An open ontology for UK trade remedy measures, tested against the TRA case register, the UK Trade Tariff and legislation.gov.uk

Every anti-dumping and countervailing duty the UK charges runs a chain. The Trade Remedies Authority opens a case, the case names commodity codes, a legal instrument gives the resulting measure force, and the measure assigns a duty rate to a named exporter. By 31 March 2026 the measures in force covered around 20.6 billion pounds of UK imports. We built an open ontology of that chain, censused every link in it, and found that it breaks in twenty-five places. All twenty-five are repaired in the repository.

What we measured

We visited all 21,008 commodity codes in the UK Trade Tariff. 405 of them carry a trade remedy measure, which comes to 9,385 distinct measures citing 49 legal instruments between them, with 990 additional codes assigning duty rates to named exporters. We harvested all 92 cases on the TRA public file, dereferenced every legal instrument locator against legislation.gov.uk and gov.uk, and resolved every published exporter name against the Global LEI System. The census is complete rather than sampled, and every headline figure is computed twice, once set-based in Python and once by SPARQL over the built graph, with a gate that exits non-zero if the two ever disagree, currently across 20 metrics.

The tariff truncates the legal instrument link at 200 characters

Thirteen of the 49 cited instruments publish a locator of exactly 200 characters, cut mid word. All thirteen return HTTP 404. Every locator shorter than 200 characters returns HTTP 200, and no locator in the set exceeds 200 characters. That is a field length limit rather than a content error, and it falls almost entirely on trade remedies notices, whose gov.uk paths are the longest in the set.

The effect is that 1,477 of the 9,385 live measures, 15.7 percent, publish a broken link to the instrument that gives them legal force. We repaired all thirteen by fetching the gov.uk collection page named in the surviving prefix and keeping the one document link that begins with the truncated string. Every repair extended to exactly one candidate, and every candidate returns HTTP 200. The before and after table ships in the repository as a CSV, and we have staged the defect report for the Trade Tariff team, whose platform is itself open source on GitHub.

The case register drops leading zeros from commodity codes

A UK commodity code is 6, 8 or 10 digits. Two TRA cases each publish six nine-digit codes, and all twelve return HTTP 404 against the tariff. Restore the leading zero and all twelve resolve, each to a trout product, which matches what both cases are about. That is a numeric cast stripping the zero from chapter 3 goods. All twelve are repaired in the repository, and a repair is accepted only where the published value fails to resolve and the restored value resolves.

The case file joins to the measures

The public file and the tariff required joining before either could be relied upon as a register, as they describe the same measures from different perspectives. Of the 92 cases on the TRA public file, 91 publish commodity codes and 79 join to at least one commodity carrying a live measure. These cases publish 2,204 codes in total, of which 1,393 carry measures in the tariff today. This remainder is legitimate, because the cases also cover terminated investigations and expired measures, and the join table is included in the repository so the residue is inspectable rather than asserted.

The exporter name field is not a name field

Duty rates are assigned to individually named exporters through additional codes, and the name is the only identity the tariff publishes. Of 982 named exporters, 955 do not resolve to exactly one entity under a legal name lookup. Six in ten values mix the company name with address fragments, 65 carry raw HTML markup or double encoded characters inside the published name, and 55 pack more than one legal entity into a single value. Citation format is equally ungoverned: 87 legal citation records match no consistent scheme, and the most cited instrument of all, at 6,738 citations, is recorded with a double space in its own name.

The same questions, asked of other jurisdictions

The UK trade regime is largely inherited, yet the inherited framework is increasingly diverging from its origins. The WTO’s publicly accessible notification tables, which require no subscription, list the United Kingdom as having notified 7 anti-dumping measures and 3 countervailing measures since 1995, in contrast to 46 measures currently in force. The UK tariff references EU parent regulations using R-codes, which resolve against the Publications Office CELLAR service, a keyless system that maintains an in-force status flag. Both EU parents cited as the legal basis for active UK measures, R1722/18 and R0999/14, are recorded by the EU itself as no longer in force, a status that affects 6 UK measures. The Canadian register serves as a contrasting example, where the CBSA measures-in-force file parses cleanly into 189 measure rows and 4,681 commodity codes with zero malformed tokens, demonstrating that the defect classes identified above are not inevitable. The Australian register refused a scripted connection outright, a fact we record rather than attempt to circumvent.

The document layer underneath the measures

Underneath the measures sits the evidence base. The 92 cases on the public file list 3,849 submissions, behind which we discovered 12,964 public documents. A register census of that document layer is the next build, addressing whether every listed document resolves, whether listed and actual counts agree, and how much of the corpus is machine-readable text rather than scanned image. The same link-integrity question that found the 200-character truncation applies one layer down.

How the ontology holds disagreement

Identity and legal citation are modelled as dated assertions by a named source, never as properties of a thing. An IdentifierAssertion records who published which identifier for whom and how many entities it resolves to. A LegalCitationAssertion records the citation code and the locator separately, because they can disagree, and a ResolutionObservation records what the locator returned on a given date. Each identifier scheme declares its own conformance rules as data, so adding a jurisdiction means adding a concept rather than editing code. SHACL carries one shape per defect class, which makes the validation report the findings table.

Everything is validated three ways: with pyshacl, with our own open-source Open Ontologies engine, and by the dual-computation gate. Validating this vertical also exposed a real gap in our own engine, which did not implement SHACL's numeric range constraints. We fixed the engine, pinned regression tests against pyshacl, and both engines now return identical violation counts. Findings that died under checking are recorded in the build report alongside the ones that survived, including one plausible-looking claim about a revoked sanctions regulation that turned out not to be a trade remedies defect at all.

Working with us

We do bounded first engagements. For an investigating authority, a customs administration or a firm that depends on remedy measures being right, that is a fixed-scope integrity audit of your register against the sources it must agree with, delivered as a reproducible pipeline you keep, with every headline figure computed two ways. The repository shows exactly what that looks like.

Kampakis and Co Ltd, trading as The Tesseract Academy. fabio@thetesseractacademy.com