Measured accuracy
On our ground-truth corpus the pipeline read 100% of the characters of its one Punjabi test page correctly, 1,206 characters of prose. One clean digital page is far too small a sample to call that a rate: it says the common letters, tippi, bindi and addak are read well when the print is clean, and nothing about a worn or photographed page.
Letters on their own
Gurmukhi recognition leans on the word around each letter. We drew 41 Gurmukhi letters one at a time in a clean Unicode font and read them back: 26 came back right. Of 15 consonants with a vowel sign, 10 did, and every subjoined letter under a consonant was misread. Real words in the same font came back far better, 34 of 40. So a list of letters, a table of single characters or a word broken across a line reads much worse than running text.
Lines read in the wrong alphabet
A page is read with the Punjabi and English models together, so an English phrase in it comes back in English. On our test page that sometimes went the other way: whole Gurmukhi lines came back as Latin letters, such as HH act Uet. Latin words on a line that is not mostly English words are now read again with the Punjabi model alone, which repairs those lines, while English lines are left exactly as they were read. Handwriting is not supported; this reads printed Punjabi.