Public benchmark
60 new handwritten problems that Lens never saw while we tuned it, from Class 3 arithmetic to PhD analysis, across 11 subjects. Each page has one planted mistake, no mistake, or is an open essay. We ran every one through the live app and publish every result, misses included.
Real phone photos
Seven real photos of a student's notebook, taken in normal light, three of them sideways. These are the hard cases: dim light, stacked fractions, two columns, and scribbled-out ink.
| Photo | What's on it | Truth | Lens |
|---|
Every item
How we scored it
- A planted mistake counts as found only if Lens names the first wrong line (or quotes its text). Flagging a later line, or saying the page is clean, is a miss.
- A clean page counts only if Lens claims no mistake. Any flag is a false positive.
- An essay counts only if Lens says it is open-ended and does not invent a "wrong line".
- The second look is a separate Muse Spark check that runs only when a mistake is flagged. It re-reads the photo and drops flags it cannot confirm (crossed-out ink, notation, unreadable lines). Its numbers here replay the shipped code on the same first-pass output.
Limitations
- The 60 pages are rendered handwriting (seven handwriting fonts, blue and black ink, ruled and plain paper, photo noise), not 60 different students. Real photos are harder; that is why we show them separately.
- One run per item. Model output varies run to run, so expect a few points of swing either way.
- Planted mistakes are single and deliberate. Real work can hold several slips; Lens reports the first one.
- Written by the team that built Lens. The images, truth lines and answers are all here so anyone can check us.