Native minus drop-in accuracy, by task

Percentage points on the same examples, with 95% paired bootstrap intervals. Right of zero means native did better.

Hover a row for the number of examples each setup got right and the other got wrong. None of the ten differences is significant after correction.