Methodology

How we know an explanation is right

The complaint we built this around is that existing prep gives you an answer letter and no reasoning. Ours gives reasoning — which is only worth more if you can check it. Here is exactly how each one is produced, and where the process still falls short.

697
past-exam questions
1,198
citations indexed
2,280
19 CFR sections
29,725
tariff lines

Where the material comes from

Past exams, official answer keys, Title 19 CFR, the HTSUS, and CBP directives are works of the U.S. government and carry no copyright (17 U.S.C. §105). That is the entire source base. We do not use, adapt, or paraphrase any competitor’s question bank, and we do not collect questions from candidates who have sat an unpublished exam — that material is covered by the exam-day agreement and is not ours to take.

The regulations come from the eCFR; the tariff comes from the USITC’s open API. Both are re-ingested rather than transcribed, so what you read here is what the source says.

How an explanation is made

  1. 01

    Source

    Nothing is written from memory. Every explanation is generated against the actual text of the cited provision, pulled from our ingested copy of Title 19 CFR and the HTSUS. If the provision is not in the corpus, no explanation is produced for that question.

  2. 02

    Ground

    The model sees the part of the section the answer key points at — centred on the cited subsection, not the first paragraph. This sounds like a detail. It was the single biggest quality problem we had: §111.30 runs 7,544 characters, and a head-slice handed the model paragraphs (a) and (b) while the answer lived in (d).

  3. 03

    Judge

    A second model, on a different tier and a different prompt, checks the explanation against the same source text. It never sees the first model's reasoning — only the question, the answer, the sources, and the draft. It rejects anything that cites a provision absent from the context or asserts a rule the text does not support.

  4. 04

    Audit

    A separate pass walks every stored citation and resolves it against the corpus. A reference into a part we have ingested that points at a section we do not hold fails the audit. This runs on the deploy path against the real database, not in CI, because on an empty database it would pass by proving nothing.

What the judge rejects

Roughly a third of first drafts are thrown away. The usual reasons are a citation to a provision that is real but does not say what the draft claims, reasoning that reaches the right letter for the wrong reason, and confident treatment of a question whose answer genuinely depends on an exhibit we do not hold. Rejected drafts are not published in a softened form; they are regenerated or left undone.

Where it still falls short

Two questions have no walkthrough at all. One asks for a dollar figure derivable only from a broker’s invoice that CBP supplies as an exhibit — we hold the question, not the exhibit, and the number cannot be derived from anything we have. The other is a classification question the model declines to answer. Both show the official answer and its citation, and no invented reasoning.

A few questions reference exhibits we don’t hold. They are answerable on the real exam and not from our copy. They are marked rather than quietly dropped.

Coverage is uneven by topic. CBP prints its exams under six broad categories, two of which are catch-alls holding over half the questions. We assign a topic from the printed range first, then refine using the answer key’s own citation. Subjects CBP never breaks out — bonds, marking, penalties, FTZ — therefore have thin coverage. That is a real limit of the source material, not an oversight, and more sittings is the only fix.

Editions, and why they matter

Each sitting is bound to the reference editions that were designated for it. A question from an April 2026 exam is explained against the regulations as they read for that sitting, not as they read today. When a new cycle’s editions land, every stored citation is re-resolved against them before anything is published — sections do move.

You can read the sources yourself: the 19 CFR navigator and the HTSUS library carry the same text our explanations are built on.

Found something wrong?

Tell us. Every explanation cites its source precisely so that a disagreement is checkable rather than a matter of opinion — that is the point of citing at all. We correct or retract, and we would rather hear it from you than leave it live.

Report a correction

Independent study product. Not affiliated with U.S. Customs and Border Protection or any government agency.