Content methodology
Where every answer comes from, and what stops a wrong one shipping
Questions are written from the text of 40 CFR Part 82, Subpart F, which is a work of the United States Government and not subject to copyright. Nothing here is copied from a certifying body, a trade association or a competitor’s bank.
Public-domain sources only
Under 17 U.S.C. § 105, a work prepared by an officer or employee of the United States Government as part of that person’s duties is not subject to copyright protection in the United States. The Code of Federal Regulations is such a work. That is what makes it legitimate to build a question bank from it and to quote it in full on a card.
Every regulatory item is drafted from a committed snapshot of the eCFR rather than from a live fetch. The snapshot in use is 2026-08-13, and the repository records the request URL and the SHA-256 of the response alongside it, so the corpus an item was written from can be reproduced byte for byte.
Where an answer is not in Subpart F, the item says so and carries a different class rather than pretending. That taxonomy is the honest core of the product.
Never used, as a matter of policy
- Any certifying organization’s question bank.
- Any competitor’s practice questions, paraphrased or otherwise.
- Copyrighted training manuals or standards documents.
- A model’s recollection of what Section 608 says.
Each item records the paragraph it was drafted from, the snapshot, the drafter and the sign-off. If anybody ever alleged that a question was copied from their bank, that record is a contemporaneous, per-item demonstration of where it actually came from.
The four source classes
Every item carries exactly one
An answer can be true for four different reasons, and treating them as one is how prep material ends up asserting that the regulation requires something it merely recommends. The class is printed on the card.
| Class | Name | When it applies | What the item must carry |
|---|---|---|---|
| REG | Regulation | The answer is stated in 40 CFR Part 82 Subpart F. | A paragraph citation that resolves on eCFR, and the paragraph quoted verbatim. |
| STAT | Statute | The answer comes from Clean Air Act Title VI or the Montreal Protocol, not from the subpart. | The statute or instrument cited by name and section, and the text quoted. |
| SCI | Science | The answer is a physical or chemical fact, not a legal requirement. | A named public-domain scientific source, cited and quoted. |
| PRAC | Practice | The answer is accepted trade practice that is not written in the regulation. | An explicit badge saying so. The build rejects a practice item whose wording claims the regulation requires it. |
Subpart F is not Part 84
The Section 608 exam is a Subpart F exam. Part 84 — the AIM Act rules on hydrofluorocarbons — is a different programme with different numbers, and mixing the two is the most plausible mechanism behind the contradictory questions technicians report finding in other prep material.
The build refuses an item that cites outside Subpart F unless it carries an explicit statute or practice class. It is not a guideline; the check fails the build.
A practice item may not pose as a rule
A PRAC item is accepted trade practice that is, by definition, not written in the regulation. The verifier reads the stem and the explanation of every practice item and rejects it if the wording claims the regulation, the rule or the EPA requires the thing. That check is mechanical, because the failure it prevents is a sentence somebody wrote in a hurry.
Five checks, in order
| Layer | What it does | Automated |
|---|---|---|
| L1 Provenance by construction | Each item is drafted from exactly one supplied paragraph of the regulation. The drafting step is never asked what it knows about Section 608; it is handed the text and asked to write a question about that text. An item that cannot name the paragraph it came from does not exist. | yes |
| L2 Mechanical assertions | Runs on every push, in seconds, with no network and no model. The citation must resolve to a paragraph the eCFR actually publishes. The quote must be an exact substring of that paragraph after whitespace is collapsed. There must be four distinct, non-empty options and an in-range answer. The citation must be inside Part 82 Subpart F unless the item carries an explicit statute or practice class. No two stems in a bank may be near-duplicates. Ids are unique and are never reused. | yes |
| L3 Blind adjudication | A second pass in a fresh context sees only the paragraph, the stem and the four options — never the recorded answer — and must choose one. Disagreement flags the item for a human. The verdict is recomputed from the recorded choice rather than trusted from a stored flag, so an answer edited after adjudication cannot carry a stale agreement with it. | yes |
| L4 Human sign-off | The one part of this that a machine may not do. A person reads the item against the paragraph and signs it. A release build refuses to include an item that has not been signed. | no |
| L5 Field signal | After launch: a report control on every question in the app, and an alert when an item’s miss rate is anomalous. A question that everybody gets wrong is usually a bad question, not a hard one. | no |
The gate is never switched off
Every shipped question carries a source class, a resolvable citation, a verbatim source quote, a snapshot date and a verification state. Continuous integration fails the build if any item lacks one of them. There is no flag to skip it for a release, and adding one would defeat the only claim this product actually makes.
Findings are written to a review queue ordered by severity. That queue is the human work list, and it is in version control alongside the items.
Where the bank stands today
The 25 items on the sample page read 25 draft. Layers one and two pass. Layer three has not been recorded against them and layer four has not happened, so none of them is eligible for a release build yet.
That is what the word on the card means, and it is why the word is on the card.
Why the exam is shaped the way it is
The number of questions, the split between the core group and the technical group, and the subject areas within each are all set by the regulation rather than chosen by a prep vendor. The mock exam is built to that structure.
Each test for each type of certification must include at least 25 questions drawn from the core group and at least 25 questions drawn from each relevant technical group. These questions must address the subject areas in appendix D of this subpart.
Appendix D names the subject areas in each group and says only that questions “should be divided in order to sufficiently cover each topic within the Group”. It does not apportion them. Superheat divides each group equally across its named areas, which is a product decision and is recorded as one rather than presented as a regulatory figure.
The regulation prescribes no time limit; certifying organizations set their own. The time allowance in the mock exam is therefore a product default too.
When the regulation changes
One question is one file in version control. A change to the regulation therefore produces a reviewable diff rather than a silent edit to a database row: the new snapshot is fetched, the verifier re-runs every quote against it, and any item whose quoted text no longer matches is flagged rather than quietly served.
The compiled bank the app ships with is a build artifact, never hand-edited. The app checks for a content update on launch when it has a connection and applies it in the background; if it has no connection, it keeps using the bank it already has, because a study session that fails because a server is unreachable is a broken product.