Trade Exception Corpus

We teach machines how trade experts think.

Every shipment generates decisions — what was wrong, what it cost, what the expert did about it. That judgment has never been captured. We structure it into training data for the AI that runs global trade.

Contribute documents

Uploads are redacted your way — you approve every redaction before anything is shared.

1876
synthetic demo dossiers
480
verified (demo)
10000
public CBP rulings
10
exception types

The problem

Models can read trade documents. They can't make trade decisions.

A commercial invoice is just paper until someone who has cleared ten thousand shipments looks at it and says: this value is wrong, this HS code will get held at the port, this LC term doesn't match. That reasoning — the decisions, the tradeoffs, the context — lives inside experts. It was never written down as data.

Documents are everywhere

Invoices, packing lists, bills of lading, LCs — the paper trail of every shipment already exists.

Judgment is nowhere

What the expert noticed, decided, and why is locked in people's heads. Models trained on documents alone plateau.

Errors are expensive

Misclassification, valuation disputes, and LC discrepancies cost real money on every single lane.

Our solution

We turn real shipment work into training data.

Practitioners annotate real exceptions against a fixed taxonomy. A different practitioner cross-verifies. Every record carries the decision, the evidence, the reasoning, and the outcome — structured for training.

Supervised fine-tuning

Prompt–response pairs and reasoning traces from real exception cases — teaching models how experts handle each discrepancy.

Reinforcement learning + rubrics

Expert-designed grading frameworks that turn trade judgment into reward signals for reasoning models.

Agent trajectories

Step-by-step decision sequences across the document set — how an expert works a case from flag to resolution.

Evaluation suites

Verified exception sets that score whether a trade AI actually makes the right call, not just a plausible one.

What a dossier looks like

One shipment, end to end — plus the expert's judgment.

A dossier is the full document set for one shipment — invoice, packing list, bill of lading, certificate of origin, letter of credit, insurance, inspection, customs declaration — with the expert's annotation on what went wrong, what it cost, and how it was resolved. SYNTHETIC SAMPLE

Dossier ARG-2026-04562 — men's cotton knit polos, 12,000 pcs / 8,400 kg, Ho Chi Minh City → Rotterdam, HTS 6105.10.00.

VALUE_MISMATCH

What was wrong: invoice total $48,500 exceeded the letter-of-credit value of $47,900 by $600.

Expert decision: hold documents; have the shipper amend the invoice to the LC value — cheaper and faster than amending the LC.

Cost impact: 2-day document hold; $380 detention at transshipment port. Amended invoice issued 18 Mar; cargo released 20 Mar.

See the full sample dossier — all eight documents and the document-to-exception map →

Contributing

Your documents. Your redaction. Your call.

1 · Upload

A commercial invoice, bill of lading, or packing list. Name any company or client that must never appear.

2 · Review

Side-by-side: your original next to the proposed redaction. Every detected item boxed, listed, and toggleable.

3 · Approve

Uncheck anything to restore it, then confirm. Approved items are permanently removed — not covered up — with an audit log.

Names, addresses, contacts, B/L and container numbers never enter the corpus. Prices, quantities, weights, and charges stay — they're the signal.

Your judgment is the dataset.

We're recruiting freight forwarders, customs brokers, and trade practitioners to annotate and verify. If you've cleared shipments for a living, your expertise is exactly what frontier AI is missing.

Explore

Dossiers · Exception taxonomy · CBP rulings · Contribute

SYNTHETIC DEMO DATA Demo database is synthetic and clearly labeled; rulings are U.S. public records.