Quill: recovering handwritten weather records

Quill uses Meta's Muse Spark 1.3 and Muse Wonder 1.0 to recover handwritten rainfall records, with checks against the original observer's arithmetic. Weather ledgers preserve generations of scientific work. Making those measurements usable takes more than scanning the paper: someone must read the handwriting and check the transcription. In 2020, 16,000 Rainfall Rescue volunteers transcribed 66,000 sheets from Britain and Ireland. Their work helped extend the official UK rainfall series to 1836. Quill builds on that work by testing how much of the reading and checking can be automated. The key is already on the page: the observer wrote a total below each year's monthly measurements. 1. Read and locate. Muse Spark reads the sheet and locates its year columns. 2. Check. Python compares each column's monthly sum with its written annual total. 3. Read again when needed. Quill crops and enlarges a failed column for another Spark read, then calls Muse Wonder if it remains unresolved. Readers see the source image, including written totals, but receive no previous answers or instructions to make a sum fit. 4. Keep the evidence. Reconciliation considers model readings and explicit decimal-point interpretations, with guards against wrong-column crops and excessive changes. The review view retains readings, statuses and image crops; CSV export includes values and statuses. On a benchmark of 169 sheets and 16,810 comparison cells, Spark's first read matched the volunteer reference on 96.75% of cells. Quill accepted 15,721 cells (93.52%) under its arithmetic or reader-agreement checks. Six accepted cells disagreed with the reference: 0.038%. These are measurements on this sample, not a guarantee of correctness. Quill explicitly flagged 976 cells for review; another 113 reference cells were blank in its output. Average model cost was $0.055 per sheet, excluding hosting and human review. A separate batch processed all 31 Stonyhurst pages in 145 seconds for $1.45, with pages running in parallel. Of 2,964 accepted monthly values, 2,958 matched the volunteer series (99.8%). An 1895-96 Kenyan daily register provides an exploratory test of a second layout: seven months match their written totals. It has no reference transcription for measuring accuracy. The aim is to help archivists and researchers recover historical measurements while keeping the source and uncertainty visible. American weather archives are a relevant next test; this submission reports results on British and Irish sheets and the Kenyan example. Quill does no task-specific fine-tuning. Existing AI transcription work by Philip Brohan and the Rainfall Rescue volunteers is credited in the README. Built during the hackathon, Nov 2-4, 2026.