Started by markers, not by a growth team

Exam marking is a strange job. You spend weeks reading scripts by students you will never meet, awarding and withholding marks under a scheme that is unambiguous once you know it and near-invisible if you do not. What becomes obvious quickly is that most lost marks are not knowledge gaps. They are habits, unshown working, an answer to the topic instead of the question, an analysis that never leaves description.

Those habits are invisible from the inside. A student who self-marks awards themselves the mark because they know what they meant. A parent looking at a score cannot tell whether a 52 was 52 marks of near-misses or 52 marks of solid work plus one collapsed topic. The information that would fix it exists only in the margin annotations that exam boards never send back.

So we started sending them back. A student sits a paper at home, an examiner marks it properly, and the annotated script returns with a topic breakdown and three things to do next. That was the whole idea in 2019 and it is still the whole idea now, 31,400 scripts and 180 partner schools later.

What we will not do

We do not mark with AI. Models are good at plausible feedback and bad at knowing whether a specific line of algebra earns a method mark; the failure mode is confident and wrong, which is the worst possible thing to hand a student. AI at GCSE Feedback sorts uploads and checks image legibility. It never touches a mark.

We also do not promise grades. Marked practice moves grades on average, and does nothing at all for a student who files the report unread. We would rather publish the average and the caveat than the testimonial alone.