The upstream Ovrin project is explicit that no accuracy claim is published unless it is reproduced by the evaluation harness. It also states that real-world corpus work is still necessary before confidence and accuracy claims can be trusted.
Why evaluation matters
To justify a document-extraction product, you need evidence that it works on the kinds of documents you care about. Synthetic corpora are useful, but they do not replace real data.
Current stage
The project has evaluation infrastructure, but it remains careful about claiming a general accuracy number before real corpus data is in place.