England's 2020 A levels: a centre model applied to individual students

What happened

With exams cancelled in 2020, schools and colleges submitted a centre assessment grade (CAG) for each student and a rank order of students in each subject ¹.

Ofqual's Direct Centre Performance model predicted the distribution of grades for each school or college from its historical results in the subject, taking account of the prior attainment of this year's students, then used the submitted rank order to assign grades to individual students. Where a subject cohort was small (fewer than 15 students across the current and historical entries), the model put more weight on the CAGs ¹.

The interim report, published on results day (13 Aug 2020), stated that 96.4% of calculated grades were the same as, or within one grade of, the CAG, and that its equalities analysis showed no evidence the process had introduced bias ¹. On 17 Aug 2020 Ofqual announced that AS and A level results, and the GCSE results due later that week, would switch to centre assessment grades ². The AIAAIC Repository records the case ³.

Failure mode

The model was checked at the level it was designed for and failed at the level it was used. The headline checks in the interim report are aggregate: prediction accuracy within one grade on historical data, and an equalities analysis by group ¹. A student's grade, though, came from their school's past and their place in a rank order, so a strong student in a historically weaker centre could be held down by results they had no part in.

The small-cohort rule made treatment depend on class size: small groups kept more of their teachers' grades than large ones ¹. That is a design choice with distributional consequences, and it needed testing as one.

Which control would have caught it

An eval gate with individual-level acceptance criteria forces the question the aggregate tests skip: for which students does the model move the grade furthest from the evidence about them, and can each move be justified? A human-in-the-loop gate on large adjustments, with a contestation route before results day, turns those outliers into reviewed decisions instead of published ones. The fundamental-rights impact assessment is where the treatment of small and large cohorts would have been argued in the open.

Patterns: Eval Gate in CI · Human-in-the-loop Gate · FRIA-as-Code

The evidence that would have existed

What an auditor could have read, and the stack layer that produces it.

Obligations it touches today

As of 2026-09-24. Mappings are illustrative, not a claim of conformity.