How Trade Core Works

You describe a product to Claude or ChatGPT in your own words. This skill's tools hand your assistant the official tariff text and the CBP rulings that decided products like yours — each with its ruling number, its date and the facts you would have to confirm. Your assistant picks the 10-digit code and shows its working. Here's how that works, where it is strong, and where you should verify.

The approach

The skill connects to your LLM workspace (Claude, ChatGPT, or any MCP-compatible environment) and provides tools for step-by-step tariff classification. There is no Proofwork application to log into and no form to fill in: you describe the product in the chat you already use. The server returns candidate headings with CBP ruling precedents, chapter notes, and classification instructions. Your assistant does the reasoning.

Proofwork is never in the inference path. The server handles deterministic work — candidate retrieval, evidence card construction, CROSS ruling lookup — while your assistant applies the GRI rules and selects the code. We never call a model, and we never see your conversation: the tools receive the arguments of each call and nothing else.

What comes back

Your assistant decides whether that ruling governs your goods. When none matches, the tools say so rather than dressing a coincidence up as a precedent.

What the tools draw on

• CBP CROSS ruling rows — real customs decisions used as precedent evidence at every step. Each row carries how far it was checked, and whether the code it cites is still a line of the current schedule. What ships, in detail.
• Evidence cards — each candidate heading arrives with ruling precedents and chapter-note excerpts, so your assistant reasons from real regulatory text rather than from memory.
• One narrowing step at a time — heading, then subheading, then the 10-digit code, each its own tool call with its own evidence, the way a customs broker works. You can see what your assistant was given at every point.

Where it is strong, and where to verify

When a CBP ruling for a closely matching product is in the reference data, the tools state that ruling's holding and the classification typically follows it. When nothing comparable is there, the tools say so plainly rather than dressing a loose match up as a precedent, and the answer is reasoned from the tariff text — which is lower-confidence. Either way the tools state what a ruling said, not whether its facts are true of your goods: check the distinguishing facts yourself, confirm against the current HTSUS, and seek a binding ruling from CBP for material shipments.

Benchmark status

Measured on 21 September 2026: Claude Opus 5 driving these MCP tools, web search off, counted right only when all ten digits match, over the 143 scorable products of our 150-product test set. The two figures are the same run twice: once with the test products' own CBP rulings available to the tools, once with every one of them withheld.

An earlier accuracy figure on this site was withdrawn: the evaluation behind it let the tools retrieve the test products' own CBP rulings, the case count was overstated, and the customer-facing tool path had never been measured. The methodology page says exactly what was wrong, what changed so it cannot recur, which cases are excluded and why, the known failure modes, and the commands that reproduce both figures.

Available tools

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