AutoEx

The research described five stages: segment the page, read the handwriting, correct what was read, match it against the examiner’s keywords, and report on the class. It ran on a laptop, in R, for one person.

The product keeps all five and moves them to where a school actually is. The segmenter runs inside the browser, the reader is a vision model, the mark book is a web page, and the whole thing opens from a link with no account and nothing to install.

A mark with no explanation is a mark that cannot be appealed. Each question lists the keywords the student actually wrote, the ones they missed, and how each one matched — exactly, by stem, or fuzzily within an edit or two, so a misspelling still earns its mark.

Change a keyword or a weight and the whole class re-scores in front of you. Marks are capped at each question’s maximum, and the transcription stays editable, because a reader that cannot be corrected is a reader nobody will trust.

The paper evaluated the reader and reported a word error rate of 5.6%. That is about fourteen words in every two hundred and fifty — enough that a human still signs off the marks, and few enough that the signing-off is a review rather than a re-marking.

Where a page may not leave the building at all, the reader can be swapped for one that runs on the device itself. It is less accurate on cursive, and it never sends a thing.