DacteloMilestone: Published:
By the Dactelo team at Taktek.
Keep the test away from training
A handwriting reader needs examples to learn from and separate examples to be tested on. Reusing the same ruling on both sides can make its results look better than they are. Dactelo keeps lines from a source ruling in the same split, so nearby writing does not slip into both training and evaluation.
The locked sample contains 2,970 line units from 110 pages, spanning nine court labels and ten decades. A first pass selects 330 lines across all 110 pages. That gives the correction effort a defined starting point. Selection alone does not make the lines verified reference text. Each still needs a reading checked against its image.
Count errors that change meaning
The standard requires at least 300 human-corrected lines from at least 100 pages before a result can qualify beyond a pilot. Machine drafts and unresolved disagreements do not count as that reference. On 28 October, the selected set had no human-corrected text values. We therefore had a plan for evaluation, rather than a completed accuracy measurement.
Character and word error rates describe how much the output differs from the reference. They can still hide a serious mistake inside an otherwise readable page. Dactelo also requires a separate account of wrong or missing names, dates, amounts, article numbers, negation and the ruling’s disposition. A low average cannot excuse a silent change to one of those items.
Use a score for the task it measures
The proposed gates include a character error rate of at most 5% and a word error rate of at most 10%, with coverage checks across the sample. These are requirements for helping prioritise human review. We have not announced that a Dactelo handwriting model meets them. Passing them would still leave the source image and published ruling as the controlling references.
For readers, the next useful evidence is a completed evaluation with its sample coverage, raw errors and material mistakes shown together. Until then, the public examples show individual reviewed readings. You can inspect those examples and the workflow without having to treat them as a benchmark for an entire archive.