Every finding here still holds when letters are misread. The study reads the standard transcription (Zandbergen and Landini's, called ZL) as the manuscript's letters. ZL is good but not perfect. So we measured how often it misreads a letter, then reran every finding on 50 copies of the text with misreads put in at that rate.
How often ZL misreads a letter
We compared ZL with the ink on words people checked by hand: words approved on the tracing site, and words Ryan checked because they looked doubtful. The doubtful words were chosen for doubt, so their rate is an upper bound.
Each finding, with misread letters
A finding holds when it stands in at least 95 of every 100 misread copies, and is fragile otherwise. The last column repeats the test at twice the doubtful words' rate.
| Page | Finding | Measured rate |
|---|
What this does and doesn't show
Misreads change the size of some numbers. The letter-to-letter predictability, for one, drops by about 0.2 bits. None of them changes what a finding says. A result that turns on the exact spelling of a few words is where a misread still matters. Those words go to the front of the tracing queue, and the study will be read again once every word is traced.