Yes, this. I don't know what technique they used. I expect it was known-value variety because they would have better automation in error detection. They may also have used a consensus model, showing work product that really wasn't known value, but was shown to enough different data entry people that the error rate was negated (the error rate requirement was pretty low as I recall) as in a group almost everyone would have done the entry correctly. A third alternative would be to send images that had passed OCR intake for the testing sample.
