Skip to content
LEAPDocs
2026.08

OCR and scan quality

Getting scans and photographs to read reliably — for syllabi and for answer sheets.

Syllabus scans

Image-only PDFs are OCR-ed before extraction, and low-confidence pages are flagged rather than silently accepted.

Proceeding on a bad OCR pass wastes the generation step downstream. If most pages are flagged, a cleaner scan — or pasting the text — produces dramatically better gates for a fraction of the effort of correcting them afterwards.

Photographing answer scripts

  • Even light. Shadow across the page is the single most common cause of failed extraction.
  • Flat on. A steep angle distorts lines enough to break line detection.
  • Fill the frame. One script per photo, edge to edge.
  • Check focus before moving on. Re-photographing later means matching it back up by hand.
  • Keep under 10 MB per file. A phone photo is well under; a raw flatbed scan may not be.

Diagnosing a bad grade

When a paper-quiz mark looks wrong, the question is always the same: did the platform read the page correctly, or did it read it correctly and mark it badly?

  1. 1
    Open the response and read the extracted text

    Compare it against what is actually on the script.

  2. 2
    If the text is wrong, it is an extraction problem

    Re-photograph that script. Correcting the mark by hand leaves the underlying record wrong.

  3. 3
    If the text is right, it is a grading problem

    Check the criterion it was matched against, then override. Your override is final and recomputes mastery.

This distinction is why grading shows both the extracted text and the matched criterion. A blank sheet once scored full marks because extraction hallucinated answers and the marks looked entirely reasonable — the extracted text was the only thing that would have revealed it.

QR codes and capture sessions

The QR code will not scan.

Print larger, or type the URL from the sheet directly. The capture session is a plain web page — no app, no sign-in.

The capture page says the session expired.

Capture tokens are deliberately short-lived. Reopen the QR code from the assessment to start a fresh session.

I photographed the same script twice.

Duplicates match to the same attempt and the most recent capture wins. Nothing is double-counted.

Uploads stopped part way through a class.

Check the session caps: 500 MB per session and 200 files. A large class of high-resolution scans can reach these. Start a second session.