# Provenance and permanent archive This repository holds the **text**. The work that backs it — the scans of the printed pages, the two OCR readings, the record of every discrepancy and the tools that did the collation — lives outside it, because those are binaries and working files with no place in git. This file says where each piece is and how to check it. A permanent copy of everything together is published on Arweave, where it cannot be modified or withdrawn; the transaction ids go in the table below as they are published. ## The pieces | Piece | What it is | Where it lives today | Size | Arweave | |---|---|---|---|---| | **corpus** | The text of the first editions, 102,729 files, of which 86,418 are published — see *What is deliberately absent* | this repository | 495 MB | see [ANCHORS.md](ANCHORS.md) | | **corrections** | The ledger of corrections applied to the text | this repository, `*.jsonl` | 272 MB | see [ANCHORS.md](ANCHORS.md) | | **scans** | 71 PDFs of the printed books, including the complete Śrīmad-Bhāgavatam | `scan_vedabase/originals/` | 2,078 MB | see [ANCHORS.md](ANCHORS.md) | | **ocr-packed** | Every page as read by both engines, one `.tar` per book | `surya_ocr/`, `.../scan_audit/ocr/` | 193 MB | see [ANCHORS.md](ANCHORS.md) | | **audit** | The ledger of discrepancies: open, arbitrated, applied | `astro_vedabase/scripts/scan_audit/*.json` | 68 MB | see [ANCHORS.md](ANCHORS.md) | | **reports** | Each difference beside the image of the scanned page | `.../scan_audit/*.html` | 126 MB | see [ANCHORS.md](ANCHORS.md) | | **tools** | The code that did the comparison and applied the fixes | both repos, `*.py` | 1.1 MB | see [ANCHORS.md](ANCHORS.md) | | **manifest** | SHA-256 of every file in the package plus a root over the whole set | `MANIFEST.sha256` | 15 MB | see [ANCHORS.md](ANCHORS.md) | Reproduce the package root with `python3 scripts/build_archive.py --manifest-only` and compare it against ANCHORS.md. To assemble the complete package: ``` python3 scripts/build_archive.py --dry-run # see what goes in python3 scripts/build_archive.py --with-tesseract ``` ## A warning to whoever checks this A gateway saying *404* does not mean the file is gone. Verify against more than one, and try again before concluding anything. This was learned the hard way here. A stratified sample of 630 of the published files was pulled from `arweave.net` and fifteen came back as *404 – Page cannot be found*. At two point four per cent, that would have meant some two thousand files lost. Every one of them was then retrieved, byte for byte identical, from `ar-io.dev`. The data was on chain the whole time; what was missing was one gateway's index of it. The same gateway contradicted itself within a minute on one file: a GET returned the nine bytes `Not found` while a HEAD immediately after returned HTTP 200 with `content-length: 795` — the file's exact size. Treat a single negative answer from a single gateway as noise, not as evidence. The reverse error is easier still. A `200` in a redirect chain proves nothing either: `arweave.net` answers a request for an unpropagated file with a 302 to a sandbox subdomain and then a page of HTML. Only comparing the bytes you received against the hash in the manifest settles it. That is what the manifest is for. **Six files were missing, and finding them is why every one was checked.** The fault described below had a third consequence that neither sampling nor reasoning had suggested: in each affected batch one file was never created at all. Not mis-recorded — absent. A verse of the Bhāgavatam in Hindi, three in Portuguese, and two fragments of *The Science of Self-Realization* were listed in the manifest as present and were not on the chain. They were uploaded on 29 Aug 2026 and verified byte for byte. Reading all 86,996 files against the chain took twelve hours and found them. Three separate samples, one of 630 files and two of a few hundred, had all come back clean of this. A sample can only tell you about the proportion it happens to catch, and six files in eighty-seven thousand is not a proportion any sample of that size will reliably catch. If an archive claims to be complete, the claim has to be tested file by file at least once. **And do not trust the upload log's transaction column.** The record kept while uploading, `UPLOAD-STATE.json`, pairs each file with the transaction that carries it, and roughly three pairings in a thousand are crossed: the entry for `sb-4.22.48.md` names the transaction that in fact holds `sb-4.22.49.md`. The uploader had matched the transactions the command-line tool returned to the paths it had been given **by position**, and that order is not guaranteed. Six files were left with no transaction at all by the same fault. It is fixed — the pairing is now made on the source path each transaction reports — but the entries written before 28 August 2026 carry the error. Nothing was lost by it and nothing published is wrong. Every file reached the chain; the drive's own file names were written by the tool, not by that code; and the manifest pairs hash with path and never touched a transaction id. If you want to know which transaction carries a given file, ask the drive or search by the hash — not that log. ## How to check that the text has not changed ``` python3 scripts/hash_manifest.py --check ``` This recomputes the hash of every file and compares it against `MANIFEST.sha256`. If the `root` matches the one anchored on Arweave for that date, the text is the same as it was when published. If it does not match, something changed — and being able to detect that is the whole point. The manifest proves the text **has not been altered**. What proves it **matches the paper** is the collation, and its record is in `audit/` and `reports/`: every discrepancy, which page settled it, and what was decided. ## Published anchors An anchor on a single date says nothing about what came after. What counts is the succession. Every time the corpus changes, the manifest is regenerated and the new `root` is anchored. **The roots and transaction ids live in [ANCHORS.md](ANCHORS.md), not here.** That separation is not tidiness, it is arithmetic. This file is part of the corpus the manifest covers, so if it stated the manifest's own `root`, writing the root down would change the root — the document would be false the instant it was saved. It happened: on 26 Aug 2026 the package root was noted here, and the manifest was stale before the file was closed. So the rule is: every number that depends on the manifest lives in ANCHORS.md, and ANCHORS.md is excluded from both manifests and from the package. It is uploaded on its own, last, once the transaction ids are known. This file carries no figure that changes, which is precisely what lets it be certified along with the text it describes. ## Provenance of the scans A PDF of a 1972 book does not establish on its own where it came from. This is what turns a file into evidence, and it is why the table below records what each scan *says about itself* rather than what its filename claims. **How this was obtained.** Two independent sources, neither of them anyone's memory. The technical columns come from the PDF metadata (`pdfinfo`, `pdfimages`). The copyright column is the text of the book's own copyright page, as read by the Surya OCR engine and quoted verbatim, misprints and all. Anyone can repeat both readings against the published scans. **Checked against the second engine.** The same method used on the text itself was used here: every copyright page was read again by Tesseract and the two readings compared on printing number and year. Of the sixty-nine scans, sixty-four have both readings and **all sixty-four agree**; the remaining five have no Tesseract pass. Three volumes appeared to differ and did not — Tesseract also caught the phrase *"the first edition widely available to the English-reading public"* from a jacket blurb in Bhāgavatam 10.1-10.3, which is publicity, not a printing statement. Two of the five without a second reading matter, and were confirmed differently. The Bhagavad-gītā copyright page was read in full rather than in fragments, and says plainly that the Collier paperback and the Macmillan hardcover are separate editions. The Śrī Īśopaniṣad page carries all three lines together: Copyright © 1969 by ISKCON Books First edition published by ISKCON Books in 1969 Fifth Printing 1972 **Read the names with suspicion.** Several files are named for an edition their copyright page does not support, and the discrepancy runs in one direction — the name claims an earlier, more authoritative printing than the book actually is. Each entry below was confirmed by reading the copyright page image itself, not only its OCR: - `Bhagavad-gita_As_It_Is-Original_1972_Macmillan_edition-SCAN` is not the 1972 Macmillan first printing. Its copyright page reads *First Collier Books Edition 1972 · Third Printing 1973*, and adds that the book *"is also published in a hardcover edition by the Macmillan Company"* — naming the Macmillan as a separate edition, not as this one. What was scanned is the Collier paperback, third printing. That matters more here than anywhere else: the 1972 Macmillan is the edition this project treats as its source. - `Beyond_Birth_and_Death-Original_1974_edition_scan` is the **third** printing. The page lists 1972, 1973 and 1974; 1974 is the last of the three. - `Sri-Isopanisad-scans-of-original-1969-edition` carries the 1969 first edition notice, but the copy scanned is the **fifth printing, 1972**. Both engines read the three lines identically on page 7. - `The_Nectar_of_Instruction-Original_1976_SCAN` is the one near-miss that turns out to be right. Its copyright is 1975 and its first printing was 1975, but the same page lists *Second printing, 1976: 100,000 copies* — and that is the copy scanned. The year in the name is correct. - `sri-caitanya-caritamrta` Ādi-līlā volumes 1 and 2 read *Second printing, 1983*, which the README already records. None of this makes the scans less useful. It makes them *honest*, which is the point: a collation is only as good as the knowledge of which printing it was made against. **The two doubtful readings, resolved.** Neither was an OCR failure. `KRSNA-the-Reservoir-of-Pleasure-1970` carries a full printing history, and the earlier extract had simply truncated it after the first line: Copyright ©1970 by ISKCON PRESS Printing History First Printing 1966 - 1,000 Second Printing 1967 - 5,000 Third Printing 1968 - 5,000 Fourth Printing 1969 - 20,000 Fifth Printing 1970 - 30,000 The 1966 is real — the text circulated as an essay from then. What was scanned is the **fifth printing, 1970**, which is what the copyright year records. `1972_Perfection_of_Yoga_2008_Original` lists two printings and no years because **the printed page itself omits them**. Both OCR engines read the page identically. Copyright 1972; which of the two printings this copy is cannot be told from that page. **Three more scans, read directly.** - `Easy-Journey-to-Other-Planets-1972-krishnaorg-SCAN` is the *First Collier Paperbacks Edition 1972* — the same Collier/Macmillan arrangement as the Bhagavad-gītā. - `Easy_Journey_to_Other_Planets-Original_India_Edition_SCAN` is what its name says, and is the one case where the name is exactly right: published by Prabhupāda himself through The League of Devotees, 1/859 Keshi Ghat, Vrindaban, printed by Surendra Printer's of Delhi, priced at one rupee, with *Copy Right: Author* and no year. It is a different book from the Collier edition, not another printing of it. - `Life_Comes_from_Life-1979_first_edition-SCAN` **is not a scan**. Its PDF was produced by Microsoft Word — 68 pages, 535 KB, not one image in it. It is a retyping, and cannot serve as evidence of any printed page. That is also why its OCR shows no copyright page: there is none to read. **A gap found while checking, and closed.** Two books had their OCR published in the containers while their PDFs were absent from the archive; both sat outside the `originals/` directory the package is built from, so the build never saw them. `Teachings_of_Lord_Chaitanya-1968_first_edition-SCAN` is a genuine scan — 351 pages at 200 ppi — and has since been added, which is why the archive holds seventy-one scans rather than seventy. `Life_Comes_from_Life` was left out on purpose: as noted above it is a Word document, not a scan, and putting it among the page images would misrepresent what it is. **A second version exists for eight books, and the published one is the right one.** `improved/` holds reprocessed copies of eight of the published scans, differing in bytes and in file size. Identical pagination is not by itself proof that the collation refers to the published file rather than to the other, so it was tested: for each of them, the OCR text was compared against the text layer of both PDFs, page by page. In every case the published `originals/` file matches the OCR at least as closely as the `improved/` one, and in no case does `improved/` match better. The two share a text layer and differ only in image processing. **Where these files come from.** Not from scanning done for this project. The evidence is in the files themselves and it is consistent. The PDFs carry their own creation dates, and they fall into a campaign rather than a trickle. Twenty-seven were made between 13 and 16 September 2012 — the Bhāgavatam and Caitanya-caritāmṛta volumes, several a day. Twenty more followed between 27 September and 9 October, two in December, eight through 2013 and five through 2014 at roughly weekly intervals, and two stragglers in 2016 and 2020. Seven carry no internal date at all; the seven groups account for all seventy-one files. Seventeen name `krishnapath.org` in their metadata. Ten name the machine that produced them: eight a CanoScan LiDE 210, one a Canon SC1801, one a Canon MG5300. That matches what the source says of itself. KrishnaPath.org describes a volunteer effort to scan and **digitally remaster** the authorised editions — removing background marks and sharpening the text so it can be searched and copied. Fifty-three of the seventy-one bear the fingerprint of that work: Adobe Acrobat's Paper Capture *ClearScan*, which replaces the scanned glyphs with synthesised fonts. **These are therefore remastered images, not raw captures.** The page images survive at the resolutions listed, and the collation against them stands, but anyone treating a letterform in these files as photographic evidence of the printed page should know what was done to it first. For the Bhāgavatam the trail is recorded elsewhere by the same hands that assembled this archive: *"The Śrīmad-Bhāgavatam scans (original 1972 edition, © BBT) come from prabhupadabooks.com / krishna.org; the rest from devotee archives."* **What remains unknown.** Which physical copy each volunteer worked from, and whose it was. That is not recorded in any file and cannot be recovered from them; it would have to come from the people who did the scanning. Given that KrishnaPath itself warns that altered copies of these free scans circulate, the practical check is not provenance by testimony but byte-identity against a known public copy — which is what was done for the seventeen Caitanya-caritāmṛta volumes, and what the hashes in this archive make possible for the rest. | Scan | Book | What its copyright page says | ppi | Source | |---|---|---|---:|---| | `Beyond_Birth_and_Death-Original_1974_edition_scan` | beyond-birth-and-death | © 1972 by the Bhaktivedanta Book Trust (B.B.T.) · First Printing, 1972: 25,000 copies Second Printing, 1973: 100,000 copies Third Printing, 1974: 200,000 copies | 600 | krishnapath.org | | `Bhagavad-gita_As_It_Is-Original_1972_Macmillan_edition-SCAN` | bhagavad-gita-as-it-is | Copyright © 1972 by His Divine Grace A.C. Bhaktivedanta Swami Prabhupāda · First Collier Books Edition 1972 Third Printing 1973 | 600 | krishnapath.org | | `Easy-Journey-to-Other-Planets-1972-krishnaorg-SCAN` | easy-journey-to-other-planets | Copyright © 1970, 1972 by ISKCON Press | 300 | CanoScan LiDE 210 | | `Easy_Journey_to_Other_Planets-Original_India_Edition_SCAN` | easy-journey-to-other-planets | *(not found in the first 30 pages)* | 600 | krishnapath.org | | `1973_Elevation_to_Krsna_Consciousness` | elevation-to-krsna-consciousness | © 1973 the Bhaktivedanta Book Trust. · First Printing, 1973: | 300 | krishnapath.org | | `Sri-Isopanisad-scans-of-original-1969-edition` | isopanisad | Copyright © 1969 by ISKCON Books · First edition published by ISKCON Books in 1969 | 600 | krishnapath.org | | `KRSNA_Book_Vol.1_1970_ISKCON_Press_edition_SCAN` | krsna-the-supreme-personality-of-godhead | Copyright © 1970 A. C. Bhaktivedanta Swami | 598 | krishnapath.org | | `KRSNA_Book_Vol.2_1970_ISKCON_Press_edition_SCAN` | krsna-the-supreme-personality-of-godhead | *(not found in the first 30 pages)* | 300 | krishnapath.org | | `Life_Comes_from_Life-1979_first_edition-SCAN` | life-comes-from-life | **NOT A SCAN and not in this archive** — a Word export with no page images; its OCR is published, the file is not | — | — | | `Light_of_the_Bhagavata-1984_first_edition_SCAN` | light-of-the-bhagavata | First Printing, 1984: 40,000 copies. | 300 | CanoScan LiDE 210 | | `The_Nectar_of_Devotion-1970_ISKCON_Press_edition-Hardcover-SCAN` | nectar-of-devotion | Copyright © 1970 by ISKCON PRESS | 300 | krishnapath.org | | `The_Nectar_of_Instruction-Original_1976_SCAN` | nectar-of-instruction | ©1975 Bhaktivedanta Book Trust. · First printing, 1975: 10,000 copies | 600 | krishnapath.org | | `On_the_Way_to_Krsna-1973_book_SCAN` | on-the-way-to-krsna | © 1973 by the Bhaktivedanta Book Trust. All rights reserved · First Printing, 1973: 100,000 Printed in the United States of America by ISKCON Press | 600 | krishnapath.org | | `The-Path-of-Perfection-SCAN` | path-of-perfection | First Printing, 1979: 175,000 copies · © 1979 Bhaktivedanta Book Trust All Rights Reserved Printed in the United States of America | 300 | CanoScan LiDE 210 | | `Perfect_Questions_Perfect_Answers-Original_1977_Edition-SCAN` | perfect-questions-perfect-answers | © 1977 Bhaktivedanta Book Trust · First printing, 1977: 220,000 copies | 300 | krishnapath.org | | `1972_Perfection_of_Yoga_2008_Original` | perfection-of-yoga | © 1972 by ISKCON PRESS. All rights reserved · First Printing: 30,000 copies Second Printing: 100,000 copies Printed in the United States of America | 600 | krishnapath.org | | `1973_Raja-Vidya_The_King_of_Knowledge` | raja-vidya | © 1973 by the Bhaktivedanta Book Trust (B.B.T.). · First Printing, 1973: 100,000 copies | 600 | krishnapath.org | | `KRSNA-the-Reservoir-of-Pleasure-1970` | reservoir-of-pleasure | Copyright ©1970 by ISKCON PRESS · First Printing 1966 - 1,000 | 150 | krishnapath.org | | `Science-of-Self-Realization-1977` | science-of-self-realization | First Printing, 1977: 50,000 copies | 300 | CanoScan LiDE 210 | | `adi1` | sri-caitanya-caritamrta | © 1974 the Bhaktivedanta Book Trust · First printing, 1974 Second printing, 1983 | 300 | Adobe Acrobat 10.1.4 | | `adi2` | sri-caitanya-caritamrta | © 1973 the Bhaktivedanta Book Trust · First printing, 1973 Second printing, 1983 | 300 | Adobe Acrobat 10.1.4 | | `adi3` | sri-caitanya-caritamrta | © 1974 the Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.0 | | `ant1` | sri-caitanya-caritamrta | ©1975 Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.1.4 | | `ant2` | sri-caitanya-caritamrta | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `ant3` | sri-caitanya-caritamrta | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `ant4` | sri-caitanya-caritamrta | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `ant5` | sri-caitanya-caritamrta | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `mad1` | sri-caitanya-caritamrta | ©1975 Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.1.4 | | `mad2` | sri-caitanya-caritamrta | ©1975 Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.1.4 | | `mad3` | sri-caitanya-caritamrta | ©1975 Bhaktivedanta Book Trust. | 300 | Adobe Acrobat 10.1.4 | | `mad4` | sri-caitanya-caritamrta | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat Pro 10.1.4 | | `mad5` | sri-caitanya-caritamrta | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `mad6` | sri-caitanya-caritamrta | © 1975 Bhaktivedanta Book Trust. · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `mad7` | sri-caitanya-caritamrta | © 1975 Bhaktivedanta Book Trust. · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `mad8` | sri-caitanya-caritamrta | ©1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `mad9` | sri-caitanya-caritamrta | ©1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 300 | Adobe Acrobat 10.1.4 | | `SB1.1` | srimad-bhagavatam | Copyright © 1972 by the Bhaktivedanta Book Trust | | Adobe Acrobat 10.0 | | `SB1.2` | srimad-bhagavatam | © 1972 the Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.0 | | `SB1.3` | srimad-bhagavatam | Copyright © 1972 by the Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.0 | | `SB10.1` | srimad-bhagavatam | First Printing, 1977: 20,000 copies · © 1977 Bhaktivedanta Book Trust All Rights Reserved Printed in the United States of America | 600 | CanoScan LiDE 210 | | `SB10.2` | srimad-bhagavatam | First Printing, 1977: 50,000 copies · © 1977 Bhaktivedanta Book Trust All Rights Reserved Printed in the United States of America | 600 | CanoScan LiDE 210 | | `SB10.3` | srimad-bhagavatam | First Printing, 1980: 20,000 copies · © 1980 Bhaktivedanta Book Trust All Rights Reserved Printed in the United States of America | 600 | CanoScan LiDE 210 | | `SB2.1` | srimad-bhagavatam | *(not found in the first 30 pages)* | 300 | Adobe Acrobat 10.0 | | `SB2.2` | srimad-bhagavatam | © 1972 the Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.1.4 | | `SB3.1` | srimad-bhagavatam | © 1972 the Bhaktivedanta Book Trust | 150 | Adobe Acrobat 10.0 | | `SB3.2` | srimad-bhagavatam | © 1974 the Bhaktivedanta Book Trust | 150 | Adobe Acrobat 10.0 | | `SB3.3` | srimad-bhagavatam | © 1974 the Bhaktivedanta Book Trust | 150 | Adobe Acrobat 10.0 | | `SB3.4` | srimad-bhagavatam | © 1974 the Bhaktivedanta Book Trust | 150 | Adobe Acrobat 10.0 | | `SB4.1` | srimad-bhagavatam | Copyright © 1972 by the Bhaktivedanta Book Trust | | — | | `SB4.2` | srimad-bhagavatam | © 1974 the Bhaktivedanta Book Trust | 150 | Adobe Acrobat 10.0 | | `SB4.3` | srimad-bhagavatam | © 1974 the Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.0 | | `SB4.4` | srimad-bhagavatam | © 1974 the Bhaktivedanta Book Trust | 300 | Adobe Acrobat 10.0 | | `SB5.1` | srimad-bhagavatam | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 299 | Adobe Acrobat 10.0 | | `SB5.2` | srimad-bhagavatam | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 299 | Adobe Acrobat 10.0 | | `SB6.1` | srimad-bhagavatam | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 299 | Adobe Acrobat 10.0 | | `SB6.2` | srimad-bhagavatam | © 1975 Bhaktivedanta Book Trust · First printing, 1975: 20,000 copies | 299 | Adobe Acrobat 10.0 | | `SB6.3` | srimad-bhagavatam | © 1976 Bhaktivedanta Book Trust · First printing, 1976: 20,000 copies | 300 | Adobe Acrobat 10.0 | | `SB7.1` | srimad-bhagavatam | ©1976 Bhaktivedanta Book Trust · First printing, 1976: 50,000 copies | 300 | Adobe Acrobat 10.0 | | `SB7.2` | srimad-bhagavatam | ©1976 Bhaktivedanta Book Trust · First printing, 1976: 20,000 copies | 300 | Adobe Acrobat 10.0 | | `SB7.3` | srimad-bhagavatam | ©1976 Bhaktivedanta Book Trust · First printing, 1976: 20,000 copies | 300 | Adobe Acrobat 10.0 | | `SB8.1` | srimad-bhagavatam | © 1976 Bhaktivedanta Book Trust · First printing, 1976: 20,000 copies | | — | | `SB8.2` | srimad-bhagavatam | ©1976 Bhaktivedanta Book Trust · First printing, 1976: 20,000 copies | 300 | Adobe Acrobat 10.0 | | `SB8.3` | srimad-bhagavatam | © 1976 Bhaktivedanta Book Trust · First printing, 1976: 20,000 copies | 299 | Adobe Acrobat 10.0 | | `SB9.1` | srimad-bhagavatam | First Printing, 1977: 20,000 copies · © 1977 Bhaktivedanta Book Trust All Rights Reserved Printed in the United States of America | 299 | Adobe Acrobat 10.0 | | `SB9.2` | srimad-bhagavatam | First Printing, 1977: 20,000 copies · © 1977 Bhaktivedanta Book Trust All Rights Reserved Printed in the United States of America | 300 | Adobe Acrobat 10.0 | | `SB9.3` | srimad-bhagavatam | First Printing, 1977: 20,000 copies · © 1977 Bhaktivedanta Book Trust All Rights Reserved Printed in the United States of America | 299 | Adobe Acrobat 10.0 | | `Teachings_of_Lord_Chaitanya-1968_first_edition-SCAN` | teachings-of-lord-caitanya | Copyright © 1968 International Society for Krishna Consciousness Inc. (ISKCON) · Library of Congress Catalogue Card Number: 68-29320 ALL RIGHTS RESERVED FIRST EDITION | | — | | `Teachings_of_Queen_Kunti-SCAN` | teachings-of-queen-kunti | First Printing, 1978: 57,300 copies · © 1978 Bhaktivedanta Book Trust All Rights Reserved Printed in the United States of America | 600 | krishnapath.org | | `KRSNA Consciousness - The Topmost Yoga System - Original 1970 edition Scan` | topmost-yoga-system | Copyright © 1970 ISKCON PRESS | 600 | PDFium | | `Teachings_of_Lord_Kapila-SCAN` | *(not in the collated corpus)* | *(no OCR pass — see note below the table)* | 592 | krishnapath.org | | `Transcendental_Teachings_of_Caitanya_Mahaprabhu` | *(not in the collated corpus)* | *(no OCR pass — see note below the table)* | 300 | krishnapath.org | | `Transcendental_Teachings_of_Prahlad_Maharaj-Original_book-SCAN` | *(not in the collated corpus)* | *(no OCR pass — see note below the table)* | 600 | Canon MG5300 series | Three of the seventy-one have no OCR column because they were never collated: they are books outside the corpus this project reproduces. They are archived as scans all the same — a printed page is worth keeping whether or not we set text against it. ## What is deliberately absent Count the files in the archive against the table above and 16,311 are missing. They are not lost, and their absence is not an accident. **The Russian translation is not published.** It is unfinished, and Arweave has no way back: a half-translated chapter put up today stays up, and standing beside a corpus collated page by page against the printed editions it would borrow a credibility it has not earned. The files are held in the repository until the translation is complete, and will be published then as a later addition. The English, Spanish, Portuguese and Hindi translations are all here. One Russian thing is here: the twenty `.jsonl` correction ledgers under `corrections/translations/russian/`. They went up before the text was held back, and on Arweave that cannot be undone. They record what was corrected, not what the translation says, so they are listed rather than hidden — but they are not the text, and the text is not here. This is worth stating plainly rather than leaving to inference, because an archive of this kind is read by people who cannot ask us anything. A silent gap looks like data loss, or like something hidden. It is neither. ## A note on character encoding Every text file here is UTF-8, and the corpus is full of Sanskrit diacritics (ā, ī, ū, ṛ, ṣ, ṭ, ñ, ś) as well as Spanish, Portuguese and Russian text in the translations. When uploading, the content type must carry the charset — `text/markdown; charset=utf-8`, not bare `text/markdown` — or browsers fall back to Latin-1 and the diacritics come out as mojibake. This cannot be corrected after the fact: a file already on Arweave can only be replaced by uploading it again and paying again. Because the content type applies to a whole upload, each section is uploaded separately with its own: | Section | `--content-type` | |---|---| | corpus, audit/notes | `text/markdown; charset=utf-8` | | corrections, audit/candidates | `application/x-ndjson; charset=utf-8` | | audit/ledger, audit/text-layer | `application/json; charset=utf-8` | | ocr-packed, `*.tar` | `application/x-tar` | | ocr-packed, `OCR-CONTENTS.sha256` | `text/plain; charset=utf-8` | | reports | `text/html; charset=utf-8` | | tools | `text/x-python; charset=utf-8` | | scans | `application/pdf` | This is why `corpus` and `corrections` are separate sections even though both come from this repository: they are different kinds of thing and they cannot share a content type. Note the syntax: `text/markdown;charset=utf-8`, with **no space** after the semicolon. With a space the shell splits the argument and the CLI keeps only the first half, so the charset never reaches the transaction and the file is served without it. Verified on 26 Aug 2026 — the first attempt, written with a space, recorded a bare `text/markdown` and rendered the Sanskrit as mojibake. ## A note on the OCR containers Everything in this archive is uploaded one file at a time, addressable on its own from any gateway — except the OCR, which travels as 42 `.tar` containers, one per book per engine. The reason is arithmetic. The OCR output is 69,799 text files, one per page per engine, with a median size of 1,874 bytes. Measured against the Turbo price API on 26 Aug 2026, an upload costs **9,174,313 winc per file plus 11,184.90 winc per byte** — a formula that reproduces the API's own quotes to 0.000% across sizes from 500 bytes to 1 MB. The per-file part is small in money, but the CLI spends 0.58 s on each file, so those pages alone were thirteen hours of upload. Packing them costs slightly *more* in credits, not less: tar pads every member to a 512-byte boundary, so 129 MB of pages become 193 MB of containers, and the 0.64 credits saved on per-file overhead are given back in padding and then some. What it buys is time — thirteen hours become seconds. Nothing else in the archive is packed, because being readable one file at a time is most of what the corpus, the scans and the reports are for. `tar` was chosen over `zip` or `gzip` deliberately. It is uncompressed and its format is specified in POSIX, so a flipped bit damages one member and the rest still extract; compression would couple every byte to every other and bet on a decompressor still existing in fifty years. The variant is `ustar`, the most conservative one that fits these paths — the longest is 112 characters, split across the 155-byte prefix and the 100-byte name field. The containers are deterministic: members sorted by path, uid and gid zeroed, owner names emptied. The same input produces the same bytes on any machine. **A container is only worth using if you need not trust it.** So `OCR-CONTENTS.sha256` travels beside the tars and lists the SHA-256 of each container followed by the SHA-256 of all 69,799 members, as `container!path`. To check one page without extracting the rest: P=isopanisad/Sri-Isopanisad-scans-of-original-1969-edition/p0001.surya.txt tar -xOf ocr-surya-isopanisad.tar "$P" | shasum -a 256 grep "ocr-surya-isopanisad.tar!$P" OCR-CONTENTS.sha256 Inside a container the path keeps the edition directory the scan came from, so the page carries the printing it was read from, not just a page number. ### One duplicate, removed The Surya run left the Bhāgavatam laid out twice: as loose volume directories `SB1.1 … SB10.3` at the top of `surya_ocr/`, and again inside `srimad-bhagavatam/`, which holds those same thirty. Verified on 26 Aug 2026: the 11,474 relative paths match exactly and all 11,474 SHA-256 match. Only the merged copy is archived — it is how every other book in the section is named. The loose copies were dropped before packing, which is why `ocr-surya` holds 24,035 files here and 35,509 in the working directory it came from. ## A note on the markup Ślokas in Devanāgarī and their IAST transliteration are `>` blockquotes, and each line of the verse ends with a **backslash**. That backslash is Markdown's hard line break: without it the pādas would be joined into one running line. It is markup, not text. Anyone extracting plain text from these files should strip a trailing `\` from each line — it is not part of the verse. The backslash was chosen over Markdown's other hard-break form, two trailing spaces, because trailing whitespace is invisible and almost every editor strips it on save. A verse break is information, not formatting, so it is written in a way that is visible and hard to destroy by accident.