Question pipeline

How we write questions

The most common lie on a test-prep site is not "this question is hard" but "this question is correct". Our own history is enough to fill a page: an external answer key wrong on roughly one question in twelve, and wrong precisely on the traps; two rounds of seventy-plus AI agents cross-reviewing each other and still letting fabricated questions through; our own checking scripts fooling us more than once. So this page is not about how careful we are. It is about mechanism: the eight gates an original question passes between drafting and release, how a student gets it taken down afterwards, and why none of those records can be edited.

Who we are Read the study guides

At a glance

Where answers come from: Solved by code for quantitative items; the model only polishes wording
Gates: Eight, automatic, fixed order, any failure rejects
Reporting: Six categories, an entry under every question, into a review queue
Retirement: An item that fails review twice is retired and never served again

This page describes the original-question pipeline. Official past papers are imported against the official answer keys and pass our own answer verification; reporting and retirement work the same way for both.

In short

For quantitative questions the answer is computed by code, not written by a model: the generator solves symbolically, each distractor is derived from a named error, and no numeral may change after polishing.
Eight automatic gates run in a fixed order; any failure rejects the item and the reason is stored per item. Review by models of the same lineage does not count as a gate, because they share blind spots.
Release is not the end: students can report under six categories, items with abnormal accuracy or clustered reports enter a review queue automatically, and an item that fails twice is retired.
Gate results, reports and decisions are appended and never deleted, so any problem can be traced back to exactly which step let it through.

4 sections1,218 words5 min read4 official sources

Part 1

Why eight gates, and not "have a teacher look"

Human review works at small scale, but it has two structural problems: reviewers tire, and nobody can spot the errors they do not know about. Every expensive lesson we learned was not "nobody looked" but "the person looking shared a blind spot with the person writing": two rounds of seventy-plus AI agents reviewing one another still let a batch of fully fabricated questions through, and an external answer key turned out to be wrong on roughly one question in twelve, wrong precisely on the traps the questions were designed around, the kind of error a reviewer following the question naturally repeats.

So we turned acceptance into gates that a machine can run repeatably and that draw on different sources: structural faults are caught by deterministic checks, wrong answers by a second, independent piece of code, ambiguity by an audit whose only job is to argue for the distractors, and model-flavoured wording by a style score. The gates themselves sit an entrance exam first: a batch of known-good official questions measures the false-rejection rate, a batch of hand-built bad questions measures recall, and if a category of bad question gets through untouched we fix that gate and rerun rather than lower the bar.

Part 2

The eight gates, in order

Cheap gates run first, expensive ones last; the first failure stops the run, and the reason is stored together with every gate already run. Each gate below comes with a real rejection.

01Gate 1

Brief and structure

Every question starts from a slot specification: syllabus reference, difficulty band, option count, archetype. The structure gate checks that the maths renders, the option count matches the slot, the options are distinct and the cited syllabus entries actually exist. Rejection: a question carried the correct value in two different options; it looked perfectly normal to the eye and was returned under "no distractor may equal the correct value".

02Gate 2

Generation: solved by code, polished by a model

For quantitative items the model does not write the question; it writes a generator: parameter sampling, symbolic solving, and each distractor computed from a named error. The model then polishes the template into natural English, and every numeral is compared before and after; one change is a rejection. Rejection: while polishing, the model "helpfully" changed a coefficient from 3 to 2. The sentence read better and the answer was now wrong. The numeral diff caught it.

03Gate 3

Deterministic provenance

The no-calculator gate thresholds arithmetic complexity (digits in products and quotients, steps to simplify a surd, size of a common denominator); the duplication gate compares vector similarity against the whole bank and within the batch. Rejections: one question required 23 times 47, pure punishment for a student without a calculator, returned as "ugly numbers"; another, presented as original, sat at near-identical similarity to a real past-paper question the model had simply reproduced from memory.

04Gate 4

Adversarial check: arguing for the distractors

A dedicated auditor does one thing: for every wrong option it tries to construct a reading under which that option is correct. If any distractor survives under a reasonable alternative reading, the question has a hole and is rejected whole. Rejection: a function question never stated the sign of a parameter; for negative values a second option was also correct, and whether a student got it "right" was luck. That is why ambiguity is our highest-priority rejection.

05Gate 5

Independent answer verification

A second model receives only the stem and options, never the generator or the answer, and writes verification code from scratch; the option it computes must match the generator. Two models from different lineages then solve the question blind: both right passes, one wrong goes to a priority human list, both wrong is rejected outright. Rejection: the generator dropped one sign branch of a square root; the verification code produced a different option, the two disagreed, and the generator was fixed.

06Gate 6

Solution check: anchored expansion

A solution may not re-derive the answer; it may only expand the derivation already verified at gate five, and then passes the solution-quality gate: the conclusion must match the answer, hedging phrases are banned, length is capped. Rejection: a solution argued fluently for several lines and concluded with a different option; the model had changed approach halfway through, and the conclusion-consistency check sent it back.

07Gate 7

Difficulty calibration and style

The blind-solving stage also yields difficulty: an item both models answer instantly with high confidence is moved down a band. The style gate scores five items (concision, British spelling, SI units, no redundant information in the stem, conventional option ordering); a low total is returned with the deductions listed. Rejection: a question used American spelling and the "Which of the following best describes" cadence of US standardised tests, out of keeping with the terse British admissions-test register, and was sent back for rewriting.

08Gate 8

Human sign-off, then practice before mock

Items that pass the first seven gates get a per-item checklist sign-off, are imported into a test environment, worked through in a real browser with a real account, spot-checked, and only then promoted to production. A new paper first spends two to four weeks in the practice centre so real answer data can speak; items with abnormal accuracy or clustered reports are pulled automatically, and only clean papers may be used for high-stakes mocks. Rejection: a paper that had cleared every gate turned out, in the browser, to have its numbering swallowed by a global stylesheet, leaving over a hundred questions unanswerable. It was caught before release.

Part 3

After release: reporting, review and retirement

Every question view carries a "report a problem" entry; students choose one of six categories and need not write an essay. Reports go straight into the database and into our review queue. If several students report the same item, or its real accuracy drifts far from the estimate (too high usually means a missing condition, too low usually means a wrong answer or stem), the system pulls it from the practice pool for review without waiting for anyone to click.

Image

Cut off, garbled or failing to load

Answer

The key is wrong

Stem

Text or formula error

Solution

The worked solution is wrong

Calculation

A numerical or arithmetic error

Missing information

The question lacks a condition or a piece of information

Fail twice, retire: an item that enters the review queue is re-run through the gates and a human check, and returns to the pool only once fixed. If the same item enters the queue a second time and still cannot pass, it is retired for good, never served again in any paper or practice set, and the retirement record is kept.

Part 4

The immutable record

Three things are appended and never edited or deleted: every gate result and rejection reason for every item; every student report and its handling status; every decision to pull, fix, return or retire an item. The record is not for show. It exists so that "who let this question through, and at which step" can always be answered. We sample and try to falsify the numbers our own gates report before we trust them, because our checkers have fooled us before, and the only thing that can catch a checker is its own record.

The mechanism does not promise that questions are never wrong. It promises that a wrong question gets found, gets handled, and leaves a trail.

Billing terms

What you are actually buying

One-off, per subject: £58.80 for one subject, £108.80 for any two. Not a subscription.
Access does not expire: Everything unlocked stays unlocked, including after the update window ends.
Updates for one year: New material in that subject is included for one year from purchase, renewable at £18 per subject per year.

We avoid the phrase "lifetime access": access itself does not expire, but new material is included for the update window, and describing that as lifetime would misstate what you get.

Go deeper

Sources

Official sources

Spotted something wrong on this page? Write to contact@frontierace.ai with the page address and the sentence in question.

FAQ

Frequently asked questions

How do you distinguish original questions from past papers?

Paper titles and source labels say which is which. Where an exam board has never released a past paper we say so, and our own papers are labelled original, never dressed up as past papers.

Why not review everything by hand?

We do, but only as the last gate. The first seven are handled by repeatable machine gates from different sources, because a human reviewer often shares the blind spot of the model that wrote the question, and humans tire.

How quickly is a report handled?

A report is stored and queued immediately; items reported by several students or showing abnormal data are pulled automatically without waiting for anyone. The outcome is written to the record once review completes.

If an item is retired, what happens to my attempts on it?

Your attempts are kept, but a retired item no longer counts towards readiness or weakness statistics, so one bad question cannot skew your picture.

Do conceptual questions (biology, chemistry) pass the same gates?

The same eight, but answer verification becomes option-by-option fact checking plus an out-of-syllabus check: two models from different lineages judge each option from the stem alone, and every concept an option touches must map to a syllabus entry. This is the weakest link in the line, so human sign-off samples conceptual items more heavily.

Are you connected to UAT-UK, Pearson or any university?

No. FrontierVUE is an independent practice platform with no affiliation, endorsement or partnership with any exam board, test operator or university; syllabus references point to the publicly available official documents.

FrontierVUE is an independent practice platform. It is not affiliated with or endorsed by UAT-UK, Pearson, OCR, the University of Cambridge, Imperial College London, or any official admissions-test owner.

FrontierVUE 是独立的备考练习平台，与 UAT-UK、Pearson、OCR、剑桥大学、帝国理工学院或任何官方入学考试主办方均无隶属或背书关系。

Judge it for yourself

Every subject has a free paper you can sit end to end, with the same worked solutions and translations as the paid ones.

Start free See pricing