Inside the Rubrical build: measuring the 52%
How OpenKit built Rubrical for the DfE: a private AI stack, continuous evaluation, and a teacher signing every grade. What it took to earn the 52% figure.
Ibrahim Mizi
· 5 min read
Updated 
The Department for Education contracted OpenKit to change how formative assessment marking works, starting with GCSE Geography. The public framing was time, because marking is the load that eats teachers’ evenings. The engineering framing was trust, because a system that touches assessed work has to leave a clear record of whose work it is and how a mark was reached, and that constraint ruled out every off-the-shelf tool before we started.
What the benchmark showed
Before building anything, we benchmarked. We took GCSE Geography long-form responses already graded by senior teachers, held them out, and scored generic AI tools against the same rubric. They matched the human grader 67% of the time, which is impressive at a demo and indefensible at moderation.
Rubrical, a private stack that retrieves from the DfE content store and is evaluated continuously with the KILN framework, reached 92% on the same held-out marks. That 25-point gap is why the build went against the Department’s own content instead of prompting someone else’s model. The programme’s public target was to halve the time, and teachers report a 52% reduction in marking time in the developer’s tests, which the Department for Education records on page 98 of its February 2026 schools white paper, Every child achieving and thriving. Feedback came back around 2.5 times more detailed than teachers had time to write by hand. Teacher satisfaction measured 95%.
The teacher signs every grade
The design principle that survived every iteration is that the model drafts and the teacher signs. Typed and handwritten responses go in, rubric-aligned draft feedback comes out, and the teacher edits, overrides, and signs each grade while the moderation trail records all of it. Nothing reaches a student unsigned.
It is slower than full automation, and it is the version a trust can defend at moderation, because every mark has a named human behind it and a record of how it was reached. In regulated work the audit trail is part of what a school is buying.
The data decided the architecture
The trust had to be able to show where student work was held and who could reach it, so Rubrical runs on a private stack in a UK Azure tenant rather than a public API. OpenKit holds ISO 27001 and ISO 9001, and the deployment operates to UK GDPR.
We ran a closed pilot in three schools. The DfE data protection impact assessment passed and the ISO 27001 audit on the deployment closed before any scale-up. Fourteen schools now run it in production behind a trust-wide moderation framework and service-level agreements, and the work continues with quarterly model refreshes against new marking data and a jointly authored impact report.
What the build teaches
Rubrical is the pattern we use when a wrong answer has consequences: a private model answering from approved content it retrieves at the point of use, evaluated against marks senior teachers had already given, with a human signing every decision that matters. The full Rubrical case study carries the rollout detail, and AI marking for teachers is the teacher-facing version. It is also the approach behind our wider work on AI in education.
What is Rubrical?
Rubrical is an AI marking assistant OpenKit built for the Department for Education, starting with GCSE Geography. It drafts rubric-aligned feedback that a teacher edits and signs. Teachers report a 52% reduction in marking time in the developer's tests, a figure the Department records in its February 2026 schools white paper, Every child achieving and thriving.
How accurate is Rubrical compared with generic AI?
On GCSE Geography responses already graded by senior teachers and held out as a test set, generic AI tools matched the human grader 67% of the time and Rubrical reached 92%. The evaluation runs again at each quarterly refresh against new marking data, so the number is re-checked as the model changes.
Why does the teacher sign every grade?
Because a school has to defend its marks to governors, parents and exam boards, and a draft nobody checked cannot be defended. The teacher can override any grade, and the trail records what was changed and by whom, so a mark can still be explained months later. Full automation would be faster and would leave nobody able to give that explanation.
Was Rubrical safe to deploy on student data?
It runs on a private stack in a UK Azure tenant, answering from the DfE content store at the point of use, so student work never goes to a public model. Teachers reach it through the trust's own sign-in, inside Google Classroom or Teams, and the deployment passed a DfE data protection impact assessment before it scaled to production.
Tell us what the build has to do.
Bespoke work usually starts inside an AI Audit and Transformation, which settles what is worth building before anyone writes code. We build the system end to end and hand over the code and the IP, so there is nothing to renew each year. If we believe another route would serve you better, we will say so.
Keep reading.
The closest pieces to this one, by shared ground.
