How Trade Classification Works

The Trade Classification skill classifies products into 10-digit HTS tariff codes with 90.5% validated accuracy across 100 benchmark cases. Here's how it works and what makes it different.

The approach

The skill connects to your LLM workspace (Claude Desktop, or any MCP-compatible environment) and provides tools for step-by-step tariff classification. You describe a product in plain English. The server returns candidate headings with CBP ruling precedents, chapter notes, and classification instructions. Your LLM does the reasoning.

Proofwork is never in the inference path. The server handles deterministic work — candidate retrieval, evidence card construction, CROSS ruling lookup — while the host LLM applies GRI rules and selects the code. Your product descriptions never leave your environment.

What powers the accuracy

• 20,000 CBP CROSS rulings — real customs decisions used as precedent evidence at every classification step.
• Evidence cards — each candidate heading comes with ruling precedents, chapter note excerpts, and structural tags so the LLM can reason with real regulatory context.
• Multi-pass pipeline — heading identification, subheading selection, and leaf code selection each get their own evidence and instructions, matching how customs brokers actually work.
• Tariff memo decomposition — products are decomposed into structured fields (article type, materials, essential character, intended use) for better evidence retrieval.

Validation

Validated on 100 completely unseen cases (10 random seeds, 10 cases each). Mean accuracy: 90.5%, with three seeds achieving perfect 10/10. The range was 80–100% per seed. No overfitting: the validation mean exceeded the development mean.

Available tools

Your LLM calls these tools in sequence. Each step narrows the classification and provides evidence for the next decision.