How we check accuracy
We do not publish an accuracy percentage yet. This page describes how the data is built and checked, and shows the results of the code checks of the current release.
How each rule is built
- Two independent AI passes read the primary sources (HTS Chapter 99 text and notes, Federal Register documents, CBP CSMS messages, USTR annexes) and each produce the rules on their own.
- Adjudication against the sources settles every disagreement, with the reason recorded on the rule.
- Code copies every number (rates, dates, HTS codes) from the source files. Models choose and structure; they never type a value.
- Code checks catch whole classes of errors before release (below).
- A watch for changes: a job checks for new HTS releases, Federal Register documents and CBP messages before each update, and every API response carries the dataset version so you can pin and compare.
What the code checks look for
- Attributes the code can settle are derived, not asked: a statistical suffix that says “new passenger car”, or a heading for goods made only of one metal, decides the branch.
- Lists are cross-checked across sources: HTS PDFs, USTR annexes, CBP messages and the USITC export.
- Dates and quotes are matched against the Change Records and the cited text; a citation whose quote is not found in the stored source is reported.
- Stacking invariants per measure, such as headings that must never apply together, run on every result.
- Conflicting primary sources are flagged for review by a broker or counsel (54 items in the review queue of this release), never settled silently. Affected results carry a review flag.
- Two engines, one answer: the API's engine reproduces the Python reference engine exactly on the reference and stress queries before a release ships.
Code checks of release 1.0.0
| Check | Items checked | Findings | Errors |
|---|---|---|---|
| Attributes settled by the HTS code itself | 4,185 | 0 | 0 |
| Stacking invariants on computed results | 7,798 | 6 | 6 |
| Stacking relations between rules | 54,553 | 741 | 0 |
| Scope lists cross-checked across sources | 222,481 | 671 | 0 |
| Dates and quoted citations against source text | 11,603 | 943 | 0 |
| IEEPA refund calculations | 560 | 0 | 0 |
| Blind regression cases | 50 | 0 | 0 |
Findings are items for review (for example, a quote the checker could not locate in a stored source); errors are failures tied to known open items in the review queue.
Next: expert review
A sample of results checked by a licensed customs broker is planned. Until it is done, results are AI-built and code-checked, and we make no claim of a measured accuracy rate. If you find an error, write to support@htscalculator.com with the request and meta.dataset_version.