Olex has spent years proving what examiner-level marking looks like for GCSE English. Now we are doing the same work for Maths and Science, and it is a different kind of challenge
English marking is built around extended, evaluative writing, where a student’s argument, structure and use of evidence all need weighing against a mark scheme. Maths and Science bring a different challenge. A one-mark recall question sits alongside a six-mark evaluative response. A graph, a handwritten algebraic proof, a biological explanation and a multiple-choice answer, all place very different demands on a marker, often within the same paper.
That difference is exactly why we do not simply point our English marking approach at a new subject and assume it translates. We rebuild the validation from the ground up for each one, and it happens in two stages: before we mark, and afterwards.
Before we mark: predicting the struggle
When people talk about AI marking, the conversation typically starts with one question:
How accurate is it?
We think that question comes too late.
Before we test accuracy, we first ask something different:
Where is it most likely to struggle?
So every paper goes through a pre-validation review, question by question, carried out by our own team before any marking results exist. We look at two things:
Marking risk – how much judgement is needed to apply the mark scheme accurately.
Input risk – how hard the student evidence is to interpret reliably.
These are different problems. A response can be easy to read but hard to mark because it needs linked reasoning or evaluation. A question can have an objective mark scheme but still be tricky because the evidence is handwritten or spread across a graph, table or diagram.
What we have found so far
We’ve now reviewed thousands of marks of GCSE Maths and Science, across Combined Science and the separate sciences, at Foundation and Higher tier.
Maths is dominated by method and process marks, where the challenge is deciding whether the required working has actually been shown. High marking risk appears to be rare.
Science is more mixed. Questions range from short recall and calculations to extended responses where several ideas must be linked or evaluated, which creates different levels of risk across Biology, Chemistry and Physics.
We expected more objective mark schemes to make Maths and Science easier to mark consistently. The reality is more complicated. Maths and Science papers bring new questions, specific acceptable answers and different mark schemes every year. Trust leaders have told us that across a large pile of scripts, fatigue means even multiple-choice answers can be mismarked or the precise requirements of a mark scheme overlooked. That may help explain some of the gaps we’ve seen between original teacher marks and independent reviews. A mark scheme can be objective without being easy to apply consistently.
After we mark: what the results tell us
Once a marking run is complete, the task is not simply to look at the headline accuracy figure and decide whether Olex “worked”.
The more useful question is:
Where did the disagreements happen, and why?
So we compare Olex’s marks against a benchmark, question by question, not just on paper totals. We look at exact agreement, agreement within one, two or three marks, average error, over- and under-marking, and how performance shifts by question type, tariff and predicted risk.
Then every meaningful disagreement is adjudicated. Was Olex wrong, or the benchmark? Was the rubric too vague or the student response ambiguous? Was the key evidence buried somewhere on the page? A raw disagreement is not automatically a failure.
What we have seen in practice
Two patterns stand out from our validation runs so far.
First, some apparent “errors” disappeared once the original benchmark marks were checked carefully against the official mark scheme.
Second, the disagreements that remained, typically clustered around the questions we had flagged as higher risk before any script was marked. As predicted risk rose, agreement fell. The questions we expected to be hardest were usually the ones that proved hardest.
When the original mark changes
Some disagreements between Olex and the original marker have led to marks being reconsidered.
On a three-mark Biology question, the teacher awarded zero and Olex awarded two. The trust’s STEM lead reviewed it independently and awarded one. When it was escalated to the trust-wide Science lead, they awarded two, matching Olex.
In Maths, Olex was repeatedly awarding more marks than the original teacher. On one four-mark question, the teacher gave one and Olex gave four. The trust’s Network Lead agreed that four was correct, because the original marking had missed credit for the student’s working. We then requested a full re-mark of the same 80-mark paper:
Original teacher: 16 | Olex: 42 | Network Lead re-mark: 48
Olex was six marks from the re-mark. The original mark was 32 away.
The reviewers weren’t shown Olex’s mark beforehand, and these were the first cases we escalated, not examples picked out of hundreds of disagreements. It’s one paper, so it isn’t proof on its own, but it shows that a disagreement doesn’t automatically mean the AI is wrong. Sometimes it means the original mark needs another look.
The validation process isn’t just helping us assess AI marking accuracy. It’s also revealing where existing marking can benefit from greater standardisation.
This turns validation from a scorecard into a diagnostic tool. It shows us where marking is reliable, where a rubric needs tightening, where handwritten or visual evidence causes problems, and where human judgement is genuinely difficult. The strongest validation isn’t the one with the most flattering percentage. It’s the one where every disagreement has been understood.
Agreeing what accurate means
So how accurate is AI marking for Maths and Science? The honest answer is that it depends on what accurate means, and trusts don’t all agree.
We’ve also found that trusts can have very different expectations of what accurate marking looks like. On a 70-mark Science paper, three trusts have described acceptable marking tolerances of three, four and seven marks respectively. That’s a range from around 4% to 10% of the available marks. It shows why establishing a shared understanding of marking accuracy matters, rather than assuming every school or trust is working to the same standard.
What this means for teachers
The validation work is also starting to shape how trusts think about using Olex in practice.
A few trusts have now purchased Olex to mark Maths & Science mocks, and we’ve started exploring what a blended approach could look like.
One option is for Olex to take on the low- and medium-risk questions, while teachers continue marking the higher-risk responses.
Importantly, this isn’t because Olex cannot mark those questions to the required standard. Because it can. It’s about making sure teachers continue to exercise the professional judgement that good assessment depends on.
The pre-validation process gives trusts a way to make those decisions question by question, rather than treating every question on a paper in the same way.
The same risk analysis that helps us validate marking accuracy, can also help trusts decide where AI can reduce workload and where teachers should continue exercising and developing their expertise.
If you’d like to see how Maths and Science will work for your department, book a demo.
