EXP-066 · live demoGet your records read →
Handwriting OCR · Met Office rainfall sheets, 1860s to 1950s

Can AI read a century of handwritten British rainfall?

Ten real ten-year rainfall sheets, one from each decade, filled in by hand at weather stations across Britain. Four open OCR models read every page, and each number is checked against the copies made by the volunteers of the Rainfall Rescue project. Pick a decade and a model; hover any cell to see who read what.

Decade

–

Read by
Read rightRead wrongMissed
This sheet: right number, right year

All ten sheets

The leaderboard

Every handwritten amount and yearly total on the ten sheets, compared with the volunteers' copy to the hundredth of an inch. A cell counts only when the right number sits under the right year; a figure slipped into the next year's column is wrong. Blank cells aren't scored.

ModelRight number, right yearWrongMissedSeconds a pageSize · licence

The sheet checks itself

Every column has its own answer key

Each year's twelve months should add up to the total written at the bottom. When a model's reading doesn't, something in that column is wrong.

How we built it

From scanned ledgers to a checked table

    Limits

    • Ten sheets is a small test: the scores show how each model behaves on real archive pages, not a ranking to three decimal places.
    • The volunteers' copies are the yardstick. They were checked by other volunteers, but they are human copies too, and a few cells may be disputed.
    • Pages were matched to their transcriptions using one model's reading (PaddleOCR-VL's), so a sheet it could not read at all could not enter the test. That may flatter it a little.
    • Each model got the same page size and the same length limit, and the prompt its own maker recommends. A model that loops is stopped at the limit, and whatever it wrote is scored.
    • We also tried Qwen3-VL-4B, a general vision model. It needs more memory than our 12 GB card has free at this page size, so it is left out rather than shown shrunk or compressed.
    • Cell positions on the images were marked by hand from the printed grid, so the coloured boxes can sit a few pixels off the writing.
    For archives, councils, insurers and back offices

    Have handwritten records to digitise?

    We build OCR pipelines that run on your own hardware, read tables cell by cell, and check themselves against the totals and rules your documents already contain, so people only look at the cells that need it.

    Get your records read →