What the primary source reports
In its public OCR annotation account, RWS describes TrainAI working with an unnamed large IT company. RWS, not XIVTech, was the implementing vendor. The account discusses characters that were curved or vertical, occluded or blurry.
RWS reports a custom annotation program, annotator training and certification, and skilled annotators. It also reports 3.5 million transcriptions, 30,000 image annotations, 32 languages, more than 500 specialists, greater than 99% data accuracy and an expanded scope. These are claims from that single vendor source, not independently verified results.
The account does not supply the accuracy denominator, test method or measurement date. Its data-accuracy figure must not be presented as OCR model accuracy. The source was accessed on 30 October 2026; a publication date was not established.
Bounding and transcription are separate specifications
The source's difficult-text conditions raise a practical specification question: what exactly must an annotator identify and transcribe?
A buyer commissioning comparable work would need to define the annotation unit. Is a region a character, word or line? How should a curved sequence be bounded? What reading order applies to vertical text? When a character is partly hidden, should the annotation preserve uncertainty or use a designated unreadable marker?
These are analytical questions prompted by the account, not claims about RWS's undisclosed instructions. A detailed contract would distinguish locating text from deciding what it says. Otherwise two annotations can disagree for different reasons while appearing to represent the same error.
Multilingual scale needs interpretable quality evidence
The reported 32-language scope makes a single combined accuracy figure insufficient for understanding every language slice. For comparable procurement, useful evidence would identify which languages and text conditions were sampled, who reviewed them and how acceptance decisions were made.