Quality, measurement and acceptance
Accuracy is a measurement, not a promise. This page states how it is measured, against what, and what of it reaches the contract.
Quality and acceptance
99.67 % character accuracy
Pre-registered blind test against the ground truth of the Deutsches Textarchiv: 17 deviations across 5,129 characters, on three pages from 1591, 1615 and 1701. 95 % confidence interval (Wilson) 99.47 % to 99.79 %; per page 99.92 %, 100.00 % and 99.28 %. Computed without post-hoc normalisation.
Measurement protocol with the full deviation list (PDF, German) - conventions, page selection and every individual finding are disclosed, including the places where the reference itself was wrong.
- Measured, not asserted. Character accuracy is measured per material class and contractually agreed as the target; acceptance testing follows the sampling procedure of the DFG Practical Guidelines on Digitisation (2022 edition). A figure from someone else's collection does not replace that measurement, which is why every project starts with a sample conversion.
- CER/WER per the OCR-D evaluation specification. On request we evaluate against your own ground truth - Unicode-NFC-normalised, compatible with the usual evaluation tools.
- Faithful to the source, no silent normalization. Historical orthography, printing errors and punctuation stay as they stand; nothing is modernized.
- Illegible passages are marked, never invented. Whatever the scan does not support appears as
<gap reason="illegible">or [illegible] in the text - never as a plausible guess. - Line-bound and image-referenced. Every text line stays tied to the source via its coordinates and is machine-checkable - no omissions, no unsupported additions.
- Raw and final version. You receive the unedited first pass and the final version; every correction stays traceable. The capture conventions agreed for a project are recorded in writing, so later deliveries can be held to the same rules.
Free sample conversion
Up to 50 sample pages from your own holdings, ideally your hardest volume year. No obligation, result with an error-rate measurement within 5 working days.