For tax reasons, the X-Men aren't human
A US federal court once had to decide whether the X-Men are human. For tax reasons.
Marvel's Toy Biz argued its action figures were non-human creatures, so they carried 6.8% duty as toys rather than the 12% charged on dolls representing humans. The judge examined 60+ figures and agreed. Same plastic. Half the duty. Decided on comic-book lore.
That reads as absurd until you look at what's happened over the past year.
The gap used to be a few percent. Now it's 10x.
After the Supreme Court threw out most of 2025's tariffs in February, 330,000+ importers are re-filing their codes. The gap between right and wrong is no longer a few percent. It's 2.5% versus 25%+.
Ford brought the Transit Connect in as a passenger van at 2.5%, stripped the rear seats to sell it as a cargo van, and owed the 25% chicken tax. Customs caught it. Ford fought, lost, and settled for $365 million.
The tariff schedule lists the codes. It does not tell you which side of the line you're on.
Customs rulings do. That's where those lines were actually argued and drawn.
The schedule lists codes. Rulings draw the lines.
So that's what we built on: retrieval over past CBP cases, re-ranked against the General Rules of Interpretation. Give it an ambiguous description and it flags the risk before you ship. Hit a blocked shipment and it hands you the cases to argue back with.
Then we tried to break it
ATLAS, a fine-tuned 70B model, scores 40% exact 10-digit accuracy on its own benchmark. We ran a training-free, open-book pipeline on the same test split: 55% with a flash model, 65.5% with a pro preview. No fine-tuning. No retraining on every tariff revision. And it cites the authority behind every answer, which a fine-tuned model cannot do.
The full breakdown
Every number and caveat, including the full benchmark methodology, is in the HTS classification report.
What's the strangest classification call you've had to defend? If you're wrestling with tariff classification at scale, .