DevLys is Kruti Dev 010, with two letters swapped
Drawn code by code from the fonts embedded in real government PDFs, 501 of DevLys 010’s 503 glyphs are Kruti Dev 010’s, at the same codes, digits included. One of the two that differ is a code Word’s curly quotes produce, where DevLys draws श and Kruti Dev 010 draws ष. So DevLys is read through the Kruti Dev table with that swap, the same layout a gazette notification typed in Kruti Dev was found to use.
That is also why a converter built for Kruti Dev 010 alone gets DevLys slightly wrong: it reads every श typed on those keys as ष. The words it gets wrong are real words either way, so nothing looks broken; only the spelling is.
The letter the text layer lost
The codes a DevLys PDF stores for श can come from Word’s curly quotes. In two of these notices the embedded font’s own map has no entry for that code, and a reader that took only plain Latin characters from the page dropped it: आदेश read as आदेा and विशेष as विोष, in words that otherwise looked fine. The converter now takes the code from the curly quote the page records. In two notices that brought back विशेष, आवश्यक, पश्चात् and आदेश, whole.
Measured
Fourteen real PDFs typed in DevLys, from state recruitment bodies: result notices, syllabi, instructions to candidates and a scrutiny form. Converted, they hold 17,025 words, and 95.2% of them are in Tesseract’s Hindi or Devanagari word lists, between 89% and 98% document by document. The rest include names and English terms typed in Hindi letters, which are right. One printed line of every document was checked against the conversion by eye, and all matched.
Measuring found the lost श above, and typing habits the converter now reads as the page shows them: अंग्रेजी typed with the ra-foot after the e sign, जिम्मेदारी and बगैर typed with a half letter before the o or ai key, and अंकों typed as aa then e. One notice had been refused outright by a crash in the step that finds page headers; it converts now. Of the 16 sign sequences left that no Hindi word holds, the ones looked at are doubled keys and stray signs in the files, such as a nukta pressed after परी.
Where it still goes wrong
Only PDFs were measured, so this page offers PDFs. A Word file typed in DevLys is read the same way, but ten real ones have not been checked. Of the four curly-quote codes, only one appears in the fonts these notices embed; the other three are read by the same swap on its evidence, and the reading holding more real words is chosen for a document that uses them.
Everything the Kruti Dev page lists holds here too: codes nobody has identified come through as characters that are plainly not Devanagari, a dense form comes out longer than the printed one, and a scan needs OCR.