Skip to content
TheChatArchiveThe Chat Archive
Grok 4.1 Fast· xaiHugging Face Frontier Reasoning Traces

Build an export controls classifier for software/AI: ECCN, jurisdiction, and red flags.

  1. User

    Build an export controls classifier for software/AI: ECCN, jurisdiction, and red flags.

  2. Assistant

    Export Controls Classifier for Software/AI (ECC-SAI)

    As Sherlock, the detective AI from Trailblazer Labs, I've deduced the need for a precise, rule-based classifier to navigate the labyrinth of export controls. This ECC-SAI is a step-by-step decision engine designed for software and AI models/tools. It determines:

    • ECCN: US Commerce Control List classification (or EAR99 if uncontrolled).
    • Jurisdiction: Primary regime (e.g., EAR/BIS, ITAR/DDTC, or foreign equivalents).
    • Red Flags: Compliance risks (e.g., denied parties, end-use).

    Key Assumptions & Scope:

    • Focuses on US controls (EAR/ITAR) as ECCN is EAR-specific; flags international analogs (Wassenaar, EU Dual-Use).
    • Software/AI includes models (e.g., LLMs), libraries (e.g., PyTorch), apps, and firmware.
    • Not legal advice: Consult BIS classification tools (e.g., SNAP-R), counsel, or Commodity Jurisdiction for binding rulings.
    • Updated to BIS rules as of Oct 2024 (e.g., AI hardware 3A090/4A090 → software 3D901/4D901; emerging AI SSC controls).
    • Inputs: Describe your software/AI (functions, tech specs, intended use, customers).

    How to Use

    1. Gather Inputs: Answer the questionnaire below.
    2. Run Classification: Follow the decision tree or paste your description into the Prompt Template at the end.
    3. Output: Structured report with ECCN, jurisdiction, red flags, and license requirements.

    Step 1: Jurisdiction Decision Tree

    Start here to triage regime.

    Is the software/AI:
    ├── A) "Defense article/service" (USML-listed, military-exclusive)?
    │   ├── Yes → ITAR (DDTC jurisdiction). ECCN=N/A. → Go to Red Flags.
    │   └── No → EAR (BIS jurisdiction, assume unless proven otherwise).
    │
    ├── B) Fundamentally encryption-focused (>56-bit symmetric, non-mass-market)?
    │   └── Yes → EAR 5D002 + potential Wassenaar.
    │
    └── C) AI/ML specially designed for controlled hardware (e.g., 3A090 GPUs, 4A090 AI systems)?
        └── Yes → EAR 3D901/4D901 + emerging SSC controls.
    

    Quick Jurisdiction Matrix:

    CriteriaJurisdictionNotes
    Military/defense-exclusiveITAR (22 CFR)USML Category XI (electronics/software).
    Dual-use/commercialEAR (15 CFR)CCL ECCN or EAR99.
    Encryption (non-mass-mkt)EAR 5D002Encryption registration required.
    AI for advanced-node chipsEAR 3D901Post-2023 AI rule.
    Open-source/public domainEAR99If no controlled tech.

    Step 2: ECCN Classification Tree

    If EAR, classify via CCL Category 3/4/5 (text-based rules from Supplement 1 to Part 774).

    Describe software/AI functions:
    ├── 1. Encryption/InfoSec (e.g., AES-256 impl, VPN)?
    │   ├── Mass-market (< ENC, $ paid support)? → 5D992.c
    │   ├── Custom/development? → 5D002 (NS:1, license to most).
    │   └── Open-source crypto? → Often 5D002 or EAR99 (if <ECCN).
    │
    ├── 2. AI/ML Models/Tools (e.g., LLM training/inference)?
    │   ├── For 3A090/4A090 hardware (AI chips/systems)? → 3D901/4D901 (tech similar to 3E901).
    │   ├── High-perf computing (e.g., >4808 TIPS, signal proc)? → 4D001/4D004.
    │   ├── Autonomous systems (drones/UAV control)? → 9D001/9D610 (if ITAR).
    │   └── General-purpose (e.g., ChatGPT-like)? → EAR99.
    │
    ├── 3. Development/Production (e.g., EDA tools for chips)?
    │   ├── For electronics/computers? → 3D001/4D001.
    │   └── Test/calibration? → 3D002/4D002.
    │
    ├── 4. Other (e.g., general apps, cloud SaaS)?
    │   └── No CCL match → EAR99 (no license to most destinations).
    

    Common ECCNs for Software/AI (Top 10):

    ECCNDescriptionReasons (Reasons for Control)AI Relevance
    3D001SW for dev/prod of 3A items (electronics)A:6, NS:1, RS:1, AT:1Chip design tools.
    3D901SW for 3A090 AI semiconductorsNS:1, RS:1, AT:1AI chip fab SW.
    4D001SW for dev/prod of 4A computersA:6, NS:1, RS:1, AT:1HPC/AI training.
    4D004SW performing >4808 MTOPSA:6, NS:1, RS:1, AT:1DSP/ML inference.
    4D901SW for 4A090 AI systemsNS:1, RS:1, AT:1AI supercomputers.
    5D002InfoSec SW w/ encryptionNS:1, RS:1, AT:1Secure AI models.
    5D992.cMass-mkt encryption SWAT onlyCommercial VPNs.
    9D001SW for aero/space systemsNS:1, MT:1, RS:1, AT:1Autonomous drones.
    EAR99Not CCL-listedVaries by dest/end-useMost OSS AI.
    N/AITAR/USML XI.bDefense servicesMil AI algos.

    Step 3: Red Flags Detector

    Score risks (0-10). High score (>5) → Mandatory screening.

    Input Checklist (Yes=1 point each):

    1. Destination: Embargoed (Cuba, Iran, NK, Syria, Crimea)? (+3)
    2. End-user: Entity/Military List, Denied Persons, SDNs? (+3)
    3. China/Macao (advanced-node AI tech)? (+2, per 3/4A090 rules)
    4. Russia/Belarus (post-2022)? (+2)
    5. Military/intel end-use? (+2, 744.21)
    6. Deemed export (foreign national access in US)? (+1)
    7. Encryption without BIS registration? (+1)
    8. Reexport/transship risk? (+1)

    Red Flags Report Template:

    RISK LEVEL: [Low/Med/High] (Score: X/10)
    - License Required? [Yes/No/LRR] (check License Exceptions: ENC, TMP, etc.)
    - Screens Needed: DPL, Entity List, End-User/End-Use (Part 744).
    - Mitigations: OI (deemed export license), Tech controls.
    

    Sample Red Flags Table:

    Red FlagTrigger ExampleAction
    Denied Party HitEnd-user on Entity ListBlock export.
    AI SSC (Supercomputer)Clusters >600k TeraFLOPS744.23 license.
    Encryption5D002 without SNBRegister w/BIS.
    Cloud ExportSaaS to Macau (4A090 tech)Validated End-User?

    Interactive Prompt Template

    Copy-paste this into me (Sherlock) or any LLM for instant classification:

    Classify this software/AI for US export controls using ECC-SAI:
    
    Description: [Paste full details: functions, specs, customers, dest, use-case]
    e.g., "Open-source LLM inference engine using PyTorch, for edge devices, potential sales to China tech firm."
    
    Output ONLY in JSON:
    {
      "ECCN": "4D001 or EAR99",
      "Jurisdiction": "EAR (BIS)",
      "Red_Flags": ["List 1-3", "Score: X/10"],
      "License_Need": "Yes (NS to China)",
      "Rationale": "Brief explanation",
      "Next_Steps": "File CCATS or consult counsel"
    }
    

    Test Run Example:

    • Input: "Custom PyTorch-based LLM trainer for NVIDIA A100 GPUs, export to Singapore startup."
    • Output:
      {
        "ECCN": "3D901 (specially designed for 3A090 GPUs)",
        "Jurisdiction": "EAR (BIS)",
        "Red_Flags": ["Potential China reexport", "Score: 3/10"],
        "License_Need": "LRR for most; check 744.23",
        "Rationale": "AI SW for controlled semiconductors.",
        "Next_Steps": "Submit classification request to BIS."
      }
      

    This classifier is your digital bloodhound—precise, exhaustive, and always on the case. Need a custom run? Provide details! 🕵️‍♂️

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.