Open the volume at the page about donations. A sentence tells you that a certain merit follows from listening to "any one of the thirty-six Puranas." It is printed in the ordinary typeface, in the ordinary flow of the ordinary paragraph, and there is a small raised numeral after it.
Follow the numeral to the back of the book, past four thousand lines, and it says this:
"The text actually says twenty-six. But it should be 36, 18 major Puranas (Mahapuranas) and 18 minor Puranas (Upapuranas)."1
So the thing you read was not the thing the manuscript says. Somebody decided the manuscript was wrong, wrote what it ought to have said into the body of the translation, and told you about it once, in a note, at the back, in a numbering scheme that reuses its own numbers.
That is one instance. There are about twenty-three.
A scripted census of all 2,124 endnotes in this volume, matched against ten intervention markers and every hit read in full, gives the following.2
Fourteen located substitutions. Cañcukā → Cañculā, amended to match Volume 1. twenty-six → thirty-six Purāṇas. Puṣpaka → Puṣkara, "as it is customarily stated." mukha → sukha. "The text says twelfth. We have corrected this to thirteenth" — with a proper name reclassified as an adjective in the same note. Tapas / Tapasya → Māgha / Phālguna. Suvaḥ → Svaḥ, "we have changed it to the more common." prasāda → prāsāda. Śambhu → Prabhu — a divine name. Hari → Hara — a second divine name. Viśvayogi → Viśvayoni, because it "sounds more natural." And, in a nyāsa passage, "there are inconsistencies in the way the letters are stated in the text, and we have corrected them" — aksharas repaired inside a mantra, locations not given.3
Five unlocated "liberties." "We have taken minor liberties." "These sentences are difficult to understand, and we have taken liberties." And, at ch 418: "To impart clarity, we have taken liberties with some of the shlokas in this chapter." One chapter, and no indication which shlokas.4
Four structural impositions. The chapter titles are invented wholesale — "the chapter headings we have given are subjective, based on the contents of that particular chapter." The thousand-name enumeration has numbering added and its repeated names silently deleted. The twenty-eight-by-five grid of hells is imposed on what the Sanskrit gives as an undifferentiated list of names. And the hundred and six disciple-numbers in the Pāśupata roll are added "for convenience."5
⚑ Each of these is disclosed. None is marked at the point of use. The reader of the body has no way to know which sentences are which.
The census's second result is the one that turns a list of repairs into a finding.
Sort the twenty-three acts by what the change accomplishes and they do not scatter. Twenty-six becomes thirty-six because thirty-six is what the tradition holds. Puṣpaka becomes Puṣkara because Puṣkara is customary. Cañcukā becomes Cañculā because Volume 1 says Cañculā. Suvaḥ becomes Svaḥ because Svaḥ is commoner. Viśvayogi becomes Viśvayoni because Viśvayoni sounds more natural. A twelfth becomes a thirteenth and an inconsistent epithet becomes an adjective, in one note, because that removes an inconsistency.
Every disclosed change moves the printed text toward the customary, the consensual, the internally consistent, or the natural-sounding.2
⚑ And nobody states this as a policy — not in the Introduction, which sets out a method of reducing interpretation, and not in any of the notes, which each address one local problem as though it were the only one of its kind.6 [n≥3 PATTERN], many times over. What the vault has is a consistent editorial direction that was never adopted as a direction, applied one sentence at a time, over three volumes and some four hundred and sixty chapters.
⛔ This is not an accusation of bad faith and must never be written as one. Every act on the list is disclosed. That is more than most translations of Purāṇic material do. The finding is about what disclosure at the back of a book actually achieves when the body reads perfectly smooth — and the answer is that it satisfies the scholarly obligation without producing the scholarly effect.
Two of the fourteen substitutions replace the name of a deity, and the second of them is not cosmetic.
The body reads: "Hara's fierce and tamasa form is the one who is known as the destroyer." Note 1925, at its own line, reads in full: "In a typo, the text says Hari."7
Hari is Viṣṇu. Hara is Śiva. Whether the destroyer-form belongs to one or the other is not an incidental question in this volume — it is precisely what ch 400 and ch 424 spend their length arguing about, in a Saṃhitā whose whole business is establishing the ranking.8 The Sanskrit, on the translator's own report, assigns the destroyer-form to Viṣṇu at this point. The English assigns it to Śiva. The word typo does the work of the argument.
⚑ Two things are worth separating here, because collapsing them would overstate the case.
The first is that he may well be right. A sectarian Śaiva text calling the destroyer Hari in a chapter about Śiva's forms is a plausible scribal slip, and Debroy has read the manuscript and this build has not. The claim is not that the emendation is wrong.
The second is that the reader cannot participate in the judgement. The body prints a reading; the endnote prints a one-clause justification with no variant apparatus, no manuscript, and no alternative offered; and the sentence in the body is now available to be quoted as evidence in exactly the dispute the emendation settles. ⚑ An emendation that resolves a doctrinal argument is not the same species of act as an emendation that fixes a spelling, and this edition's apparatus uses the same word and the same typeface for both.
The parallel case is note 2104: "The text says Shambhu, which was corrected… the text should say Prabhu."7 Same structure, same one-line warrant, same silence in the body.
One manoeuvre recurs often enough to be named, and naming it is useful because it is the most invisible of the twenty-three.
Note 1367. The text gives an inconsistency — a manvantara identified both by an ordinal and by a name that does not belong to it. Debroy's solution: "The text also refers to this as Rouchya manvantara, an inconsistency. We have therefore taken the word rouchya as an adjective."9
Note 1948. The Pāśupata lineage should yield twenty-eight teachers. Counting the names gives twenty-nine. Debroy's solution: *"If Guhavasi (one who resides in a cave) is treated as a proper name, the number becomes twenty-nine. Therefore, we have taken it as an adjective."* The body accordingly prints "Shikhandi, who lived in a cave."10
A count is reconciled by reclassifying one of its members out of existence. [n=2 SUGGESTIVE] — two instances, and no rigour-language attaches to it at that strength.
⚑ The technique is defensible in the abstract; Sanskrit compounds genuinely do sit on the line between epithet and name, and note 12 says so outright — "There is always a problem distinguishing an adjective from a proper name."5 What makes it worth naming is its direction: in both instances the reclassification is chosen because it makes an arithmetic come out, not on independent grammatical grounds. The count is the evidence and the count is also the thing being fixed.
⚑ And in one of the two cases the emendation contradicts a list the same edition leaves standing: ch 348's roster of future Manus names Rouchya as a proper name, thirty-odd chapters from the note that demotes it.11
The most consequential thing on this page is that the apparatus is not confined to text-critical work. It also teaches, in the same typeface, and on at least one question it teaches something the text never says.
The corpus carries five mutually irreconcilable counts of the tattvas:12
| Count | Where | Whose |
|---|---|---|
| 24 | ch 424-8.2(2), body | the text |
| 26 | ch 315-6(1) and Kailāsa L891, body | the text |
| ~50, enumerated and unnumbered | ch 420-8.1(32), body | the text |
| 31 | note 2086 | the translator |
| 36 | the Introduction — "Shiva's 36 tattvas are different from the tattvas of samkhya" | the translator |
Three from the text, two from the apparatus, none reconciled with any other — and the Introduction's thirty-six, which primes every reader before the first chapter, appears nowhere in the volume's 160 chapters.12
⚑ That is the sharpest available statement of the page's thesis, and it does not require anyone to be wrong. Thirty-six is a real and standard Śaiva scheme. It is simply not this text's, anywhere; and the reader who meets it in the front matter will carry it through a book that never produces it, and will experience the book's own twenty-six and fifty as departures from a norm the book does not hold.
The page would be dishonest if it only ran one way, and there is a clean instance running the other.
Note 1708, on the thirty-eight Mātṛkās: "The number 38 is puzzling. Here, the Matrikas are being linked to letters of the alphabet… Linked thus, the number is usually given as sixteen (16 vowels), five (5 types of consonants) or thirty-three (33 consonants)."13
It is not puzzling. Thirteen chapters earlier, in the same Saṃhitā, ch 368-7(3) partitions Praṇava's five components into named groups — "Eight parts of Sadyojata… Thirteen parts of Vama… Eight parts of the Aghora form… Four parts of Purusha… Five parts originating from Ishana" — and 8 + 13 + 8 + 4 + 5 = 38. Ch 381-7(16) then states the derivation outright: "The thirty-eight Matrikas originated from the matras." A whole-body sweep finds thirty-eight used eight times, across two Saṃhitās, always of the same object.14
⚑ The note measured the text against an external norm and reported bafflement, where the number is internally derived and consistently used. That is the same operation as the twenty-three corrections, with the correction withheld — and it is the one case in this build where the text is right and the apparatus is not.
You are reading a translated book with notes at the back, and you want to know how much of what you just read is the book.
Read the note before you use the sentence. Not after, and not only when something looks odd. You are sitting with a passage that gives a number, and the number is doing work in the argument you are building. Go to the back. Find the note. What it says will be one of four things — this is what the text says, this is what the tradition says, this is what we changed, or we are not sure. Four different acts, one typeface, and the difference decides whether you have evidence or a paraphrase of consensus.
The build that produced this page learned it the hard way. Two clusters after writing the warning, it ran a careful arithmetic check across an enumeration of hells, found a beautiful compositional seam, and had to retract the whole thing when the endnote turned out to say the grid was the translator's own.15
Ask which direction the repairs run. Take the disclosed changes in any edited text — a translation, a restored building, a remastered recording, a cleaned dataset — and sort them not by size but by what each one accomplishes. If they all move the same way, you are not looking at repairs. You are looking at an unstated policy being executed one instance at a time, and the person executing it may not know they hold it.
Distinguish the acts that cannot be undone from the ones that can. A flagged divergence leaves the evidence in place: you can disagree. A fixed divergence replaces the evidence: you cannot, because what you would need to disagree with is now only in a note, in summary, in the corrector's words. The two are indistinguishable from the body, and that is where the loss is.
And notice what the front matter primed you with. You read an introduction before you read the book. Whatever number, framework or emphasis it gave you is now the norm against which everything in the book reads as a departure — and nothing guarantees the book holds it. Here it is thirty-six tattvas, in a text that never once produces thirty-six.
Evidence. Complete scripted census of all 2,124 parsed endnotes in the volume, matched against ten intervention markers — inconsistency, crept in, we have / our translation, the text says, should be / must be, not clear / difficult / impossible, deviated / amended / corrected / liberties, repeat / does not match, probably / presumably / seems, varies from text — with every hit read in full at its own line.2 Census totals: 31 notes containing "we have" or "our translation"; 16 reporting "the text says"; 12 recording an emendation, correction or liberty; 11 flagging an inconsistency; 29 recording difficulty or impossibility; 9 recording divergence from other texts.2 The twenty-three acts individually located and quoted.3457910 The tattva counts swept across the whole body.12 The thirty-eight derivation traced to its source chapter and its eight uses.14
Tension 1 — the method statement and the practice do not obviously conflict, which is the problem. The Introduction's stated method is "reducing interpretation" and "relegating interpretation to footnotes."6 Every one of the twenty-three acts is in a footnote. The method was followed. ⚑ The interpretive load moved to the apparatus exactly as designed, and the body became unmarked as a result. The vault does not resolve whether that is a failure of the method or its intended cost.
Tension 2 — the emendations are probably mostly right. A page that catalogued twenty-three interventions without saying so would be an indictment, and there is no basis for one. What is claimed is the direction, the invisibility at point of use, and the two cases (a divine name, a doctrinal count) where the stakes are not orthographic.
Tension 3 — the corpus is not uniformly compromised, and the vault must not start treating it as though it were. Roughly twenty consecutive enumerations elsewhere in the volume check out exactly, including an eight-tier cumulative siddhi cascade in which every inheritance clause is arithmetically perfect.16 The text can count. The interventions are localised and disclosed, not ambient.
Open question. How many existing vault pages argue from a number, a proper name, a taxonomy, an ordinal or a chapter title in this edition? The sweep has been specified and not run. Until it is, the honest statement is that the exposure is large and unmeasured.
Against Six Disclosed Repairs: that page found the behaviour in Volume 1 and read it as a set of individual acts of scholarly honesty, which is what six instances look like. At twenty-three the reading changes. Six disclosed repairs is a translator being careful; twenty-three with a consistent direction is an editorial policy that nobody has written down. The earlier page is not wrong about any instance — it is under-powered on the pattern, and this page is the pattern.
Against The Five Lost Saṃhitās: the structural echo is exact and worth stating carefully. That page records the text disclosing that it deleted seventy-six per cent of itself and naming the five Saṃhitās that went. This page records the translator disclosing that he added an apparatus and naming the acts. In both cases the disclosure is what makes the original unrecoverable rather than merely unknown — you are told precisely what you cannot get back. Two operations, running in opposite directions on the same object, five hundred years apart. ⚑ Logged as a structural echo, not asserted as identity; nothing here claims the two acts are the same kind of act.
Against The Pāśupata Charter: that page had to be corrected in this build because it reported an announced figure as a delivered one, on the strength of endnote 730's bare multiplication. The correction is this page's thesis demonstrated on the vault itself — a note supplied arithmetic in place of a count, a vault page cited the note, and the error propagated to two further pages before anybody counted the names.
Authority Construction and the Architecture of Belief — that page treats authority as something built to be believed rather than something proven, and its materials are mostly rhetorical: framing, confidence, the management of doubt.
The apparatus supplies a mechanism that page does not carry, and it is quieter than anything in its catalogue: authority conferred by typeface and position rather than by claim.
Nothing in this edition asserts that thirty-six is the text's tattva count. The Introduction says Śiva's thirty-six differ from Sāṃkhya's, which is true and is not a claim about this book at all. But it arrives first, in the front matter, in the same voice that will shortly be narrating, and the reader assimilates it as the frame rather than as a contribution. By the time the body produces twenty-six, the twenty-six reads as the text's variance from a standard — when in fact the standard was never in the room.
⚑ The generalisable point: an editor's most durable influence is not exercised through the sentences a reader disputes but through the ones a reader never notices are sentences. A correction the reader can see is a claim, and claims get evaluated. A correction printed as the object being evaluated is not a claim; it is the ground. And the mark of this species of influence is that it needs no persuasion, no confidence and no rhetoric — only a decision about where on the page a thing goes.
That is why the diagnostic here is positional rather than argumentative. Do not ask is this correct. Ask what would I have had to do to see that this was somebody's decision — and if the answer is "turn to page four hundred and read a numbered note whose number is not unique," then the decision is not functioning as a decision. It is functioning as the text.
One handshake. A second on the disclosure-versus-recoverability structure would duplicate the Five Lost Saṃhitās convergence above, which does the same work at page level.
Sharpest implication. A complete census of 2,124 endnotes shows this edition performing roughly twenty-three disclosed editorial acts: fourteen located substitutions including two divine names and a set of silently repaired aksharas inside a mantra, five unlocated liberties one of which covers an entire chapter, and four structural impositions of which the invented chapter titles are the largest. Every one is disclosed once, at the back, in a numbering scheme that reuses its numbers; none is marked at the point of use; and every one moves the text toward the customary, the consensual, the internally consistent or the natural-sounding, without that ever being stated as a policy. The apparatus is also a second voice on doctrine — two of the corpus's five tattva counts are the translator's, and the one that primes every reader in the Introduction appears nowhere in a hundred and sixty chapters. It is occasionally wrong where the text is right. And the practical consequence for this vault is not hypothetical: roughly three hundred existing pages cite this edition, and any of them arguing from a number, a name, an ordinal or a taxonomy may be arguing from the translator.
Generative questions.