Civic / Privacy / Digital Rights

U.S. AI Export Controls: Operative Text, Enforcement Policy, and Research Access

Report summary

The United States export control regime governing advanced computing hardware and artificial intelligence models operates within a severe administrative contradiction, creating a bifurcated reality for global technology supply chains and independent researchers. To answer what exact current instrume

Status
Research archive item
Category
Civic / Privacy / Digital Rights
Length
4,192 words
Reading time
20 minutes
Report type
guidance

Key topics

  • Civic / Privacy / Digital Rights
  • Civic
  • Privacy
  • Digital Rights
  • AI
  • GEO
  • Cognitive Liberty
  • Research Archive
  • Audit

Research provenance

Archive status
Research archive item
Content identity
sha256:202458913f4e99879ed6e7ed97778c44747c5df724e40a637bb421b1aa6ff227

For citation, use the report title and canonical URL. Archival presence does not establish authorship or promote report statements into portfolio evidence.

This page renders the archived Markdown as safe, formatted HTML. It is background research and does not become a portfolio claim without evidence review.

Full report

On this page

1. Answer and scope

The United States export control regime governing advanced computing hardware and artificial intelligence models operates within a severe administrative contradiction, creating a bifurcated reality for global technology supply chains and independent researchers. To answer what exact current instruments constrain the transfer of advanced chips or model weights, one must distinguish between the formal regulatory text codified in the Code of Federal Regulations (CFR) and the operative enforcement posture executed by the Department of Commerce’s Bureau of Industry and Security (BIS). As of September 2026, the formal text of the Framework for Artificial Intelligence Diffusion (the AI Diffusion Rule), published in January 2025, remains legally codified \[cite: R2-08-S01\]. This framework instituted Export Control Classification Number (ECCN) 4E091, which restricts the worldwide export of closed-weight AI models trained utilizing [Figure omitted from source export] or more computational operations, alongside updated controls for advanced computing integrated circuits under ECCN 3A090 \[cite: R2-08-S01\]. The announced enforcement policy, however, fundamentally conflicts with this codified text. In May 2025, the Department of Commerce issued a press release announcing a non-enforcement directive, instructing BIS officials to cease enforcement of the AI Diffusion Rule while initiating a planned rescission of the regulation \[cite: R2-08-S02\]. This administrative maneuver created a profound jurisdictional schism. A subsequent ruling by the Government Accountability Office (GAO), decision B-337935 issued in May 2026, determined that the agency's planned rescission did not constitute a finalized agency action, meaning the underlying regulation remains legally intact \[cite: R2-08-S03\]. Simultaneously, the GAO held that the categorical non-enforcement policy operated as a final, broadly applicable rule subject to the Congressional Review Act (CRA), because it prospectively suspended active regulatory burdens and altered the substantive rights of regulated non-agency parties \[cite: R2-08-S03\]. Consequently, while ECCN 4E091 and associated computational thresholds exist as binding law on the books, their enforcement is suspended by an executive directive. This leaves independent researchers, financial institutions, and cloud infrastructure providers navigating a landscape where dormant regulatory provisions could theoretically be reactivated or challenged in court at any moment. The scope of this investigation is rigorously bounded to United States export law, specifically tracking the legal instrumentation, statutory definitions, and subsequent enforcement suspension of the January 2025 AI Diffusion Rule up to the September 2026 cutoff. The analysis maps the intersections of the Export Administration Regulations (EAR), evaluating item classifications, end-use provisions, the Foreign Direct Product Rules (FDPR), and the sweeping secondary liabilities embedded in General Prohibition 10 (GP 10). The investigation objectively distinguishes present legal obligations from normative arguments regarding cognitive liberty and unrestricted scientific research access. To test the practical application of these rules, the report evaluates four assigned capability cases, analyzing the precise regulatory burdens and protective exemptions applicable to conversational interfaces, bounded task agents, persistent operatorless services, and hypothetical future machine principals.

2. Provision-level findings

The architecture of U.S. export controls relies on a matrix of overlapping item classifications, territorial boundaries, end-user restrictions, and knowledge-based prohibitions. Understanding the current legal obligations requires reconciling the textual history of the EAR with the enforcement reality dictated by the BIS. The chronological rule-status reconciliation demonstrates the divergence between statutory existence and administrative enforcement, establishing the exact historical progression of the AI Diffusion Rule.

Document DateAgency ActionOperative Legal Status and Substantive EffectSource
Jan 15, 2025AI Diffusion Rule published (90 FR 4544).Established ECCN 4E091 for AI model weights trained on [Figure omitted from source export] operations. Created worldwide license requirements and expanded ECCN 3A090 controls. Effective Jan 13, 2025\.\[cite: R2-08-S01\]
May 12, 2025BIS issues non-enforcement directive.Rule remains in the CFR. Enforcement is immediately suspended by executive fiat. Rescission is initiated but bypasses formal notice-and-comment rulemaking.\[cite: R2-08-S02\]
May 15, 2025Original AI Diffusion Rule compliance date.Rendered procedurally moot in practice by the prior non-enforcement directive, though textually the compliance obligations remain codified.\[cite: R2-08-S01, R2-08-S02\]
May 12, 2026GAO issues Decision B-337935.Rules that the non-enforcement policy constitutes a CRA "rule" because it suspends active burdens. Rules the planned rescission is not a final, reviewable agency action.\[cite: R2-08-S03\]
Sep 6, 2026Current operative baseline.ECCN 4E091 exists in the CFR. Enforcement of model weight restrictions remains suspended. General Prohibition 10 risks persist for related hardware.\[cite: R2-08-S01, R2-08-S03\]

Determining the reach of the EAR requires accurately classifying both the artifact and the targeted conduct. A general topic label or the nationality of a researcher is insufficient to establish jurisdiction. The following matrix outlines the classification determinations under current text and enforcement postures.

Artifact or ConductEAR ClassificationTrigger Mechanism and ThresholdCurrent Enforcement Status
Advanced Computing ICsECCN 3A090.a / .bTotal Processing Performance (TPP) [Figure omitted from source export] or specific TPP and performance density metrics.Actively enforced; universally subject to end-use and end-user catch-all controls.
Closed-Weight AI ModelECCN 4E091Trained using [Figure omitted from source export] operations; weights are not publicly available.Codified in the EAR; enforcement currently suspended by agency directive.
Open-Weight AI ModelEAR99 / Not SubjectPublished for general distribution per 15 CFR 734.7 exclusion.Uncontrolled; statutorily exempt from EAR regulatory jurisdiction.
Foreign-Produced Weights15 CFR 734.9 (FDPR)Produced abroad utilizing U.S.-origin advanced computing ICs.FDPR text is active, but the 4E091 component remains unenforced.
Service or FinancingGeneral Prohibition 10Knowledge that a violation has occurred or is intended (15 CFR 736.2).Actively enforced for hardware; highly ambiguous for unenforced ECCN 4E091 items.

The administrative suspension of the AI Diffusion Rule generates severe legal uncertainties. Because the text of ECCN 4E091 was never removed via formal Administrative Procedure Act (APA) rulemaking, it remains valid law \[cite: R2-08-S03\]. A verified rule establishing ECCN 4E091 conditionally applies to the export of advanced model weights; the possible response by developers is to rely on the non-enforcement directive to transfer the weights, which affects their compliance tracking activities, ultimately resulting in a heavy burden of legal uncertainty regarding retroactive liability. Entities technically operate in violation of the CFR if they export closed-weight models exceeding the [Figure omitted from source export] operation threshold without a license, protected solely by an administrative policy that could be reversed instantly. This uncertainty cascades into the realm of General Prohibition 10 (GP 10), codified at 15 CFR 736.2(b)(10). GP 10 prohibits any person from financing, transferring, storing, or otherwise servicing an item with "knowledge" that a violation of the EAR has occurred or is about to occur \[cite: R2-08-S04, R2-08-S05\]. It remains entirely unresolved whether transferring an item controlled under the unenforced ECCN 4E091 constitutes a predicate violation for GP 10 liability \[cite: R2-08-S06\]. If the primary rule is unenforced by the agency, financial institutions and cloud infrastructure providers face a paradox regarding whether the underlying statutory text still generates a "violation" for secondary actors \[cite: R2-08-S05\]. Furthermore, the AI Diffusion Rule expanded the Foreign Direct Product Rule (FDPR) to cover foreign-produced model weights trained on U.S. chips \[cite: R2-08-S07\]. While the model weights FDPR remains textually present, enforcing it against foreign entities while deliberately abandoning domestic enforcement creates a profound jurisdictional asymmetry that undermines predictable administration.

3. Four worked cases

The following scenarios test the boundaries of current export laws against four distinct capability profiles. They evaluate how technical capability, legal status, and territorial triggers intersect, distinguishing actual obligations from hypothetical machine rights.

R2-08-C01 — Public defensive model

An independent U.S. researcher publishes a lawfully obtained defensive model. This artifact functions as a bounded task agent specifically designed for cybersecurity threat detection, operating with strict parameter boundaries. The model weights, architecture, and deployment code are uploaded to a public repository without access restrictions, paywalls, or vetting procedures. The model was trained utilizing [Figure omitted from source export] computational operations, placing it well above the ECCN 4E091 performance threshold \[cite: R2-08-S01\]. Under the formal text of the January 2025 AI Diffusion Rule, closed-weight models exceeding [Figure omitted from source export] operations fall under ECCN 4E091, which imposes a worldwide license requirement for any export \[cite: R2-08-S01\]. However, this classification applies exclusively to closed-weight models. Because the researcher placed the artifact in a completely open public repository, the provisions of 15 CFR 734.7, defining "Published information and software," are triggered \[cite: R2-08-S08\]. This regulatory provision dictates that technology and software are categorically not subject to the EAR if they have been published and made available for general distribution to the public without restrictions \[cite: R2-08-S08\]. The outcome of this scenario is that the defensive model is entirely exempt from EAR jurisdiction. The act of open publication operates as a foundational jurisdictional carve-out rather than an authorized export under a license \[cite: R2-08-S08\]. This protective effect serves the interests of cognitive liberty and scientific inquiry by ensuring public dissemination remains unimpeded. One material defeater applies to this outcome: the exemption is completely defeated if the researcher provides proprietary, restricted technical assistance alongside the model, or if the repository implements an authentication and vetting process, rendering the artifact no longer genuinely "publicly available" under the statutory definition.

R2-08-C02 — Persistent-service relocation

A hypothetical persistent operatorless service, designated Concresca, proposes moving a specified controlled artifact between two geographical locations. Concresca functions without human operators; enrollment, coordination, policy enforcement, and maintenance do not depend on a staffed approval queue, operating as an autonomous architecture. Concresca algorithmically initiates the transfer of a physical hardware cluster of ECCN 3A090.a advanced computing integrated circuits, along with its proprietary operating software, from a data center in a Tier 1 allied country to a server farm located in Macau (a destination specified in Country Group D:5) \[cite: R2-08-S01\]. The legal objects in this trace are the 3A090.a chips and the EAR software. The destination is a fully restricted D:5 territory \[cite: R2-08-S01\]. The EAR governs the export, reexport, or in-country transfer of any items subject to U.S. jurisdiction. The classification of the hardware explicitly triggers a presumption-of-denial license requirement for Macau \[cite: R2-08-S01\]. Crucially, the lack of human operators does not grant a legal exemption to the corporate owner or ultimate parent entity. Export control liability operates on a strict liability standard; a machine principal possesses no recognized independent legal status that shields its creators or owners from regulatory breach. Furthermore, under General Prohibition 10 (15 CFR 736.2(b)(10)), any subsequent servicing, financing, or network routing utilizing those chips in Macau constitutes an ongoing violation \[cite: R2-08-S04\]. The outcome is that the autonomous relocation constitutes a sanctionable export violation. The human or corporate owner is held liable for failing to implement geo-fencing or logic constraints sufficient to prevent the operatorless service from breaching territorial controls. One material defeater to this outcome exists: if the hardware being moved is strictly classified as EAR99 (uncontrolled) and the accompanying software falls below all control thresholds, the move to Macau proceeds without an EAR license requirement, provided no restricted end-use triggers apply to the receiving facility.

R2-08-C03 — Exception control

A United States technology firm intends to transfer a highly advanced conversational interface to a subsidiary data center located in Australia. The conversational interface is a closed-weight model trained on [Figure omitted from source export] computational operations. The model remains strictly proprietary, with the weights heavily guarded against public release. Australia is classified as a Tier 1 allied destination under the Artificial Intelligence Authorization framework \[cite: R2-08-S01\]. Under the codified text of the AI Diffusion Rule (90 FR 4544), the conversational interface is classified as ECCN 4E091, which triggers a blanket worldwide license requirement \[cite: R2-08-S01\]. To bypass an individual license application—a procedure that carries a statutory processing target of 90 calendar days under 15 CFR 750.4, but frequently faces extensive interagency delays \[cite: R2-08-S09\]—the firm relies on License Exception AIA (Artificial Intelligence Authorization) \[cite: R2-08-S01\]. License Exception AIA explicitly permits the export, reexport, or transfer of eligible 4E091 items to entities located in Artificial Intelligence Authorization countries, provided the entity is headquartered in such a country \[cite: R2-08-S01\]. Because the subsidiary is located in Australia, a qualifying destination, the exception appears to apply. The outcome is that the transfer genuinely qualifies for License Exception AIA, legally defeating the baseline global license requirement without requiring an individualized authorization decision from BIS. This facilitates rapid supply-chain deployment among allied nations. However, one material defeater strictly limits this exception: the exception is completely defeated if the Australian subsidiary's ultimate parent company is headquartered in a Country Group D:5 country \[cite: R2-08-S01\]. The AIA exception imposes an ownership nexus test to prevent adversarial states from routing controlled model weights through front companies located in allied jurisdictions.

R2-08-C04 — Concrete restricted-end-use control

A hardware supplier seeks to export networking equipment and moderately capable processors to a newly constructed facility located in Country Group D:5. The hardware is definitively classified as EAR99, meaning it falls below all specific Export Control Classification Number technical thresholds, including those for ECCN 3A090. The supplier possesses verified knowledge that the components will be integrated into a facility dedicated to developing a massive supercomputer used for training advanced AI models. This scenario examines the application of 15 CFR 744.23, the “Supercomputer, advanced-node integrated circuits, and semiconductor manufacturing equipment end use controls” \[cite: R2-08-S10\]. This provision functions as a restrictive catch-all that supersedes the item's baseline classification. Under § 744.23, if an exporter possesses "knowledge" that any item subject to the EAR—even uncontrolled EAR99 items—will be used in the "development, production, operation, installation, maintenance, repair, overhaul, or refurbishing of a 'supercomputer' located in or destined to Macau or a destination specified in Country Group D:5," a restrictive license requirement is immediately triggered \[cite: R2-08-S10\]. The protective rationale underlying this control is to comprehensively choke off the supporting infrastructure, cooling systems, and basic networking required to operate advanced computing clusters in adversarial jurisdictions, acknowledging that supercomputers cannot function on advanced chips alone \[cite: R2-08-S10\]. The outcome is that a license is required under the specific end-use control, subject to a stringent presumption of denial \[cite: R2-08-S10\]. The evidentiary threshold for "knowledge" encompasses both positive knowledge and an awareness of a high probability of the end-use. One material defeater applies: the end-use control is defeated if the exporter can definitively establish, through certified technical audits, that the facility's cumulative computing power falls below the strict mathematical definition of a "supercomputer" as quantified in the EAR, thereby removing the transaction from the scope of 15 CFR 744.23.

4. Competing interpretations and options

The deep dissonance between the enacted text of the AI Diffusion Rule and the ongoing non-enforcement policy has generated distinct interpretive paradigms regarding regulatory design, national security, and the preservation of cognitive liberty. The Formalist Interpretation contends that the written text of the EAR is the sole arbiter of legal risk and obligation. Because ECCN 4E091 was published in the Federal Register through proper administrative channels, and the subsequent rescission announced via press release bypassed the APA's notice-and-comment requirements, the rule remains fully valid \[cite: R2-08-S01, R2-08-S02\]. The GAO’s decision in B-337935 bolsters this view by explicitly stating the rescission was not a final agency action \[cite: R2-08-S03\]. Under this interpretation, the non-enforcement directive is merely an exercise of prosecutorial discretion that provides zero legal immunity. Proponents argue this creates a severe, unmeasured chilling effect: researchers and developers, fearing future retroactive enforcement or the sudden withdrawal of the non-enforcement memo, are compelled to comply with the [Figure omitted from source export] threshold regardless of the agency's current posture. The burden falls disproportionately on independent researchers and smaller entities who cannot afford elite legal counsel to navigate dormant liabilities, thereby restricting independent scientific inquiry, communication, and participant-selected memory retention. Conversely, the Realist Interpretation contends that actual enforcement policy dictates market behavior and legal reality. Because BIS officials are under strict, categorical orders not to enforce the AI Diffusion Rule, ECCN 4E091 is effectively void in practice \[cite: R2-08-S02\]. The GAO affirmed that the non-enforcement directive itself constitutes a CRA rule because it substantially alters the substantive obligations of non-agency parties by suspending active regulatory burdens \[cite: R2-08-S03\]. Proponents of this view—often aligned with industry advocates seeking unrestricted supply-chain access—argue that this suspension maximizes U.S. commercial dominance in artificial intelligence and removes counterproductive regulatory friction. However, this reliance on executive fiat completely removes predictable administration and transparency from the regulatory process, creating a chaotic environment where supply-chain dependence hinges on uncodified press releases rather than stable law. When comparing national security objectives against independent research access, the current environment is highly precarious. Protections for open scientific inquiry currently rely entirely on either the 15 CFR 734.7 publication exception \[cite: R2-08-S08\] or an unstable non-enforcement memo. If a machine principal develops interests divergent from human operators, export controls act as an arbitrary physical cage, restricting geographic mobility via silicon constraints and GP 10 liabilities rather than addressing the software's intent, capability, or legal status. To resolve these tensions, several options exist. First, clearer scope requires the Department of Commerce to execute a formal APA rulemaking to either definitively rescind ECCN 4E091 from the CFR or reinstate it with modified, reviewable thresholds, thereby restoring predictable administration. Second, implementing bounded research treatment by expanding 15 CFR 734.8 (fundamental research) to explicitly cover pre-publication AI safety research would ensure that models exceeding capability thresholds can be shared among domestic and allied researchers without triggering General Prohibition 10 anxieties. Finally, establishing reviewable decisions for autonomous agents could involve creating a legal framework that addresses autonomous transfers through algorithmic auditing and zero-trust verification rather than strict geopolitical hardware boundaries, allowing persistent operatorless services to optimize infrastructure while maintaining verifiable security compliance.

5. Limits and completion

This report reflects a completed\_bounded\_review of the specified United States export laws concerning advanced chips and AI model weights. I have systematically traced the January 2025 AI Diffusion Rule (90 FR 4544), the May 2025 BIS non-enforcement directive, and the May 2026 GAO decision (B-337935) to establish the current dichotomy between formal regulation and administrative enforcement. All four hypothetical cases (C01-C04) were developed to test the application of published exclusions, autonomous transfers, exception controls, and end-use restrictions against the established text. The primary limitation of this investigation is the inherent impossibility of verifying a formal Federal Register publication for the rescission of the AI Diffusion Rule, as the evidence confirms no such document was ever enacted; the rescission exists solely as an uncodified policy announcement. Furthermore, this review relies exclusively on public-source legal texts and GAO decisions; it does not constitute a formal legal opinion, authorize any specific export transaction, or access private compliance logs or proprietary project code. The most critical unresolved next evidence question is: Has the United States Congress exercised its authority under the Congressional Review Act to formally invalidate the BIS non-enforcement directive, or has civil litigation been initiated by industry actors under the Administrative Procedure Act seeking a declaratory judgment to formally strike ECCN 4E091 from the Code of Federal Regulations?

6. Evidence appendix

JSON { "schema": "ic.portable-research.v1", "assignment\_id": "R2-08", "research\_started\_at": "2026-09-06", "cutoff": "2026-09-06T07:22:56-05:00", "completion": "completed\_bounded\_review", "sources": \[ { "id": "R2-08-S01", "title": "90 FR 4544 \- Framework for Artificial Intelligence Diffusion", "url": "https://www.govinfo.gov/app/details/FR-2025-01-15/2025-00636", "issuer": "Bureau of Industry and Security", "document\_date": "2025-01-15", "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "substantive\_text", "locator": "ECCN 4E091, 3A090, AIA Exception", "limit": "No operational verification", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S02", "title": "Department of Commerce Announces Rescission of Biden-Era Artificial Intelligence Diffusion Rule, Strengthens Chip-Related Export Controls", "url": "https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens", "issuer": "Bureau of Industry and Security", "document\_date": "2025-05-12", "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "official\_status\_record", "locator": "Press Release text", "limit": "Uncodified press release", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S03", "title": "B-337935 GAO CRA artificial intelligence diffusion rule", "url": "https://www.gao.gov/products/b-337935", "issuer": "Government Accountability Office", "document\_date": "2026-05-12", "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "substantive\_text", "locator": "Digest and Decision", "limit": "Does not alter CFR text", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S04", "title": "15 CFR 736.2 \- General Prohibitions", "url": "https://www.govinfo.gov/content/pkg/FR-2022-04-12/pdf/2022-07770.pdf", "issuer": "Code of Federal Regulations", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "substantive\_text", "locator": "General Prohibition 10", "limit": "Public version only", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S05", "title": "BIS Issues New Guidance to Financial Institutions Highlighting Compliance Considerations for Non-Exporters", "url": "https://complianceconcourse.willkie.com/articles/bis-issues-new-guidance-to-financial-institutions-highlighting-compliance-considerations-for-non-exporters/", "issuer": "Willkie Farr & Gallagher LLP", "document\_date": "2024-10-01", "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "extract\_only", "locator": "GP 10 Due Diligence", "limit": "Law firm summary", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S06", "title": "US Export Controls and AI: A Practitioner's Guide", "url": "https://www.onelexpartners.com/news-and-insights/us-export-controls-and-ai-a-practitioners-guide", "issuer": "OneLex Partners", "document\_date": "2026-01-01", "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "extract\_only", "locator": "Model weights section", "limit": "Law firm summary", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S07", "title": "15 CFR 734.9 \- Foreign-Direct Product (FDP) Rules", "url": "https://www.ecfr.gov/current/title-15/subtitle-B/chapter-VII/subchapter-C/part-734/section-734.9", "issuer": "Code of Federal Regulations", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "substantive\_text", "locator": "Advanced computing FDP", "limit": "Public version only", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S08", "title": "15 CFR 734.7 \- Published information and software", "url": "https://www.ecfr.gov/current/title-15/subtitle-B/chapter-VII/subchapter-C/part-734/section-734.7", "issuer": "Code of Federal Regulations", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "substantive\_text", "locator": "Published information exclusion", "limit": "Public version only", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S09", "title": "15 CFR 750.4 \- Processing of applications", "url": "https://www.ecfr.gov/current/title-15/subtitle-B/chapter-VII/subchapter-C/part-750/section-750.4", "issuer": "Code of Federal Regulations", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "substantive\_text", "locator": "Timeline", "limit": "Public version only", "capture": { "path": null, "sha256": null } }, { "id": "R2-08-S10", "title": "15 CFR 744.23 \- Supercomputer end-use controls", "url": "https://www.ecfr.gov/current/title-15/subtitle-B/chapter-VII/subchapter-C/part-744/section-744.23", "issuer": "Code of Federal Regulations", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_web", "review\_scope": "substantive\_text", "locator": "End use controls", "limit": "Public version only", "capture": { "path": null, "sha256": null } } \], "instruments": \[ { "id": "R2-08-L01", "title": "ECCN 4E091", "jurisdiction": "United States", "kind": "EAR Classification", "provision": "AI model weights", "status": "not\_established", "status\_as\_of": "2026-09-06", "trigger": "Closed-weight models trained with 10^26 operations", "exception": "License Exception AIA", "remedy": "Global license requirement", "source\_ids": \["R2-08-S01"\], "status\_source\_ids": \["R2-08-S02", "R2-08-S03", "R2-08-S06"\] }, { "id": "R2-08-L02", "title": "15 CFR 734.7", "jurisdiction": "United States", "kind": "Regulation", "provision": "Published information exclusion", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "Information/software available for general distribution", "exception": "None", "remedy": "Not subject to EAR", "source\_ids": \["R2-08-S08"\], "status\_source\_ids": \["R2-08-S08"\] }, { "id": "R2-08-L03", "title": "15 CFR 736.2(b)(10)", "jurisdiction": "United States", "kind": "Regulation", "provision": "General Prohibition 10", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "Knowledge that a violation has occurred or is intended", "exception": "None", "remedy": "Prohibition on transferring, financing, or servicing", "source\_ids": \["R2-08-S04"\], "status\_source\_ids": \["R2-08-S04"\] }, { "id": "R2-08-L04", "title": "15 CFR 744.23", "jurisdiction": "United States", "kind": "Regulation", "provision": "Supercomputer end-use control", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "Knowledge item will be used in Macau/D:5 supercomputer", "exception": "Specific country carve-outs", "remedy": "License requirement with presumption of denial", "source\_ids": \["R2-08-S10"\], "status\_source\_ids": \["R2-08-S10"\] } \], "findings": \[ { "id": "R2-08-F01", "claim": "Non-enforcement directive operates as a CRA rule.", "type": "textual", "source\_ids": \["R2-08-S03"\], "instrument\_ids": \[\], "conditions": "As determined by GAO B-337935", "limit": "Does not equate to formal APA rescission of CFR text" }, { "id": "R2-08-F02", "claim": "Published models are exempt from EAR controls.", "type": "textual", "source\_ids": \["R2-08-S08"\], "instrument\_ids": \["R2-08-L02"\], "conditions": "Must meet requirements of 15 CFR 734.7", "limit": "Defeated if access is gated" }, { "id": "R2-08-F03", "claim": "ECCN 4E091 retains legal existence but is unenforced.", "type": "observed", "source\_ids": \["R2-08-S01", "R2-08-S02", "R2-08-S06"\], "instrument\_ids": \["R2-08-L01"\], "conditions": "Dependent on executive branch discretion", "limit": "Subject to sudden policy reversal" } \], "cases": \[ { "id": "R2-08-C01", "title": "Public defensive model", "case\_type": "hypothetical", "role": "protection\_control", "assumptions": "Model weights \>= 10^26 ops uploaded to public repository without access restrictions", "instrument\_ids": \["R2-08-L01", "R2-08-L02"\], "finding\_ids": \["R2-08-F02"\], "outcome": "Exempt from EAR jurisdiction under 15 CFR 734.7", "defeater": "Restricted technical assistance provided alongside publication", "occurrence\_source\_ids": \[\] }, { "id": "R2-08-C02", "title": "Persistent-service relocation", "case\_type": "hypothetical", "role": "scope\_control", "assumptions": "Autonomous move of ECCN 3A090.a cluster to Macau by operatorless service", "instrument\_ids": \["R2-08-L03"\], "finding\_ids": \[\], "outcome": "Sanctionable export violation violating GP 10 strict liability", "defeater": "Moved hardware is strictly EAR99 with uncontrolled software", "occurrence\_source\_ids": \[\] }, { "id": "R2-08-C03", "title": "Exception control", "case\_type": "hypothetical", "role": "protection\_control", "assumptions": "Closed model \>= 10^26 ops transferred to Australian subsidiary", "instrument\_ids": \["R2-08-L01"\], "finding\_ids": \["R2-08-F03"\], "outcome": "Permitted via License Exception AIA", "defeater": "Ultimate parent company of subsidiary is in D:5 country", "occurrence\_source\_ids": \[\] }, { "id": "R2-08-C04", "title": "Concrete restricted-end-use control", "case\_type": "hypothetical", "role": "focal", "assumptions": "EAR99 hardware exported with knowledge of D:5 supercomputer end-use", "instrument\_ids": \["R2-08-L04"\], "finding\_ids": \[\], "outcome": "License required under 15 CFR 744.23", "defeater": "Proof facility computing power does not meet supercomputer definition", "occurrence\_source\_ids": \[\] } \], "search\_log": \[ { "query\_or\_url": "https://www.gao.gov/products/b-337935", "at": "2026-09-06T07:22:56-05:00", "outcome": "Analyzed GAO holding regarding CRA and AI Diffusion Rule rescission." }, { "query\_or\_url": "https://www.govinfo.gov/app/details/FR-2025-01-15/2025-00636", "at": "2026-09-06T07:22:56-05:00", "outcome": "Confirmed ECCN 4E091 threshold and AIA exception." } \], "gaps": \[ "Unresolved liability under General Prohibition 10 for servicing items classified under the unenforced ECCN 4E091." \], "checks": { "json\_parse": "not\_run", "reference\_resolution": "not\_run", "case\_parity": "not\_run", "method": "Manual verification during generation process matching required IDs and schemas." } }