exam.ninja

Exam.Ninja · Editorial policy

Methodology and editorial policy

Last reviewed 29 September 2026

How the questions on Exam.Ninja are specified, drafted, verified, validated, reviewed and corrected, where AI fits into that, and who is responsible for them. It applies to every question bank on the site, including DipIMC.Ninja and MSRA.Ninja.

Who is responsible for the questions

Each question bank has a named editor. The editor decides what is published, answers for every question in the bank, and has the final say on corrections.

AI-assisted, editor-led: how the pipeline works

Exam.Ninja uses AI inside a structured, multi-stage authoring and verification pipeline. It is not a chatbot writing questions on request. Generation is constrained by the exam's published blueprint and a formal authoring specification, grounded in named primary sources, and passed through independent verification, deterministic validation and human editorial oversight before a question reaches a candidate. Each stage has a different job, and a question that fails any of them does not move on.

  1. Blueprint-led specification. Work starts from coverage, not from a prompt. The bank is measured against every unit of the exam's published curriculum and domain structure, and new items are specified where coverage is thin: the unit, the domain, the difficulty band and the decision the question must test.
  2. Source-grounded drafting. Items are drafted with AI assistance against the current UK source hierarchy for that topic: national guidelines and consensus statements first, then legislation, then standard texts at edition, chapter and page. The draft must carry its citations, with an anchoring passage wherever the source allows, so every claim in the explanation is tied to a source a candidate can open.
  3. Independent verification. A separate verification pass re-opens every cited source, confirms the page numbers, sections and links, checks that the keyed answer follows from the cited text, and looks for a second arguable answer. Where a newer guideline and an older text disagree, the item is aligned to current guidance and the explanation says so.
  4. Deterministic validation. Structural rules that do not depend on any model are then enforced in the database: option counts for the exam, a keyed answer inside them, a domain belonging to that exam, references in a standard form, and for a station a complete decision tree in which every path ends and every option is scored and explained.
  5. Editorial accountability. The bank's named editor sets the specification, owns the source hierarchy and answers for every published item. Publication is a recorded, versioned event, never an automatic side effect of generation.
  6. Human-in-the-loop revision. After publication, candidates' flags and question reviewers' audits feed a revision loop. Reviewers can work through an item with an AI assistant that sees the item, the flags and its audit history, but never who raised them. The assistant can only propose a revision; the proposal is checked against the same validation rules and the item's current version, and a named human reviewer decides whether it is applied. Every applied change increments the version number and is written to an audit log.

Restraint under uncertainty. The pipeline is designed to withhold rather than overstate. An item whose sources conflict, whose answer is arguable, or which fails verification is rewritten or held back, not published with a caveat. A published item that is later shown to be wrong is corrected or withdrawn, and the change is logged.

Human judgement stays in charge. AI accelerates drafting, verification and review; it does not decide what is true. The editor and reviewers do, with the primary sources open. This page describes the methodology rather than the underlying models or infrastructure.

How a question is written

How answers are sourced

Every explanation ends with "Where this comes from": the sources the question was written from, precise enough to check.

Where an older textbook and a newer guideline disagree, the explanation says so and follows the current guideline. Ten Second Triage replacing the triage sieve in 2023 is one example.

Checks before a question is published

Each question is checked independently before it is published: the check re-opens every cited source, confirms page numbers and links, and looks for a second arguable answer. Before a question is published it is checked against a set of automated rules: the right number of options for the exam, a keyed answer that is one of them, a domain that belongs to that exam, and references recorded in a standard form. A station must also be a complete path: every decision leads to another step or to the end, and every option carries feedback and a score. Anything the rules report is fixed before publication.

Review after publication

Corrections and withdrawals

A question found to be wrong is corrected or withdrawn, not quietly left in place. A withdrawn question stops being served but is kept on record, so past attempts still make sense. If you think something is wrong, flag it in the app or contact us.

Keeping up to date

Guidelines change and so do exam arrangements. Pages about an exam's dates, fees and format say when they were last checked against the organising body's own pages, and each bank shows when its content was last updated. Where the organising body's current rules differ from anything here, the organising body is right.

Independence

Methodology FAQ

Does Exam.Ninja use AI to write questions?

Yes, as one stage of a larger pipeline. Items are drafted with AI assistance to a blueprint-led specification and against named primary sources, then independently verified against those sources, validated by deterministic rules and held to account by a named human editor.

Can AI publish or change a question on its own?

No. Publication is a recorded editorial step. After publication, an AI assistant can only propose a revision; the proposal must pass the same validation rules against the item's current version, and a named human reviewer decides whether it is applied. Every change is versioned and logged.

What happens when sources conflict?

Current UK national guidance takes precedence over older texts, and the explanation says where they differ. An item whose answer remains arguable, or whose sources cannot be reconciled, is rewritten or held back rather than published with a caveat.

Does this page name model vendors or infrastructure?

No. It describes the methodology and the standards that shape each question, not the underlying models or infrastructure.

Not clinical guidance

Exam.Ninja is for exam revision. In clinical practice, follow your own organisation's guidelines.

Choose your examContact us