The UK A-Level Grading Algorithm Fiasco

With exams cancelled, teachers provided a centre assessment grade for each student and a within-subject rank order. Ofqual sought comparable outcomes and was directed to avoid an uncontrolled rise in results relative to previous years. The statistical standardisation process used historical centre outcomes and information about students’ prior attainment, with rank order helping place candidates within the distribution. The model was not a neural network scoring exam scripts; it was a statistical procedure designed for the emergency. Ofqual’s technical materials treated small cohorts differently because there was less statistical evidence to rely on, blending centre assessment information and statistical evidence on a sliding scale. It is inaccurate to say every small class was completely exempt or that this rule automatically protected every independent school. Institution size, subject entry and historical data interacted. Initial calculated results were issued in England on 13 August 2020. Ofqual’s later analysis found that about 39% of A-level entries had a calculated grade below the CAG, about 2% had a higher calculated grade, and 59% matched; percentages are rounded. Separately, Ofqual reported that 96.4% of initial A-level grades were the same as or within one grade of CAGs. These are entry-level comparisons, not a count of students who failed. After concerns over anomalies and public confidence, on 17 August government and Ofqual announced that students would receive their centre assessment grades for GCSE, AS and A levels. Under the announcement, students received their CAGs, but a calculated grade that was higher remained in place. The final result was not always the CAG. This case is useful for understanding group-level standardisation versus individual evidence. A model can match a distribution target while failing to represent what a particular student would have demonstrated in an exam. The controversy also illustrates how public legitimacy, appeals and school-level historical patterns matter when high-stakes scores are assigned.