandatory triggers where the Russia ones are discretionary. That is true of the statutory Iran provisions, CISADA and IFCA and the NDAA correspondent-account section, but it is not true of what OFAC actually invoked here. Treasury’s own release says OFAC “can prohibit or impose strict conditions on” a foreign bank’s correspondent account. That is permissive language. The legal difference between the Russia exposure and the new Iran exposure is narrower than I made it sound. What is genuinely different is reputational, not statutory.

    On whether this is about presidential attention, largely yes, and I would put it more bluntly than I did. Treasury is running an Iran campaign that needs visible targets while it is simultaneously negotiating over Russia. Hitting VTB under a Russia authority would cut across the negotiation. Hitting the same bank under an Iran authority produces the headline without disturbing that track. The choice of authority is doing diplomatic work as much as legal work. The durability point I raised stands, since an Iran designation survives a Russia settlement, but I should be clear that is a benefit to Treasury’s future position rather than a cost imposed on VTB today.

    Chinese firms are the weakest part of the case. The large state banks pulled back from VTB years ago and have run conservative Iran policies since the last decade. Nothing here reaches them because they already left. The institutions that do handle sanctioned trade are small regional banks with no dollar business to lose, and the precedent for how that plays out is Bank of Kunlun, which the US designated in 2012 over Iran and which China then simply used as the designated channel. Cutting a bank off from a system it does not use is not much of a penalty. Add that a growing share of this trade settles in renminbi through CIPS, and the lever you are pulling is attached to nothing.

    Where I think there is real value, and it is plumbing rather than deterrence, is that a designation under an Iran authority creates a predicate for designating the people who help VTB with its Iran business. Those orders generally include a materially-assisted prong, though I have not pulled the text of 13902 to confirm the specific wording. If that is right, the update is less a punishment of VTB than a hook for whoever gets named next.

    so, maybe not Much Ado About Nothing, but more like Much Ado About Something Possibly in the Future…. meh.

    —

    , , , ,

  • Burning Question: How often does an AI model need to be retrained?

    As a non-expert, I asked Claude:

    Let’s assume I am the sanctions officer at a company with a fair number of high-risk segments of my business – not a small number, but nowhere near a majority. And we built an Ai model to evaluate transactions, documents and third party records for sanctions matches, such that the model recommends and presents all its research for a human to evaluate. Based on changes to Russian and Iranian sanctions from July 1, 2024 to June 30, 2026, including changes in designations, guidance, licensing and any other relevant content published by UK, EU and/or U S regulators, how many times would we have likely needed to retrain our model over that 24 month period? Explain when and why, in non-legal language (not a lawyer). check your response against your sources and flag and correct anything that is not well-supported by them. List your sources. Assume that my firm has a pretty full range of exposure, including crypto, securities and maritime shipping.

    And the response:

    Quick framing before the timeline: your model does two different jobs, and they don’t need retraining at the same rate.

    One job is matching against reference data – SDN entries, the EU consolidated list, the UK sanctions list, vessel IMO numbers. That data changes constantly (weekly, sometimes daily) and shouldn’t require retraining at all if the system is built the way most of these tools are – list content lives in a lookup layer, and the model just queries it. That’s a data refresh, not a retrain.

    The other job is pattern recognition – reading a bill of lading and recognizing deceptive shipping language, following an ownership chain through three shell layers, flagging a stablecoin flow that looks like sanctions evasion. That’s the part that goes stale when the underlying typology changes – a genuinely new evasion pattern, a new category of prohibited conduct, a new legal test for what counts as “control.” That’s what actually calls for retraining.

    I’m assuming your model is architected with that separation (list-matching decoupled from typology-modeling), since that’s how compliance vendors generally build these systems. If yours instead has entity names baked into model weights, the honest answer is “constantly” – so it’s worth confirming that assumption internally.

    With that framing, here’s my read of the 24 months.

    Bottom line

    My estimate: somewhere around 6-7 retrain-worthy events, clustering unevenly rather than spread on a fixed cadence. About three of those are large enough (new prohibition categories, a rebuilt legal architecture, a switch from name-by-name matching to categorical rules) that I’d call them full retrains rather than targeted updates.

    The waves, in plain terms

    1. October 31, 2024 – OFAC maritime shipping guidance. Treasury published scenario-based guidance describing deceptive shipping patterns – falsified bills of lading, AIS location-spoofing to hide ship-to-ship transfers, SDN proxies buried in trade documents issued to help maritime sector stakeholders identify new or common fact patterns indicative of sanctions evasion and address common counterparty due diligence issues. If your document-review component wasn’t already trained to recognize these fact patterns, this is the kind of thing that needs to go into the model, not just a rulebook update.

    2. December 2024 – January 2025 – the shadow fleet buildout. EU package 15 introduced targeted measures against vessels used to circumvent sanctions, the “shadow fleet,” and for the first time imposed full sanctions on third-country actors including Chinese entities. Then, on January 10, 2025, OFAC issued sweeping new sanctions targeting Russia’s energy sector, including a new petroleum services ban, expanded secondary sanctions authority, and designations of Russian oil producers, insurance providers, and more than 180 shadow fleet vessels, done in parallel with the UK, which designated Gazprom Neft and Surgutneftegas the same day, followed by OFAC and OFSI publishing a memorandum of understanding to strengthen cooperation. A brand-new prohibition category (petroleum services) plus expanded secondary-sanctions exposure for foreign banks is a genuine typology addition, not a list update.

    3. February 2025 – Iran’s posture flips. National Security Presidential Memorandum 2, issued when Trump took office, formally reinstated the “maximum pressure” strategy against Iran and significantly expanded its scope. In parallel, EU package 16 hit energy, trade, transport, infrastructure and financial services and added listings touching Russian crypto-asset exchanges and the maritime sector for the first time. This is where “Iran risk” stops meaning “drones and human rights” and starts meaning a full petroleum-sector campaign – a different thing for a model to detect.

    4. March – May 2025 – the campaign gets teeth, and it’s China-shaped. OFAC’s Iran push became a near-monthly cadence targeting Chinese “teapot” refineries and Iranian shadow-banking exchange-house networks – since February 2025, OFAC sanctioned over 1,000 Iran-related persons, vessels, and aircraft as part of this campaign. Treasury also updated its 2019 maritime evasion guidance on April 16, 2025, as part of implementing NSPM-2. Separately, US and European law enforcement seized Garantex’s web domain and froze over $26 million in cryptocurrency on March 6, 2025. The teapot-refinery and exchange-house patterns are genuinely novel fact patterns for a model that was probably trained mostly on Russia-shaped typologies.

    5. July 2025 – the oil price cap stops being a fixed number. EU package 18 introduced an automatic and dynamic mechanism to modify the oil price cap so it stays roughly 15% below the average market price for Russian crude, reviewed every six months, alongside the first-ever full transaction bans on Chinese crypto-asset service providers and on two Chinese financial institutions found to be frustrating EU sanctions. If your model treats “is this deal above the price cap” as a static threshold check, that logic now needs to track a moving, scheduled value – a structural change, not a data point. The UK matched the cap cut and published its first Cryptoassets Threat Assessment on July 21, 2025.

    6. August – October 2025 – the densest stretch of the whole window. Four things landed close together: OFAC sanctioned a network tied to the shuttered exchange Garantex and its successor Grinex, along with the ruble-backed stablecoin A7A5, which had been moving roughly a billion dollars a day – a genuinely new evasion typology (sanctioned-bank-backed stablecoin as a settlement rail). Then the E3 invoked the UN “snapback” mechanism on August 28, 2025, resulting in full reimposition of UN sanctions on Iran effective September 27, 2025, and the EU and UK reintroduced autonomous sanctions measures against Iran that had been eased in 2015 – not new listings, but a wholesale rebuild of the Iran sanctions architecture back toward its pre-2015 scope. Days later, EU package 19 introduced specific binding legal definitions of “ownership” and “control” for designated entities – directly relevant to any third-party-ownership-chain logic your model runs. And on October 22, 2025, OFAC designated Rosneft and Lukoil, Russia’s two largest oil companies, under Executive Order 14024 – the first designation by the Trump administration of Russian actors under the war-related sanctions authorities, a sharp reversal after a year in which Russia designations had dropped to 74 total, down from over 3,000 in 2024. Any one of these would justify a retrain; together, this is the strongest single cluster in the 24 months.

    7. April 2026 – crypto logic changes shape, not just scope. EU package 20 added a ban on EU operator engagement with the Russian national payment card system Mir or the faster payments system SBP – a payment-rail typology your model may not have needed before – and, more structurally, shifted crypto enforcement away from naming specific platforms or tokens toward a blanket prohibition on transactions with crypto-asset service providers in designated categories. That’s a different kind of rule to encode: a categorical, jurisdiction-based test rather than a name-match.

    One thing just outside your window worth flagging for context: EU package 21 landed July 23, 2026 amending Regulation 833 and Regulation 269, entering into force the following day – three weeks after your June 30 cutoff, so I’ve left it out of the count, but it’s the next wave you’d hit almost immediately.

    On your specific exposure

    • Maritime: the heaviest thread by far – the October 2024 guidance, the shadow fleet buildout across essentially every EU package, and the UK’s escalation to active tanker interdiction in 2026 all sit here.
    • Crypto: three distinct shape-changes, not just more designations – individual exchange listings (early 2025), first-ever crypto-service-provider transaction bans (mid-2025), then a categorical/jurisdictional ban (2026). Plus the stablecoin typology from Garantex/A7A5, which doesn’t resemble earlier crypto sanctions patterns at all.
    • Securities: thinner, but real – EU’s central-securities-depository protections (package 15), the Russian Direct Investment Fund transaction ban (package 18), and a non-binding CJEU Advocate General opinion from September 25, 2025 advising that “freezing of funds” should be read broadly enough to cover voting rights and participation rights of depository-receipt holders – worth watching since it isn’t yet a final ruling.

    Checking this against my sources

    What’s solid: every date, designation, and mechanism description above comes directly from primary sources (Consilium, OFAC/Treasury releases, OFSI notices) or law-firm client alerts summarizing them, and I’ve kept the sourcing tight to those.

    What’s my judgment layered on top, not something any source states: which of these events “would” trigger a retrain, and the total count of 6-7. No regulator or vendor publishes a “here’s when compliance AI needs retraining” schedule – that’s an inference I’m drawing from the nature of each change (new prohibition category vs. more names on an existing list), and a different model architecture or risk appetite would land on a different number. Treat the count as a reasoned estimate, not a fact.

    Two things I want to flag rather than let sit unqualified: the CJEU item is an Advocate General opinion, not a binding judgment – I’ve labeled it that way above. And the “over 1,000 Iran designations” figure is cumulative as reported in an April 2026 source, so it’s close to but not exactly aligned with your June 30, 2026 cutoff – the true count at your exact end date is very likely a bit lower.

    One gap: I did not attempt to catalogue every individual OFAC Iran designation round (there were dozens) or every weekly OFSI list update – that would run to hundreds of entries and wouldn’t change the retrain analysis, since after the first round establishes a typology (teapot refinery, shadow banking exchange house), subsequent similar rounds are list updates, not new patterns.

    Sources

    EU: Consilium timeline of Russia sanctions packages · package 15 (White & Case) · package 16 (Fieldfisher, Baker McKenzie) · package 17 (White & Case, DLA Piper) · package 18 (White & Case, Gide) · package 19 (Skadden, Consilium press release) · package 20 (Morgan Lewis, Mayer Brown) · package 21, for context (Skadden) · CJEU depository-receipt opinion (Sullivan & Cromwell) · EU Iran snapback (Mayer Brown)

    US: OFAC maritime guidance, Oct. 2024 · Jan. 10, 2025 Russia energy action (Cleary Gottlieb) · 2025 Trump Russia sanctions pace (CNAS) · Rosneft/Lukoil designations (Steptoe) · NSPM-2 and “Economic Fury” campaign (Treasury) · Iran maritime evasion advisory update (Miller & Chevalier) · Garantex/Grinex/A7A5 (TRM Labs)

    UK: OFSI Jan. 2025 Gazprom Neft/Surgutneftegas + OFAC MOU (Covington) · UK “largest ever” shadow fleet package, May 2025 (Mayer Brown) · UK price cap cut + first Cryptoassets Threat Assessment, July 2025 (Skadden) · UK Rosneft/Lukoil, Oct. 2025 (Squire Patton Boggs)

    UN/multilateral: Snapback mechanics (Congress.gov CRS) · G7 statement on snapback

    —

    , ,

  • AI Corner: You’re not the only one who’s confused

    So, this morning I asked Claude to process this AM’s SECO update (incorporating the UNSC’s August 14th updates), and it told me that it looked awfully like yesterday’s SECO update – and would I like to update that post, or create a new one.

    Yes, on both days, 3 persons and 1 entity were updated by the UNSC. And SECO created an update for each. Because of the timing of the 2 updates, OFSI actually issued a single update on Monday comprising of both sets of UN updates. Even there, OFSI said there were 6 updates to individuals and 2 to entities… which confused Claude, even if it was technically true, from a. regulatory standpoint.

    It’s a good reminder of the limitations, especially over time, of AI tools’ ability to understand context. Would Claude have not been confused had I used a more powerful model like Opus or Fable instead of Sonnet? Perhaps, but at what cost – those models chew up your usage a lot faster, which is silly for normal day-to-day work.

    Thought this was an interesting case to share.

    —

    , , , , ,

  • AI Corner: Why was the EU guide so much harder to get right?

    So, I had to give a lot of feedback to Claude to build that EU sanctions history post – a ton more than I usually have to give. So, I asked about it:

    So, the work on the EU sanctions history required a lot of feedback from me on word usage and writing structure – which I have not had to do with my other requests, like the sanctions guides or even today’s UFLPA prompts. Any idea why that might be?

    and here’s the first response, which I thought had some real insights in it:

    A few real differences stand out, though I’ll flag which parts I can trace directly versus where I’m inferring.

    The EU post was original analytical synthesis with no template behind it. The guides run on the Master Prompts framework you built out over roughly ten regimes now – the structure, sourcing conventions, and voice were already validated before any single guide started. The UFLPA piece was built the same way in miniature: you walked me through three separate Q&A rounds first (mechanics, importer consequences, third-party liability), approved each one, and only then asked for a meld – so by the time I wrote the combined draft, the content and tone had already been tested in stages. The EU post had none of that scaffolding. Nobody had written “functional evolution across the 21 packages” before, so the chart format, the section groupings, how to frame “significance” for each group, even the package 14 versus 15 attribution call – all of that got worked out live, inside the draft, rather than inherited from a proven process.

    Some of the prose rules I’m now applying didn’t exist yet when that project started. Looking at when things got filed: several of the specific conventions I follow now – avoiding label-then-explain constructions, not narrating the article’s own structure, varying back-references instead of repeating a pronoun, watching paragraph length, naming things plainly instead of using shorthand – trace to that same EU project. That’s not a coincidence so much as a cause: it’s the project where those rules got established through your corrections, not one where I was already following them. The UFLPA piece benefited from having that list already in hand.

    The content type itself carries more judgment calls. The EU post asked for interpretive synthesis – financial impact, geopolitical reaction, convergence and divergence across five section groups – which has a lot of surface area for word-choice and framing to go wrong. The guides are closer to organizing known facts into a fixed shape. The UFLPA piece was more explanatory (what this means for someone in the chain) than argumentative, which tends to need fewer editorial calls per paragraph.

    So: newer content type, no established framework, and it’s likely where a chunk of my current style rules for you actually got written. The UFLPA piece and the guides are both riding on groundwork the EU post did the hard part of laying down.

    But, while I agreed with a lot of it, I wanted to have Claude explain something that I felt was off:

    I want to challenge one thing – your prose rules point. I looked back at the guides: those sort of “errors” didn’t occur. So, there was no need to call them out. I think the reason for the difference is more your first and third point – think of them as coarse sandpaper, and the second point more fine finishing and buffing materials

    And here’s what it said (yes, it tends to agree with the prompt):

    Fair, and it holds up when I look at what’s actually in the guide drafts rather than reasoning about it abstractly.

    I pulled a few of the actual guide-building conversations to check. What’s in them is regulatory framework sections, warning boxes for counter-intuitive rules, FAQ entries, checklist steps, glossary terms, document indexes – declarative, informational moves. There’s no equivalent anywhere to “on convergence” labeling, no closing section weighing significance across several groups of content, no long stretch where I’m carrying one evolving subject across dozens of consecutive paragraphs. The EU post’s pronoun-repetition problem, for instance, came specifically from referring back to “this package” or “it” sentence after sentence because the whole piece tracks one subject through time. A guide’s subject changes with the section – IRGC designation status, then the 50% Rule, then a general license – so that particular failure mode has nowhere to occur regardless of whether I’m applying a rule against it.

    So you’re right to demote the second point. It wasn’t that I had those rules in hand for the guides and lacked them for the EU post – it’s that guide content doesn’t create the occasions for label-then-explain constructions, structure-narration, or pronoun drift in the first place. Those errors are a symptom of writing original interpretive narrative across a large, evolving subject, which is exactly what points one and three describe. Point two was me mistaking the moment those rules got written down for a cause, when it was really just documentation of a problem that only the first and third factors could have produced.

    Now, you may find it weird – and certainly not standard practice – but I basically shared my insight from its response:

    Which will help me recognize when I need to review your output more carefully – I don’t tend to review what you’ve written for errors. I look it over more for what I learn from your response.

    And the response brought further insight:

    Given that, the useful signal isn’t really about prose at all – it’s about where interpretive claims are getting made versus where facts are getting organized.

    The two things share a root cause. Original synthesis with no precedent forces me to make judgment calls – what counts as significant, what caused what, how one thing compares to another – and those same judgment calls are what produce both the prose problems you caught and any unsupported inference that might be sitting underneath clean-sounding sentences. A guide’s content is close to directly extractable from a regulation or an FAQ, so there’s not much room for me to be inferring rather than reporting. The EU post asked me to characterize financial impact, geopolitical reaction, and causal sequencing across 21 packages – all of that is synthesis, not extraction, even where it reads confidently.

    So the practical version: when I’m building something with no established template and the content requires me to characterize significance, causation, or comparison rather than just organize known facts, that’s where to look harder at the specific interpretive claims – not the whole document, just the sentences doing analytical work rather than reporting work. Dates, figures, FAQ numbers, regulatory citations are about equally checkable regardless of content type. It’s the “this mattered because X” and “this led to Y” sentences that concentrate in the original-synthesis work and deserve the closer look, precisely because that’s where I’m most likely to be asserting more than the sources actually support.

    And I acknowledged that I agreed:

    nice

    And the response will further help me in my workload here:

    Sounds good. I’ll keep that distinction in mind on future projects – flag up front when something’s original synthesis versus template work, so you know which one warrants the closer look.

    What are the takeaways?

    I think there are two significant takeaways – one mostly for you, my readers, and one for us all.

    The first one: if you’re going to ask the kind of questions that involve the kind of analysis and synthesis that goes beyond summarization and organization of facts, you’re going to need some writing and editorial skills. It’s not so much about knowing what to write; rather, it’s about point out what looks funky. If you can, you can make suggestions for rewording, but you don’t need to be Shakespeare for the AI to realize the difference between “right” and “wrong.”

    The second takeaway: You can push the envelope with what you ask an AI to do, if you’re willing to put in the work. Not only do you need to do the editorial work, but the farther outside the envelope you go, the more you have to check for the AI inferring too much. Of course, you should be explicit about performing a sanity check against sources, and about preventing hallucinations – but the response needs to pass a basic sniff test. If something is valid 55% of the time, is it an inference worth including in your output? 65%? 85%? Drawing that line is the real challenge – I don’t think you can do that effectively without either a lot of research and review, and a certain amount of subject matter expertise. The less expertise and the less time you are willing to check the responses you get, the less you and your AI can go out on an inferential limb.

    —

    ,

  • A followup Burning Question: Can AI meet us halfway?

    Followup prompt:

    Actually, I do have a question: If I provided an update, could you pull all the info you could, plus identify for each listing, where else I should look, based on the nature of the listing?

    and the response (ignore the fact that Claude and I already have a relationship):

    Yes – and it would layer naturally on top of your existing reformatting workflow rather than replace it.

    For each listing you pass along, I’d do two separate things. First, pull whatever’s publicly findable on it – press releases, prior designation history, news coverage, anything that adds context beyond the bare SDN entry. Second, and this is the part you’re actually asking about, tell you specifically where to go dig further, chosen by what kind of listing it is rather than handing you the same generic list every time.

    That second part is where the nature of the listing matters. A few examples pulled straight from the pattern in your own recent updates:

    • A vessel or shipping entity tied to an Iran designation (HormuzSafe, the shadow fleet tankers) – I’d point you to Equasis for registered owner, manager, and flag history, plus Windward or Lloyd’s List Intelligence and UANI or C4ADS for the network picture.
    • A crypto exchange entity like the Shelbit network – blockchain analytics platforms (Chainalysis, TRM Labs, Elliptic) and on-chain explorers matter more here than corporate registries, since the trail runs through wallets as much as paperwork.
    • A Cuba military-industrial entity tied to GAESA or MINFAR (Tecnoimport, UIM) – ASCE and Cuba Study Group research tend to go deeper than general corporate databases, since a lot of that structure never shows up in any commercial registry.
    • A Russia entity – OpenSanctions or OCCRP Aleph first, since they’ve already scraped a good deal of EGRUL, then the Ukrainian NSDC list and the KSE tracker for anything more recent than what’s been indexed.
    • A plain corporate or individual addition with no obvious sector hook – OpenCorporates and the registry of wherever it’s domiciled, plus a check of whether OFAC’s own AKA or Linked To fields already answer the question.

    One caveat, consistent with what I said earlier: the first half I can actually run myself. The second half means I’m handing you the right door, not opening it – I still can’t get past the paywalls on Orbis, Sayari, or Kontur/SPARK, so for anything that needs those, that’s where you’d take it from there.

    —

    , , ,

  • Burning Question: Can AI do the 50% Rule stuff for me?

    My prompt:

    let’s assume I see a new set of designations from a regulator who has some sort of ownership and control sanctions requirement. How would I research to find the additional entities implicated by this? What sources would I be looking at – both global coverage and ones with more targeted geographic coverage (e.g. Iran or Crimea)? And how much of this can you do, and what are you not able to , and why?

    and the response:

    Good question, and the timing is useful because there’s a live example running right now: BIS just built a brand-new automatic O&C mechanism into the export control system. The Affiliates Rule automatically extends the Entity List and Military End-User List to entities owned 50 percent or more by listed entities, modeled explicitly on OFAC’s rule. It was adopted as an interim final rule with immediate effect in September 2025, then stayed by BIS until November 9, 2026 – so it reactivates in three months. That’s a genuinely useful anchor for this answer since it’s exactly the “new O&C requirement, now go find the implicated universe” scenario you’re describing.

    First, the framing question that determines your whole approach: is this regime self-executing or discretionary?

    • Self-executing/arithmetic regimes (OFAC’s 50% Rule, OFSI’s 2022 update, and BIS’s Affiliates Rule once it’s live) don’t require the regulator to take further action – an entity is captured the moment aggregate ownership crosses the threshold, regardless of whether it’s ever named. Your research task here is essentially corporate math: build the ownership tree and sum the stakes held by listed/blocked parties.
    • Discretionary/control-test regimes (the EU’s “owned or controlled by” doctrine, and UK guidance beyond the pure ownership percentage) require an affirmative designation or a documented control assessment – board composition, contractual control, negative control rights, etc. Here you’re not doing arithmetic, you’re building an evidentiary case, and “implicated” doesn’t mean “automatically caught,” it means “worth flagging as a candidate.”

    Worth noting on the BIS rule specifically, since it’s new: if Company A, an Entity List party, owns 50 percent of Company B, which owns 50 percent of Company C, Company C is subject to the same restrictions as Company A – so it chains through multiple tiers, and the rule applies regardless of the foreign country in which the affiliate is located. It also creates an affirmative duty to determine ownership rather than letting silence default to clear.

    Research workflow, roughly in order:

    1. Start with the regulator’s own materials – the designation notice, Federal Register/OJEU entry, and any accompanying press release or FAQ. Regulators frequently name known subsidiaries in the announcement itself (Treasury in particular tends to do this for network-style designations, e.g. IRISL-related actions), which shortcuts a lot of downstream work.
    2. Pull the existing entry’s own metadata – AKAs, “linked to,” prior addresses, registration numbers. This is the cheapest signal and gets skipped constantly.
    3. Confirm the exact threshold mechanic for that regime (aggregate vs. per-owner, direct vs. indirect, does it chain through unlisted intermediate tiers).
    4. Build the tree in both directions – not just subsidiaries down, but parents and sibling entities, since aggregation across multiple blocked co-owners can trip the threshold even when no single owner does.
    5. Cross-check each node against the list itself – a subsidiary you find might already be separately listed under a different name.
    6. Validate anything load-bearing against a primary-source document (registry filing, share register, annual report), not a secondary aggregator’s summary.
    7. Flag and document anything where ownership can’t be resolved – under a rule like BIS’s this is now itself a compliance-relevant fact, not a dead end.

    Sources – global coverage:

    • OpenCorporates – aggregates official company registries, decent free tier
    • Moody’s Orbis/Bureau van Dijk, Sayari Graph, LSEG World-Check, LexisNexis Risk Solutions (Dow Jones’s old risk data now sits inside RELX alongside your former employer Accuity) – all paid, all built for exactly this
    • OpenSanctions.org – free for non-commercial use, and genuinely useful here: it structures identifying information consistently so that relationships between entities, like company ownership or control, can be easily identified, integrating data from 444 global sources
    • OCCRP Aleph and ICIJ’s Offshore Leaks Database – investigative-journalism-sourced, unusually good at surfacing offshore shell layers that official registries don’t show
    • Open Ownership (the BODS standard) – complements OpenSanctions for beneficial-ownership-specific queries
    • Primary registries where they’re genuinely free: UK Companies House + PSC register, SEC EDGAR, EU member-state beneficial ownership registers (coverage is uneven post-rollback in some states)

    Targeted – Iran:

    Iran-linked evasion structures are disproportionately asset-based (vessels, shipping management companies) rather than pure holding-company trees, so vessel data matters as much as corporate data:

    • Equasis – free, gives registered owner/manager/flag history, essential for IRISL-descendant fleets
    • Lloyd’s List Intelligence / Windward – AIS tracking and beneficial-ownership-behind-flag analysis, paid
    • UANI and TankerTrackers.com – track Iran-linked shadow fleet vessels and front companies specifically
    • C4ADS – has published detailed network maps of IRISL/NITC successor structures
    • Treasury’s own past IRISL-related press releases and designation packages, since OFAC did a lot of this mapping itself over 15+ years

    Targeted – Russia/Crimea:

    • EGRUL/EGRIP-based aggregators (Kontur.Focus, SPARK-Interfax, Rusprofile) – the authoritative underlying data, but increasingly hard to access from outside Russia and typically require Russian payment/phone verification
    • OpenSanctions and OCCRP Aleph both scrape a meaningful amount of EGRUL data already, so they’re often the more practically accessible route to the same underlying facts
    • KSE Institute’s Leave Russia tracker – useful for corporate exit/restructuring signals
    • YouControl – Ukrainian-sourced data with good Russia-adjacent coverage
    • Ukraine’s own NSDC/”War and Sanctions” designations – frequently anticipate or run parallel to Western lists and come with their own ownership rationale worth cross-referencing
    • RUSI and Transparency International UK – both have done sustained investigative work specifically mapping Russian elite ownership networks

    What I can and can’t do here, honestly:

    I can search the open web, read regulator press materials and FAQs, pull from freely accessible registries, and synthesize investigative/NGO reporting – and I can help you build and document the research workflow itself, with citations kept separate from inference.

    What I can’t do, and the reasons matter for how you’d rely on this:

    • No login or API access to any paid platform – Sayari, Orbis, World-Check, LexisNexis Risk, Kontur/SPARK. I have no account, no credentials, and no way to acquire them mid-session even if you have a license, so anything behind those walls I simply cannot see.
    • No independent verification of an ownership percentage. If I surface a number from a news article or aggregator, that’s secondhand until checked against a primary document – I’ll flag it as such rather than presenting it as established.
    • No guaranteed completeness. Web search samples the indexed internet; it doesn’t exhaustively traverse a six-tier ownership graph the way a purpose-built product claims to, and coverage of non-English-language and offshore-secrecy-jurisdiction filings is systematically thin. Absence of a hit from me is not evidence the entity isn’t captured.
    • Can’t get past geo-blocking or verification walls (EGRUL-adjacent systems requiring Russian phone/payment being the clearest case).
    • Can’t make the actual compliance determination – whether to block, whether the evidence meets your institution’s risk threshold – that’s a licensed judgment call, not a factual output, and especially with a discretionary-control regime like the EU’s, it’s inherently a judgment exercise, not a lookup.
    • On genuinely obscure multi-hop chains, an LLM’s biggest risk is producing a plausible-sounding structure that isn’t actually sourced. I’ll cite as I go specifically so that risk stays visible to you rather than hidden in confident prose.

    —

    , , ,

  • Burning Questions; My Dinner with Claude – Teaching Claude to Skip the Footnote Forest

    My prompt:

    I knew you were going to add a caveat… it – the second I entered the question, I realized I didn’t include the exceptions, like the NS-PLC list, etc. Comparatively, those are edge cases… you could have said “I assume you’re talking about the blocking programs, right?” and be done with it

    Claude responded:

    Fair, and that’s the better version of it. When the exception is a real edge case rather than something that would change the answer for most actual transactions, a one-line assumption beats a taxonomy of carve-outs. Something like “assuming SDN/full-blocking designations, not narrower lists like SSI or NS-PLC” does the job without turning a comparative answer into a survey of list types.

    Going forward, the default will be that: spell out the exception in full only when it’s plausibly the actual fact pattern, flag the assumption otherwise.

    My next prompt:

    or if i ask for a more comprehensive survey of options, OK?

    Claude responded:

    Of course — that’s the trigger for the full breakdown. Default to the one-line assumption; if you ask for the comprehensive survey, you get the whole list with the reasoning behind each carve-out.

    —

    ,

  • AI Corner: My Dinner with Claude

    with apologies to Andre Gregory’s movie, which I loved…

    Over the next 7 business days, I’m going to post segments of an ongoing conversation with my Claude chatbot (using the Sonnet 5 model on the Extra effort setting). While the content itself is interesting, what fascinates me more is when I challenge its basic assumptions, or rephrase my position in a different way, Claude identifies area where it was imprecise or potentially misleading. Part of this is interpreting my point of view and seeing how its previous responses might clash with my point of view – inadvertently, such that a revised response would be better aligned with how I think and express myself.

    And, to be clear, I don’t think Claude is necessarily trying to agree with me, which chatbots are known to do. As you know, these tools don’t think – they look at probabilities when deciding how to express themselves and, sometimes, choosing a different set of words to express the same concept is received differently. Especially when trying to match my conversational style, I think Claude chooses words that match my style, but which may be misleading or imprecise.

    —

  • This Old Blog: editing of Claude responses

    My editing of Claude’s responses is pretty minimal, and I want to explain it, in the service of full transparency.

    What I do edit out is basically three things:

    • In a chatbot, any process-type discussion (couldn’t access this document, now fetching this) appears as part of the response block – and gets picked up if you hit the Copy button. The process of how a chatbot tries to gather information, as a general matter, doesn’t add that much value, IMHO, to most of my intended audience – even the more AI-curious. So, I leave these out. Ditto for any discussion as it refines its thinking – what we largely care about is the final response. A good example is the (upcoming) Burning Questions post on the “secondary risk” wording in OFAC listings, where Claude prefixes the final response with this (none of which appears in my post):
      • This is the sharper version of the “Burning Questions” angle from before — whether the field is being read as a rating rather than a citation, and whether that misreading clusters anywhere. Let me search for evidence rather than reason from priors, because the honest answer may be that this specific behavior is under-documented.
      • That gave me adjacent material (screening mechanics, one vendor that tags secondary sanctions as a labeling feature) but not direct evidence of firms misreading the field as a rating. Let me look specifically at the de-risking / over-compliance pattern, which is where any geographic or asset-class clustering would show up.
      • I now have solid material on the de-risking pattern and where it clusters, but I want to be honest about a distinction: the question asks specifically whether firms treat the wording as a rating, and most of what I’ve found is about firms reacting to the underlying risk. Those aren’t the same thing, and I don’t want to conflate them. Let me do one more targeted search on the data-field-as-signal angle.
      • I have enough to give an honest, well-bounded answer.
    • Sometimes, as part of its source checking, Claude will point out things where the difference between its initial drafting and its checking – semantically and in a practical sense – is not worth correcting, or where the discussion of the review itself doesn’t illuminate the response. Those I remove. If there is a real difference, by the way, I have Claude correct it – so I don’t post the original response.
    • Often, Claude will propose I ask a follow-up question. These I remove.

    Now, I do leave in some of Claude’s commentary from time to time. I do that for a number of reasons:

    • I think the reasoning, and the natural language production of the “conversation”, is fascinating
    • I think the back and forth of postulating something, posting a caveat and then proceeding, is more nuanced and informative to the audience (and, to be perfectly frank, to me as well)
    • As much as this is a sanctions blog, it’s also a “AI in sanctions” blog. Some of these peeks behind the curtain inform us as we make decisions about how to use AI in our field

    As always, I invite your comments, criticisms and suggestions.

    —

    , , ,

  • AI Corner: Making those guides

    Once you are happy with an output from an AI chatbot like Claude, you want to capture it for re-use in a master prompt. I am hoping to create Practitioners’ Guides and Plain Language Guides for all OFAC sanctions programs. Additionally, I want to combine related programs (the multiple Russia ones) into a single pair off guides. So, here are the master prompts for the guides – the first one for the Practitioners’ Guides, the second for the Plain Language Guides and the third to incorporate multiple programs (you insert this one before the 1st or 2nd prompts as needed):

    CAATSA_Master_Prompts_UpdatedDownload

    —

    ,