Yahya - 

I think it probably best if the data entry people did not know they are entering a duplicate.  I can't think why you would need to know which was entered in what order.  If you are entering the duplicates at a later time, however, then everyone already knows which are the duplicates. 

Is 10% a large number?  It seems it is the number of duplicates which matters.  It seems that you want enough to know what percent of each kind of error occurs within that number of entries that you will duplicate.   

Not all errors are equal.  Some matter more than others.  Don't you want to know the precent of errors committed in the test-duplicate group for each kind of error, especially the most egregious ones? 

To know the number of duplicates you want, you might think about estimating proportions (proportion of errors to no errors), so that say 0.97 proportion of errors and 0.03 no errors gives you a 3% error rate. 

See https://newonlinecourses.science.psu.edu/stat414/node/212/

and

 https://newonlinecourses.science.psu.edu/stat414/node/264/

I assume your population could basically be considered infinite.  But I don't know your whole situation. 

Now if you find a percent of errors, either overall, or for a category, I suppose that would mean a percent of disagreements.  A disagreement might mean that the original was wrong, or the duplicate was wrong, or both were wrong.  If both were wrong the same way, that could look like no error.  But with multiple digit entries, that seems unlikely.  A disagreement where both are wrong may be quite possible.  It seems that perhaps the number of errors that would have been committee on a single entry might be slightly larger than half of the disagreements, but that probably needs more thought.  I can see now why you would want to know which are the duplicates. In a disagreement, you could then check the original completely for correctness, but a code that makes it obvious to the data entry person which is the duplicate may possibly still not be the best idea. 

This just looks at the problem.  Double entering everything might basically solve it.  But this information may tell you if that is needed, or will do enough to help.  So I see its merits. 

Enjoy your project.

Cheers - Jim

PS -

Oops. ."Already entered data." How did I miss that? Well, maybe the people doing the duplicate data entry don't need to know that the data are duplicates! Maybe a character in a previously blank column could be added, and that might not make the record that different as a task, and something you can compare. If not otherwise the same as the original, you check to see if the original is correct. You will only miss cases where both have the same error(s). Unlikely.

Then the proportion of errors in the original file is (almost) exactly what you want. Well, I mean the proportion of errors in the original file, but counting only the ones that had a duplicate. That would be 10% of the original file if you do not change that.

Thank you James R Knaub for your input, that gave me a better insight into the matter.