Blueprint · design and status
Every U.S. sanction since 2000, resolved to industries, products, and direction of restriction.
The study develops a scalable, LLM-empowered multi-agent framework to delineate the influence of each sanction the United States has enforced since 2000 — which countries were directly targeted, which industries and product categories in a sanctioned country were affected, and whether each product category was restricted on the inflow or the outflow side.
The unit of observation
One record per sanction–year–country–industry–product cell
The database unfolds each sanction into a panel that global-value-network research can join directly to trade and production data, with every cell bound to the official passage that evidences it.
- SanctionLegal instrument · EO, statute, rule
- Year2000 – 2025
- CountryDirectly targeted state
- IndustrySIC, 4-digit
- ProductHS, 4-digit
- DirectionInflow / outflow
Deliverables
A database, and the workflow that builds it
1 · The sanctions exposure database
A comprehensive, off-the-shelf database of U.S. sanction impositions since 2000 in the sanction–year–country–industry–product structure, capturing the implications of U.S. sanctions for every participant country in global value networks — built from the complete Federal Register record rather than from any secondary compilation.
2 · The scalable agentic workflow
A documented, replicable method for delineating country-specific economic exposure to sanctions with multi-agent LLM systems: the screening-gate architecture, the dual-vendor verification loop, and the governance protocol that keeps every machine decision auditable and every value-laden choice with the human investigator.
Table 01 · Program status
The corpus lane, stage by stage
Status reflects the program record as of early August 2026. A stage is marked complete only where a verified, retained artifact supports it; no screening gate has yet executed over the corpus.
| Stage | Where it stands | State |
|---|---|---|
| Frozen source frame | 788,470 Federal Register publication records, 2000–2025, collected with archived query provenance; 26/26 annual source audits passed. | Complete |
| Canonical evidence projection | An 11.99 GB canonical projection of the frame, validated exhaustively by two independently built validators from two model vendors — field-for-field identical reports, zero defects. | Complete |
| Gate 1 — metadata identity | Formal specification frozen after adversarial review; fixture pilot passed with zero failures across 21 fixtures / 28 run groups. Corpus execution awaits its authorization chain. | Frozen · pilot passed |
| Gate 2 — deterministic signal screen | Design under way in recorded sittings with the PI: the country vocabulary is ruled (“fullest possible set” — full ISO structure, historical names, demonyms), the economic-vocabulary scope ruling is pending, and pile-size measurements precede any spend. | In design |
| Gate 3 — issuing-agency gate | Dual-vendor saturated research on agency histories: round 1 complete on both blind seats, synthesis complete; two registry conflicts await PI ruling. | In research |
| Corpus text acquisition | A supervised, resumable crawler fleet is acquiring verified document text for the ~656,000 records without a locally verified text artifact — over 80,000 documents captured at the current pause point, under a PI-governed rate policy with hash-chained instruction records. | In flight |
| Gates 4–7 — series, families, complete reads | Designed in the ratified gate ladder; formal specification follows Gate 1 closure. Complete-read gates are priced from live per-document read costs measured in the companion lane. | Designed |
| Final coding | Construct criteria ratified (v1.1, August 2026) after dual-vendor research, cross-seat synthesis, and PI governance review. Coding into the target record structure begins once the gates close. | Criteria ratified |
The companion lane
A sealed reading contract, proven live on both vendors
A deliberately separate second project develops the reading machinery on a hand-collected corpus of 401 sanction documents: a sealed dual-vendor reading contract under which two models read the same page blind and must agree — with typed uncertainty, bounded retries, and refusal handling — before any text is accepted. Its live pilot has completed end-to-end on both vendors, and its measured per-document read costs price the complete-read gates of the corpus lane. The two lanes share no working state; artifacts cross only as exact-hash pins.
Proof of concept: the JIBS revision database
Applying an earlier iteration of this framework, the team has already produced a lesser version of the sanctions database in support of a revision at the Journal of International Business Studies. That exercise established that the pipeline runs end-to-end — from official text to coded, evidence-linked sanction records — and motivated the scaled, gate-governed rebuild documented on this site.
Boundary
What this site does and does not claim
This is a research preview. Aggregate counts appear only where a retained, audited artifact supports them; no dataset, row-level ledger, or API has been released, and screening results will be published only after the gates close under their pre-registered acceptance statistics and the principal investigator’s countersignature.