When AI Could Help Read a Difficult Script

Experiment
Gemini and Claude transcribe 300-page archival registers for full-text searching, with readings checked against the images.
What I learned
  • a searchable transcription can still be too unreliable to cite
  • dates needed checking even after revisions to the process
  • processing a register took about 12 hours and $75 in API costs
Results
  • processed entire handwritten registers into searchable text
  • kept model outputs for comparison and error checking
You'll need
Gemini, Claude, Open Claw
Format
downloaded document images; two hours to set up; automated run of ~12hr/register
Jig by
Fred Gibbs, History

Could an LLM read the Gothic secretarial hand used in late fourteenth-century records? I had tried the specialized handwriting-recognition platform Transkribus, including training a model with 60 transcribed documents. In February 2026, a PARES image uploaded to Gemini gave me a better result. That was enough reason to try a larger batch.

Spain’s PARES provides digitized images from the Archive of the Crown of Aragon. Being able to search those handwritten registers for names and places would change how I could work through them.

From one image to a register

An example of a folio from an ACA register, ACA CR R2053 f4r. The script has been notoriously difficult and abbreviations are frequent.

An example of a folio from an ACA register, ACA CR R2053 f4r. The script has been notoriously difficult and abbreviations are frequent.

I began combining transcriptions from Gemini and Claude. By March, I was using Open Claw to download the images, send each to both models, merge their transcriptions, and save the result in a text file. A final instruction combined the files into a CSV.

Register 1819 was the first complete register I processed. Register 2053, the third, looked noticeably better. I had refined the prompts, and the models had changed between February and March; this comparison didn’t separate their contributions to the improvement.

Results and limits

A 300-page register takes about 12 hours to process, at approximately $75 in API costs. Keeping the two models’ outputs made recurring errors easier to spot and left a record I could check when the combined transcription looked questionable.

The result is useful for full-text searches. It isn’t reliable enough to cite without returning to the image. Dates remained inconsistent even after I revised the process, although I was surprised by how well the models read the script and expanded its abbreviations.

For my purposes, the immediate gain was being able to find a name or place across an entire register. The search gives me somewhere to look; the manuscript still has to supply the reading.

Using This Experiment

This experiment is licensed under Creative Commons BY-NC-SA 4.0. You're welcome to share and adapt it for noncommercial purposes, with credit, as long as you share your adaptation under the same license.

Suggested citation

Fred Gibbs. "Pipelines for Medieval Handwriting Recognition." AI Workbench, Amaranth, University of New Mexico, April 9, 2026. https://amaranth.unm.edu/ai-workbench/research/atr-paleography.html. Licensed under CC BY-NC-SA 4.0.