How Chiron Works

How Chiron Works

Built by a factory

Chiron was not hand-built. It was produced by AssemblyZero, an AI-native software factory: the human directs the work, the factory writes the software, and the human reviews and approves what ships. Chiron is the evidence that this produces real, working software.

The content loop

For most domains, questions are made the same way software is made in an AI-native process — generate, review, approve:

1. Generate. The factory drafts a question from an authoritative source passage, carrying the exact citation.
2. Review. A human curator reads each draft against its source. This is the human-in-the-loop gate: nothing reaches a learner without a person's approval.
3. Approve into the bank. Approved questions enter the study bank with their explanation and source.
4. Revise, when needed. If the curator rejects a question with a specific instruction ("make the distractors less obvious", "cite the sub-section"), the factory rewrites it from the same source and returns it for re-review — so the curator directs the generator, not just gates it.
5. Study. Learners drill approved questions and see the answer, the reasoning, and where it comes from.

What each question carries

A Chiron question is more than a stem and four or five options. Before it reaches a learner it is tagged and checked:

• Cognitive level. Each question is tagged with its Bloom's-taxonomy level, its Webb's Depth-of-Knowledge level, and its pedagogical goal, and each option is tagged with the structure of the misunderstanding it represents (SOLO). How Questions Are Classified explains all three and why the wrong answers are graded too. This is what lets Chiron say, as data rather than opinion, that a question is mere recall — and drop it. On the AIGP bank, that tagging drove the removal of 51 recall-only questions.
• A grounding passage. Most questions show the exact source sentence the answer rests on, extracted verbatim from the source. The extractor refuses to paraphrase: if it cannot find the answer stated word-for-word in the source, it declines rather than inventing a plausible-looking quote. So the learner can verify the answer against the source instead of trusting it.
• A verified citation. Every banked question is checked to have a complete, real source citation. Automated audits refuse to let a question ship if its provenance is missing, if its explanation leans on a source that is not cited, or — for the commercial domains — if it touches a copyrighted source it should not.

The result is that the sophistication is not just in generating questions; it is in the instrumentation and the checks around every question.

Two kinds of domain

• Generated and curated — Patent Agent, AIGP. The factory drafts; a human curates.
• Verbatim public pool — Amateur Radio. The FCC exams are drawn from fixed, public question pools published by the NCVEC. These load verbatim, with no generation and no review, because they are already the official exam questions. The pool's own question identifiers are kept as the citation.

Sources and copyright

The commercial domains use public-domain sources only, enforced when questions are generated and again when they are published. Copyrighted material is studied by the author but never fed to the generator for a sellable question.

Where it runs

The study-and-review website runs on the edge (a Cloudflare Worker) and is the review desk. Generation runs through the author's own Claude command-line tool. The factory sits on the author's side; the website is where questions are reviewed and studied.