Every August, students discover that an A grade in maths required 78 per cent of the marks while an A in physics needed 62, and that both figures differ from the year before. The instinctive reading is that something has gone wrong, or that standards are slipping. The accurate reading is that the system is doing exactly what it was designed to do: boundaries move because grades are deliberately anchored to the cohort, not to the paper.
Start with the marking itself, which is the least mysterious part. When scripts leave the exam hall they are scanned, cut into individual questions, and distributed electronically to examiners, most of them serving teachers, who mark on screen against a mark scheme refined at a standardisation meeting. An examiner might see thousands of responses to a single question and nothing else. Quality control runs continuously: "seed" scripts already marked by senior examiners are slipped anonymously into each marker's queue, and anyone whose marks drift beyond tolerance is stopped, retrained or removed, with their earlier work re-marked. The output of all this is a raw mark out of, say, 100 per paper. What that raw mark is worth in grade terms is a separate decision, and it comes later.
That decision happens at an awarding meeting, held by each exam board — AQA, Pearson Edexcel, OCR and WJEC Eduqas in England and Wales — once marking is essentially complete. Senior examiners look at archive scripts from previous years, this year's scripts at candidate boundary marks, and, decisively, statistical predictions. The judgement they are making is not "how many marks should an A require in the abstract" but "at what mark on this specific paper does this year's work match last year's A". No two papers are equally difficult, however carefully they are written. If this year's paper turned out harder, the same quality of candidate scores fewer raw marks, and the boundary drops to compensate. A boundary falling from 68 to 61 is not generosity; it is the correction that keeps the grade meaning the same thing.
The statistics that discipline this process are the core of comparable outcomes, the approach Ofqual, the exams regulator for England, formalised around 2011. Each board receives predictions of the proportion of its entry expected to achieve each grade, built from the cohort's prior attainment — for A-level, principally their GCSE results two years earlier. If this year's chemistry entry arrived with a stronger GCSE profile than last year's, the prediction allows more top grades; if weaker, fewer. Boards must land close to these predictions or justify the divergence to Ofqual. Examiner judgement still operates at the margins, and small-entry subjects lean more on scrutiny of actual scripts, but for large-entry A-levels the statistical prediction is the anchor.
Why the cohort, not the paper
The rationale is a trade-off between kinds of fairness. Norm referencing, used until the 1980s, fixed the percentage of each grade in advance, so your grade depended on beating other candidates in your year — a strong cohort was simply punished. Pure criterion referencing, grading only against written descriptors, sounds fairer but cannot cope with the fact that papers vary in difficulty and examiners' severity drifts, which is broadly how GCSE grades inflated year on year through the 1990s and 2000s. Comparable outcomes splits the difference: it holds the value of a grade steady across years by assuming that a cohort with similar prior attainment deserves a similar grade distribution, then letting the boundaries absorb all the noise from question papers.
The cost is that the system is opaque to the people inside it. A student cannot know in the hall what mark they need, teachers cannot promise that a given performance secures a given grade, and a genuinely better-taught cohort will find its improvement partly discounted, because the statistics expect it to look like previous cohorts. The approach also failed spectacularly when stretched beyond its design: the 2020 algorithm debacle, when exams were cancelled and standardisation was applied to individual centres rather than the national cohort, showed what happens when statistical anchoring is asked to grade people rather than to calibrate papers. Within its intended use, though, the machinery is defensible. Roughly 800,000 A-level scripts a year are marked by tens of thousands of examiners across four boards, and the grade an admissions tutor sees means close to the same thing in 2026 as it did in 2019. The boundaries move precisely so that the grades do not.
