Stroke Level Mathematical Cognition
Vision-language marking of handwritten working, down to the first error
Why this name
“Stroke level” says how deep the reading goes, not just the final answer, but every handwritten line of working. “Mathematical cognition” says what it does with it: the model doesn't just transcribe the strokes, it follows the reasoning they encode.
The problem
Real learning happens in the working-out, but that's precisely what most tools can't read. A student photographs a page of handwritten algebra and today's apps see only the last line. They can't tell whether the method was sound, where the reasoning slipped, or why marks would be lost. Teachers, meanwhile, spend hours marking working by hand.
What we're building
A vision-language engine that reads handwritten maths and science from a single photo, follows the logic line by line, and marks the method the way an exam board does (M1, A1, A0), locating the first error (argmin over steps) and explaining the fix like a good teacher would.
Key capabilities
Reads real handwriting: messy digits, symbols, fractions, and diagrams from a phone photo.
Stroke-level parsing: follows the working line by line, not just the final answer.
Mark-scheme aligned: awards method and accuracy marks the way examiners do.
First-error localisation: finds exactly where the reasoning broke, and why.
Teacher time back: automates the slow, repetitive part of marking with examiner-grade consistency.
Research focus
Recognising and interpreting handwritten mathematical and scientific reasoning.
Step-wise verification of solutions and first-error localisation: argmin_t error(step_t).
Aligning automated marking with real exam-board mark schemes (AQA, Edexcel, OCR, WJEC).
Where it fits at ILM AI
The engine that upgrades Ilmino's “Snap & Solve” and automated marking from checking answers into understanding how a student got there.
Scriptura: feedback on every step, not just the answer.