Where the learning actually happens
Real learning happens in the working-out. A student does not understand algebra because they wrote the right number at the bottom of the page; they understand it because of the steps that got them there. Yet the working-out is precisely what most tools cannot read. A learner photographs a page of handwritten algebra, and today's apps see only the last line. They cannot tell whether the method was sound, where the reasoning slipped, or why marks would be lost.
Meanwhile teachers spend hours marking working by hand, tracing each step, deciding where a method mark is earned and where accuracy is lost. It is slow, repetitive and easy to do inconsistently at the end of a long day.
Project Scriptura is ILM AI applied research into what we call stroke-level mathematical cognition. It is the engine behind ilmino's ability to read handwritten reasoning, not just recognise a final answer.
What "stroke-level mathematical cognition" means
The name carries the whole idea in two halves.
Stroke level describes how deep the reading goes. Not just the final answer, but every handwritten line of working: each digit, symbol, fraction and diagram a student actually put on the page.
Mathematical cognition describes what the model does with what it reads. It does not simply transcribe strokes into text. It follows the reasoning those strokes encode, treating the page as a sequence of logical steps rather than a picture to be captured.
The difference matters. Transcription tells you what was written. Cognition tells you whether it makes sense, and that is what marking depends on.
The vision-language engine
Scriptura is a vision-language engine that reads handwritten maths and science from a single phone photo, follows the logic line by line, and marks the method the way an exam board does. Its capabilities break down into four parts that work together.
Reading real handwriting
Real work is messy. Digits run together, symbols are ambiguous, fractions stack unevenly and diagrams sit in the margin. Scriptura is built to read that reality, from an ordinary phone photo, rather than the clean printed input most systems assume. This is the foundation: if the engine cannot read genuine handwriting, nothing else it does can be trusted.
Stroke-level parsing
Once the page is read, the engine parses it as working, following the argument line by line rather than jumping to the final answer. It reconstructs the chain of steps the student took, so that each line can be considered on its own terms and in relation to what came before.
Mark-scheme aligned marking
Scriptura awards marks the way examiners do, using method and accuracy marks such as M1, A1 and A0. A student who chooses a valid method but makes an arithmetic slip still earns the method credit, exactly as an exam board would allow. This is the point that separates a marking tool from an answer checker: it rewards how a student reasoned, not only whether the last line matched.
First-error localisation
When something goes wrong, the engine finds exactly where. It locates the first error, effectively an argmin over the steps: the earliest line at which the reasoning broke. It then explains the fix the way a good teacher would, so the learner sees not just that they lost a mark but why, and what to do differently next time.
The research behind it
Scriptura sits across several active research problems, and we are candid that this is applied research feeding a product rather than a solved task:
- Recognising and interpreting handwritten mathematical and scientific reasoning, where the input is genuinely noisy.
- Step-wise verification and first-error localisation, so the engine can say not only that an answer is wrong but where the logic first failed.
- Aligning automated marking with real exam-board mark schemes, including AQA, Edexcel, OCR and WJEC, so the marks it awards match the ones a student would actually receive.
Throughout, the AI assists and the teacher reviews. Every mark can be overridden, and teacher judgement is final.
Where it fits in ilmino
Scriptura is the engine that upgrades ilmino's Snap & Solve and automated marking. It moves both from checking answers to understanding how a student got there. A snapped question no longer returns a verdict on the final line alone; it returns a reading of the whole method, marked step by step, with the first error found and explained. The same engine gives automated marking its examiner-grade consistency.
What it means for your institution
For a department, the value of stroke-level marking is practical rather than abstract.
Marking workload
The slow, repetitive part of marking, working through each line of each script, is exactly the part Scriptura automates. That is time returned to teachers: less of the evening spent tracing method marks by hand, more of it spent on the teaching and feedback that only a person can give.
Consistency across a department
Marking standards drift from one member of staff to the next, and from the first script of the evening to the last. An engine aligned to real mark schemes applies the same standard to every script, in every classroom. Early-career teachers and specialists mark to the same line, which is difficult to achieve any other way and shows up directly in the fairness of what learners receive.
Teachers in control
Automated marking is a first pass, not a verdict. Teachers see the marks and the located errors, review them, and can override any of them. The engine does the repetitive reading; the professional judgement stays with the person who knows the class. Teacher judgement is final, by design.
In short
- Reads the working, not just the answer. Every handwritten line, from an ordinary phone photo.
- Marks like an exam board. Method and accuracy marks aligned to AQA, Edexcel, OCR and WJEC.
- Finds the first error. The earliest point the reasoning broke, explained the way a teacher would.
- Gives time back. The slow, repetitive part of marking, automated with examiner-grade consistency.
- Keeps staff in control. Teachers review every mark and can override any of them.
If you would like to see Scriptura inside ilmino and discuss what stroke-level marking could mean for your department, book a demo.
The full picture of ILM AI's applied research sits alongside the product at ilmai.co.uk.
