Skip to content

A Note on This Edition

A note on this edition

This is the text of Benoy Kumar Sarkar's translation of the Sukraniti, published in 1914 by the Panini Office, Allahabad, as volume XIII of The Sacred Books of the Hindus. It was made from the scan of the book made for the Digital Library of India and held by the Internet Archive (item in.ernet.dli.2015.242127), in two steps: the pages were read by machine, and every line of the machine's reading was then proofread against the scan of its page.

What is here. The Foreword and the five chapters (Chapter IV in its seven sections), with the translator's footnotes. The title pages, the dedication and the table of contents are left out (the library makes its own), and so are the two indexes, a Sanskrit one in Devanagari and an English one, whose references are to the pages of the printed book. Two printed pages, 34 and 35 of Chapter I (about verses 506–515), are missing from the scan and so from this text. As the translator notes at the head of Chapters I and II, the numbers at the start of each paragraph are those of the lines of Oppert's edition, not of ślokas.

How it was made. Each page of the scan was cleaned and read with the Tesseract engine (version 5, English model); the Foreword was typed from the scan. Then each of the 11,233 lines the machine had read was compared, word by word and mark by mark, with the printed line in the image of the page, and corrected to what is printed: 5,513 lines needed a correction, 4,587 were already right, and 46 printed lines the machine had missed were added. The proofreading was done by AI models (Anthropic's Claude) looking at the page images, under written rules (transcribe as printed, keep the printer's own misprints, write a diacritic only where it can be seen, never guess); a sample of their work was checked again against the scan. Running heads and page numbers were then taken out, the lines of each paragraph joined, and each footnote linked to its mark by its printed number.

The Sanskrit words that the notes quote in Devanagari type (1,244 of them) were read in a separate pass: every line holding one was enlarged from the scan, read by AI models with the Tesseract Sanskrit model's reading as a hint, and the 334 hardest lines (blurred, cut at the edge, or in the tables) were read a second time from larger crops by a stronger model. The words still in doubt after that were compared with the Sanskrit text of the Śukranīti (the verses each note explains, from the text on Sanskrit Wikisource) and with the independent transliteration of these notes on Wisdom Library, and the scan was looked at again where they disagreed; most were settled that way. They are given as printed, with the printer's spellings (कर्म्म, धर्म्म) and misprints; where the English adds a plural s to a Sanskrit word (विषयs), so does this text.

How good the machine's reading was. Measured against the proofread text over the whole book, the machine had 1.7% of the characters of the main text wrong, 3.3% of those of the footnotes, and about one word in ten overall, most of them in Sanskrit names (whose diacritics the coarse scan nearly loses) and in footnotes quoting Devanagari.

What remains to be done.

  • Devanagari. The Devanagari type is small and often blurred in this scan. Four words remain doubtful and are followed by [?]: the reading given is the one that best fits the visible letters, the translator's gloss and the verse. Where the printed note differs from the verse it explains (पर्ण for वर्णि, वध for बन्ध, रवि for रक्ति), the print is kept.
  • Uncertain readings. 585 lines (5%) were marked uncertain by the proofreading: worn letters, faint diacritics, smudged figures in the tables, and footnote marks too small to read (whose numbers were then taken from the order of the notes on the page). Six words are [illegible].
  • Footnote marks inside a line. Some marks printed in the middle of a line were not recorded; the 41 notes concerned are placed at the end of the text of their page instead of at their word.
  • Tables. The tables of measures and weights, with their rows of leader dots, are given row by row; the number of dots is not exact.

When the text matters, check it against the scan.