Notes on marking, evidence and trust.

How we measure marking accuracy

Most AI marking claims are surveys. A vendor asks teachers whether they found the marking useful, reports a percentage, and moves on. That number tells a school nothing about whether the marks were right.

We measure something narrower and harder. Every script Luminect marks keeps the AI's original band scores. When a teacher reviews that script, their confirmed marks are stored alongside — not on top of. That gives us two independent judgements of the same piece of work, and the distance between them is the only accuracy figure worth publishing.

Why we report band scores, not scripts

A script counts as touched if a teacher changed any single thing on it — a comma in the feedback, one band in one dimension. Reported that way, agreement looks far worse than it is, because a script with four dimensions marked correctly and one adjusted counts as a failure.

So we report the share of individual band scores teachers left unchanged. It is a steadier figure, and it is the more honest answer to the question a head of department is actually asking: how much of this marking did my teachers accept?

What we withhold, and when

Below a minimum sample of vetted scripts, we publish no accuracy figure at all. "Teachers kept 100% of our marks" across a handful of scripts is an artefact of a small sample, not a finding, and a school that acts on it has been misled. The counter on our home page switches itself on once the sample is large enough to mean something.

The same discipline applies to the rest of the strip. Volume figures are real counts, never padded. Demo schools and internal test accounts are excluded, because they share a database with real ones and would quietly inflate everything. Teacher-hours saved is the one derived number, and it ships labelled as an estimate.

Why every AI mark stays teacher-reviewable

Nothing Luminect generates — a mark, a worksheet, a diagnosis — reaches a student before a teacher has reviewed it. This is a rule in the platform, not a preference a school can switch off.

The commercial case for removing that step is obvious, and we have declined it deliberately. Three reasons.

A mark is an act of professional judgement

When a school reports a grade to a parent, a teacher stands behind it. Automating the labour of marking is legitimate. Automating the accountability is not, and no school should be asked to defend a number nobody in the building produced.

The review step is the measurement

Every teacher correction is a labelled disagreement between the AI and a professional who knows the student. Remove the review step and the accuracy figure in the previous article becomes unmeasurable. The oversight is not friction sitting on top of the product; it is the instrument that tells us whether the product works.

Wellbeing flags travel one way

The same principle governs the wellbeing layer, more strictly. When Luminect notices signs of distress in a student's work, it routes them privately to the counselling team. Never to the class teacher. Never onto the academic profile. And the system never advises the student — it has no business counselling a child, and it does not try.

A safety net, not surveillance. The distinction is who receives the signal, and what happens to it afterwards.

Examiner insights: what separates bands in practice

A rubric tells you what a band requires. It does not tell you what students actually do to miss it. Those are different documents, and the second one is where marking gets hard.

Behind Luminect's English marking sits a corpus we built and verify by hand, organised paper by paper across Cambridge First Language English and Second Language English — Directed Writing, descriptive and narrative Composition, the informal email, and the essay, article, report and review tasks.

Four lenses, per paper

Each task is documented four ways: the weaknesses examiners flag repeatedly, what actually separates a stronger response from a weaker one, the advice examiners give for improvement, and what the best responses do that the merely competent ones do not. Alongside those sit the concrete error patterns — the specific, recurring moves that cost marks.

That structure matters more than its length. It is the difference between a marker who can say a response is Band 3 and a marker who can say why it is not Band 4 yet, and what the student should do about it. A student who receives the first learns nothing. A student who receives the second has a next step.

Provenance is part of the work

Every entry in the corpus carries where it came from and how it was checked. An AI marker trained on unattributed advice is repeating folklore. Ours is grounded in what examiners have published about real cohorts, verified by teachers who have marked those papers.

We do not publish the corpus itself. Schools running Luminect see its effect in every piece of feedback returned to a student.