Incumbents map your supply chain and raise an alert. Command Instinct resolves the specific shipment at the transit node — the correct per-license test, the pass/fail, and if it fails, the compliant substitute that moves instead.
Describe a shipment. The engine selects the correct China-exclusion test for the governing license, checks the authorized party, contract venue, joint venture structure, diluent partition, payment channel and vessel status, and returns a ranked execution matrix. The verdict logic is the real per-license standard, not a demo stub. The panel beside it meters each decision on our RTX 4080 node: a local model writes the verdict's audit memo, and the panel shows the energy that work drew from the GPU.
Under the August and September 2026 licenses the PRC exclusion takes two forms, and neither is an origin test alone. A screen that only checks component nationality under-blocks both: a PRC-owned distributor shipping non-Chinese parts is barred under every categorical license, and under the nexus licenses wherever it is a Venezuelan or U.S. entity. It also over-blocks the nexus licenses, where PRC origin by itself is not a trigger. The engine branches on which test the governing license actually applies.
Any PRC person is barred outright, and so is any entity, wherever organized, that a PRC person owns, controls or shares a JV with. Origin screening catches the first half. Ownership tracing is needed for the second.
PRC origin is not a trigger. Barred only where a Venezuelan- or U.S.-organized entity is owned, controlled by, or in a JV with a PRC person. Requires beneficial-ownership data, not a bill of materials.
GL 47B (diluents) bars neither China nor Russia — the narrowest screen of the set, and the only lane diluents may legally travel.
Both China tests above turn on who ultimately owns a supplier — OFAC's 50% rule, counting direct and indirect ownership together. In August 2026 the one U.S. government registry that could answer that was permanently shut down for U.S. companies. The requirement stayed; the data left.
FinCEN permanently eliminated beneficial-ownership (BOI) reporting for U.S. companies and will delete the records already filed. The EU's public ownership registers were restricted by court order back in 2022. No public dataset now resolves layered ownership for the 50% rule — exactly the data both China tests need.
The engine is being built to trace the ownership graph layer by layer, the way OFAC does: once blocked persons own 50% or more of an entity, that entity is blocked and its own stakes count in full at the next layer, aggregated across paths. That is the capability the vanished registry can no longer provide. In this prototype, ownership is entered by hand in the form above; verified ownership data will plug in from a specialist source (Sayari, Kharon) behind it.
Source: U.S. Treasury / FinCEN final rule of Aug 11, 2026 (BOI reporting eliminated for U.S. companies). Decision-support, not legal advice.
Power is now a binding constraint on AI. Each verdict on this page runs real work on our RTX 4080 node: a local language model writes the audit memo, and the node meters that work from the GPU itself, in watt-hours per verdict, per thousand tokens and peak draw. CO₂ uses a sample grid factor until a live marginal emissions feed is connected. When the node is offline, the panel shows a representative curve, labelled Sample.
Each verdict records what it cost to compute, so a compliance team running thousands of shipment checks sees the energy of its own AI decision by decision, not as one line on a cloud invoice.
The power behind AI data centers runs through equipment and minerals with the same exposure this engine already resolves: who owns the supplier, where the part came from, and whether a listed party sits in the chain. Small modular reactors follow the same path.
When the panel badge reads LIVE, the readings are measured on our GPU node for the verdict you just ran. When it reads Sample, the node is offline and the curve is representative.
Does a small AI model on our GPU call the winner of a Valorant round better than a simple rule, and what does each call cost in energy? Take a round from a real ranked match. The model on our RTX 4080 node calls it, answering directly or reasoning first, and the node meters that work the same way it meters the verdicts above.
| Team | Side | Rounds won | Loadout | Banked |
|---|---|---|---|---|
| Blue | · | · | · | · |
| Red | · | · | · | · |
| Who calls it | Right | Wh per call |
|---|---|---|
| Rule: bigger buy wins | 60% | none |
| qwen2.5 3B | 54% | 0.0008 |
| qwen2.5 14B | 55% | 0.0030 |
| qwen2.5 3B, reasoning | 54% | 0.0117 |
| qwen2.5 14B, reasoning | 59% | 0.0450 |
50 recent NA ranked matches, Oct 28, 2026. The 3B used about 74% less energy than the 14B for the same accuracy. Reasoning first cost about 15 times the energy: the 3B gained nothing, and the 14B rose from 55% to 59%, still short of the rule. Measured on the RTX 4080 node: GPU board power every 100 ms, idle subtracted, 1,067 calls back to back. A single live call reads higher, since the node keeps its meter open 0.6 s after each call to catch the GPU's reading lag.
Each dot is a model's share of rounds called right, with its 95% range; the shaded band is the rule's. Energy is measured per call on our RTX 4080 node, on a log scale: a 14B reasoning call uses about 56 times the energy of a direct 3B call. The table above has the same numbers.
| Who calls it | Right | Wh per call |
|---|---|---|
| Rule: team ahead in gold wins | 77% | none |
| qwen2.5 3B | 73% | 0.0008 |
| qwen2.5 14B | 77% | 0.0035 |
| qwen2.5 3B, reasoning | 50% | 0.0116 |
| qwen2.5 14B, reasoning | 76% | 0.0410 |
The question: from a ranked game's state at 15:00 (gold, kills, towers, dragons, void grubs, heralds, levels and creep score), which team wins. 500 recent NA Challenger games, Oct 29, 2026; 494 lasted past 15 minutes. The 14B only matched the gold rule, and the 3B came close for about a quarter of the energy. Reasoning first cost 12 to 14 times the energy and gained nothing; the 3B, reasoning, picked blue in 88% of games, a lean it kept even though its answer line names no color. Measured the same way as Valorant, on the same GPU.
The same 1,067 rounds, leaving out 12 with equal buys. Below 5,000 credits the bigger buyer wins about half the time; above 10,000 it wins about four rounds in five. Each dot has its 95% range: the dip at 1,000 to 2,500 credits is in the data, but its range is wide.
| Valorant Radiant, wins this act | Players | Median rating | Rating per win | Hours played, est. | GPU energy, est. | Rating per kWh |
|---|---|---|---|---|---|---|
| 10 to 22 | 60 | 332 | 20.2 | 18 | 1.1 kWh | 299 |
| 22 to 31 | 61 | 329 | 13.2 | 27 | 1.7 kWh | 194 |
| 31 to 40 | 60 | 328 | 9.5 | 38 | 2.4 kWh | 139 |
| 40 to 55 | 61 | 348 | 7.7 | 48 | 3.0 kWh | 116 |
| 55 to 164 | 61 | 417 | 5.7 | 72 | 4.5 kWh | 92 |
Oct 29, 2026, NA. Valorant: all 303 Radiant players this act, in five groups by wins; hours are estimated from wins at an assumed 55% win rate and the 34 minute median of 30 recent ranked matches. GPU energy is those hours times 63 W, the average our RTX 4080 laptop GPU drew over one 28 minute Competitive match: GPU only, not the whole PC or monitor, and one machine, so an estimate for anyone else. The lightest Radiant group gets about three times the rating per kWh of the heaviest. Groups only: no player is named.
Research demo, not for profit and not part of any product. Match data comes from the Riot Games API as team totals from recent ranked matches (per round in Valorant, at 15 minutes in League of Legends), with no player data. This demo isn't endorsed by Riot Games and doesn't reflect the views or opinions of Riot Games or anyone officially involved in producing or managing Riot Games properties. Riot Games, and all associated properties are trademarks or registered trademarks of Riot Games, Inc.
Exiger, Interos, and Altana answer “is there risk somewhere in my chain?” Command Instinct answers “does this shipment clear this license right now, and if not, what moves instead?” Complementary altitude, not a competing one.
A deterministic routing decision at the transit hub — substitute part, hub location, freight capacity, revised transit time — where incumbents stop at a warning.
The same node flags sanctioned components and compromised or untrusted-OEM hardware — the second axis that closed-loop telecom routing depends on.
Categorical vs nexus China test, the established U.S. entity gate, dispute venue, new JV bars, diluent partition, FGDF payment channel and reporting cadence. Encoded from the license text, not approximated.