U.S. Customs Rulings Online Search (CROSS)

• CROSS is a centralized database of CBP rulings that interprets tariff laws through structured records and HTS codes.
• It organizes thousands of legal decisions to support compliance, dispute resolution, and advanced machine learning research for tariff classification.
• Recent applications of the CROSS dataset include fine-tuning LLMs, yielding notable performance improvements and cost savings over proprietary models.

The U.S. Customs Rulings Online Search System (CROSS) is the official, centralized repository managed by U.S. Customs and Border Protection (CBP) for accessing customs rulings related to imports into the United States. CROSS provides searchable text of administrative rulings that interpret and apply the Harmonized Tariff Schedule (HTS) as well as other customs statutes and regulations. It is a cornerstone resource in global trade compliance, serving customs officers, legal professionals, and the importing/exporting community by supplying precedential guidance for tariff classification, valuation, country of origin determinations, and admissibility questions.

1. System Overview and Functionality

CROSS aggregates thousands of legal and regulatory decisions produced by CBP, with each record typically corresponding to a unique customs ruling. These records often include the ruling number, date, subject, description, applicable 10-digit HTS code, textual reasoning (resembling a legal decision’s “chain-of-thought”), and citations to statutory authority. The HTS code structure is hierarchical: the first six digits are internationally harmonized, while the subsequent four digits are U.S.-specific. CROSS assists not only in regulatory compliance but also as an evidentiary source for dispute resolution and policy research (Yuvraj et al., 22 Sep 2025).

Researchers can access the CROSS database through its web-based interface or by programmatic scraping, and they can retrieve specific precedent by HS/HTS code, ruling number, or free text queries. After the introduction of comprehensive dataset formatting and prompt engineering, the CROSS corpus has also supported the creation, evaluation, and fine-tuning of advanced machine learning models for classification and reasoning tasks (Yuvraj et al., 22 Sep 2025).

2. Data Structure and Preparation for Machine Learning

The CROSS repository contains highly unstructured, verbose text. To enable downstream automation and research, records are systematically parsed and restructured. In recent benchmarking efforts, a browser automation agent was deployed to scrape and preprocess thousands of HTML documents, yielding a dataset of 18,731 rulings mapped to 2,992 unique HTS codes. Each data sample is converted to a prompt–response format:

• The prompt consists of structured metadata (ruling number, title, summary) and a regulatory instruction.
• The response requires the model to generate (a) a concise product description, (b) a reasoning trace justifying the decision, and (c) the final 10-digit HTS code (Yuvraj et al., 22 Sep 2025).

This structure reflects human customs officer workflows and provides rich semantic supervision for training LLMs and other classifiers.

3. Machine Learning Applications and Benchmarks

CROSS underpins state-of-the-art research into automated tariff classification. The Atlas project introduced the first benchmark for HTS code classification, derived directly from the CROSS dataset (Yuvraj et al., 22 Sep 2025). Key technical details include:

• Model: Fine-tuning of the LLaMA-3.3-70B model using supervised token-level sequence modeling.
• Objective: For each sample $(x, y)$, minimize

$$\mathcal{L}_\mathrm{SFT}(\theta) = -\sum_{t=1}^m \log p_\theta(y_t | x, y_{<t})$$

Here, $x$ encodes the prompt and $y$ the desired output (description, reasoning, code).

• Training Regimen: 5-epoch, 1,400-step SFT on 16 × A100-80GB GPUs with AdamW and cosine learning rate schedule.

Performance metrics include:

• Fully correct 10-digit HTS classification (critical for U.S. entry)
• Partially correct 6-digit classification (international standard)
• Average correct digits per ruling

Atlas (LLaMA-3.3-70B, fine-tuned) achieved 40% fully correct 10-digit accuracy, 57.5% 6-digit accuracy, and 6.3 average correct digits—exceeding performance of proprietary GPT-5-Thinking and Gemini-2.5-Pro-Thinking by 15–27.5 percentage points on both full and partial accuracy (Yuvraj et al., 22 Sep 2025).

4. Comparative Insights, Strengths, and Limitations

CROSS operates as a centralized, static repository focused on transparency, legal compliance, and case precedent. In contrast to decentralized, privacy-preserving platforms like DEFenD (which use blockchain to increase data trust and auditability in freight declaration (Vos et al., 2018)), CROSS’s strengths are its regulatory legitimacy and structured linkage of rulings and statute.

Benchmarks derived from CROSS reveal persistent challenges in HTS classification:

• Even state-of-the-art LLMs attain only 40% full-code accuracy, underscoring the combinatorial and legal complexity of the problem (Yuvraj et al., 22 Sep 2025).
• A major bottleneck is the translation of richly detailed narrative rulings into precise HTS codings, a task that demands both retrieval and chain-of-thought reasoning.
• Cost and privacy: Atlas demonstrated up to 5× cost savings over GPT-5-Thinking and 8× over Gemini-2.5-Pro-Thinking, with full on-premise deployability—critical for compliance and sensitive industries (Yuvraj et al., 22 Sep 2025).

5. Integration with Decision Support and Advanced Analytics

CROSS rulings serve as critical data for downstream decision support in various research domains:

• Systems for automated text-based commodity classification (combining retrieval, deep learning, and knowledge graph methods) use CROSS as a source of labeled guidance (Lee et al., 2021, Shubham et al., 2022, Lee et al., 2023).
• Cross-linking CROSS with scenario analysis, cost-benefit/multi-criteria analyses, and simulation (as detailed for cargo screening (Siebers et al., 2013)) could enable trade-off analyses between compliance, inspection resource allocation, and service standards.
• Research in active learning, adaptive fraud detection (e.g., ADAPT, GraphFC), and domain adaptation (DAS) often takes CROSS-styled legal rulings as input or reference for supervised or transfer learning (Kim et al., 2020, Mai et al., 2021, Singh et al., 2023, Park et al., 2022).

A plausible implication is that richer integration—linking real-time scenario modeling, predictive algorithms, and regulatory precedent—would transform CROSS from a lookup tool to a decision-support platform.

6. Research Directions and Future Challenges

While the Atlas benchmark demonstrates the feasibility of using LLMs to automate HTS classification from CROSS rulings, fundamental limitations remain:

• Retrieval/Reasoning: Integration with retrieval-augmented generation (over the 17,000-page HTS or ruling library) could reduce hallucinations and improve long-tail code accuracy (Yuvraj et al., 22 Sep 2025).
• Data Utilization: Leveraging both the structured and unstructured ruling content may enable more fine-grained and legally robust classification, especially when combining chain-of-thought and contrastive learning.
• Model Alignment: Direct Preference Optimization (DPO) and hybrid loss strategies are cited as promising mechanisms to increase sensitivity to small semantic/code differences crucial for compliance (Yuvraj et al., 22 Sep 2025).
• Scaling/Deployability: Research is encouraged into smaller models for edge deployment, privacy-preserving reasoning, and efficient adaptation to regulatory changes.
• Benchmarking: With only 40% accuracy at the 10-digit level, the task remains challenging and is now positioned as a new community benchmark for global trade and machine reasoning (Yuvraj et al., 22 Sep 2025).

7. Conclusion

CROSS is a foundational digital infrastructure for U.S. customs and global trade classification, functioning both as a legal reference and an emerging machine learning benchmark task. Its integration into LLM-centric compliance automation, fine-tuning strategies, and hybrid decision-support platforms holds promise for streamlining trade, increasing transparency, and reducing the frequency of misclassification-driven shipment delays. Nevertheless, the legal, combinatorial, and retrieval challenges crystallized in large-scale evaluations ensure ongoing demand for further research and technology development in this domain.