DacteloBy the Dactelo team at Taktek.
A first comparison on 35 Lebanese court pages
On 21 October 2026, Dactelo compared Tesseract Arabic with three Claude readers on 35 court pages. The selection contained 30 handwritten pages and five typed controls, spanning 1953 to 2015 and five courts. It gave the project a concrete starting point before choosing a training architecture.
The Claude readers agreed much more closely on typed text than handwriting. Pairwise agreement was 97–98% on typed pages and 55–69% on handwritten pages. These figures measure similarity between outputs. Without independently corrected reference text, two readers can agree on the same mistake.
The pilot preserved the outputs and comparison method. Its useful result was a diagnosis: handwriting required image-based correction and a separate evaluation set. It did not establish a recognition accuracy rate or demonstrate that a new model would outperform available systems.