The HS6 classification benchmark

Every item that crosses a border carries a Harmonized System code. The first six digits are internationally standardised. The national tariff line that sets the duty rate, the trade agreement treatment and any permit requirement is built on top of those six. Get the six wrong and the rest of it is wrong too. That is a misdeclaration, and the exposure sits with the importer.

The nomenclature is legal text. It turns on distinctions that do not appear in the way a product is described on a commercial document: knitted against woven, whether a preparation is put up for retail sale, whether a machine's principal function is the one its name implies. The task is a translation between the language of trade and the language of the tariff schedule. Most of the difficulty sits in that gap, and the numbers below show it.

I went looking for a published accuracy figure in this space with a dataset, a scoring rule, a sample size and a date attached to it. I did not find one. If you are carrying the compliance risk, a bare percentage gives you nothing to work with.

So the set, the scoring rule, the confidence bands and the failures all get published, one page per run, at a URL that does not change. A run that goes backwards gets published with the regression visible. The series is the point, and it starts at one.