Most translations are silent about their own surgery. You get the smooth English and no record of where the original resisted.
Debroy's Śiva Purāṇa is unusually noisy about it, and the noise is worth collecting, because every seam argument this vault makes about the text rests on being able to tell what the Sanskrit did from what the translator did.
Across Volume 1 he discloses six interventions. Not six errors he flags in the text — six places where he changed or supplemented what was in front of him and told you so.
| Endnote | Ch | What he did |
|---|---|---|
| 787 | 73 | "the text repeatedly says nadi (river). We have corrected it to Nandi."1 |
| 788 | 73 | "Since this shloka is terse, we have expanded it a bit, to make the meaning clear."1 |
| 859 | 85 | *"This sentence is unnecessary repetition."*2 |
| 877 | 89 | "The text states Vaishnava. This is a typo. A few sentences later, it becomes clear that this is meant to be Ashman."3 |
| 891 | 90 | "These sentences are convoluted, and some liberties have been taken."4 |
| 904 | 94 | "The text reads Pushan but should probably read Pushti."5 |
Six admissions, spread across twenty-one chapters, none referring to any of the others.
The first is the most consequential and it should be looked at closely.
The text repeatedly says nadi (river). We have corrected it to Nandi.
Repeatedly. This is not one slip. The manuscript tradition Debroy is working from has, in a run of sentences, a word that he judges to be wrong — and he silently substitutes the right one throughout, then tells you in a note at the back.
Now consider what that does to a lens.
The vault's vocative lens works by finding addresses that do not fit their scene — "O great sage!" in a room with no sage — and treating them as fragments that arrived from elsewhere.6 The lens is only usable because Debroy generally flags anomalies rather than smoothing them.
Endnote 787 is a case where he smoothed. If he had done that silently, the anomaly would have vanished and the vault would have read the repaired passage as clean text.
So the method's reliability is a property of the translator's disclosure habits, not of the manuscript.
The second disclosure sits in the same chapter and is a different species.
Since this shloka is terse, we have expanded it a bit, to make the meaning clear.
Nothing was wrong here. The Sanskrit was compressed, and he added words so the English would read.
That is ordinary translation practice and it is honestly declared. It is also the disclosure most dangerous to this build's methods, and for a reason that has nothing to do with error.
A repeated-block argument compares two passages and asks whether they are near-verbatim. The vault's strongest such finding is the twelve-item question appearing in chapters 33 and 37 in different mouths.7 That comparison is made in English.
If one instance of a block sits in a terse passage that was expanded and the other does not, the two will read as less similar than the Sanskrit is. Expansion is invisible in the output and it degrades exactly the measurement the lens depends on.
Endnote 788 does not undermine any specific finding. It establishes that the degree of verbatim similarity is not directly observable from this edition, and that the honest form of a duplication claim is: near-identical in Debroy's English, at a spot he has not flagged as expanded.
Two of the six are straightforward emendations and they are worth separating from the others because their epistemics differ.
Endnote 877 — "The text states Vaishnava. This is a typo." And note the justification he gives: "A few sentences later, it becomes clear that this is meant to be Ashman." He is emending from internal consistency — the later text disambiguates the earlier.
Endnote 904 — "The text reads Pushan but should probably read Pushti." Note the hedge: should probably. He is less certain here, and he says so.
The pair shows a translator calibrating. When the correction is forced by context he states it flatly; when it is inference he marks it as inference.
That is exactly the behaviour that makes the rest of the apparatus usable, and it is why the vault can treat a flagged inconsistency as evidence rather than as a translator's opinion. He does not flatten his own confidence levels.
Here is the distinction that governs every seam claim this build makes.
Flagged. He leaves the anomaly in the English and writes a note: this is an inconsistency. The reader can see the damage. Five frame violations are of this kind.6
Fixed. He removes the anomaly and writes a note: we have corrected it. The reader sees clean text and learns about the damage only from the back matter.
The two produce opposite artefacts and the vault's lenses only work on the first.
Which means every argument of the form this passage shows a seam has to answer a prior question: at this spot, did Debroy flag or fix? If he fixed, the surface has been repaired and the absence of a visible anomaly proves nothing.
And there is a worse version of the question that the six disclosures cannot answer. These are the repairs he told us about. A translator who discloses six is more scrupulous than most; he is not thereby exhaustive. The nadi/Nandi correction is disclosed because he noticed it and thought it worth a note — and the class of corrections a translator makes without noticing they are corrections is, by construction, undisclosed.
So the six are a floor on the intervention count, not a ceiling, and the honest form of every finding in this build is: visible in Debroy's English, at a spot with no disclosed repair.
That is a weaker claim than it looked like before this page. It is also the only one the evidence supports.
⚑ ENRICHED 2026-08-11 from Volume 3, where the whole endnote layer has now been censused — and the count moves from six to roughly twenty-six.
A complete scripted pass over all 2,124 of Volume 3's notes, matched against ten intervention markers with every hit read in full, gives around seventeen divergences fixed against around nine flagged and left standing.V3a Flagged-versus-fixed is not an occasional asymmetry in this edition. It is its normal operating mode, and this page's finding — assembled from a single pair thirty-three notes apart — is confirmed at scale.
And Volume 3 supplies a far cleaner pair than the original, because both halves answer the same question.
Identical problem. Identical pair of diagnoses. Opposite treatment. One thousand one hundred and eighty-four notes apart, and neither references the other.V3b
⚑ A third instance sits on one page: note 1242 fixes (Puṣpaka → Puṣkara, silently in the body) while note 1250 flags and leaves alone — "This is usually stated as Chaitraratha in Mandara… and Savitra in Suparshva" — and the body keeps the text's own reading and never gets Savitra at all.V3c
⛔ With Volume 1's notes 929 and 962, the asymmetry now has three pairs across two volumes and a census behind it. [n≥3 PATTERN] — and the operative consequence is unchanged and sharpened: the reader of the body cannot tell, at any point, which behaviour was applied.
⚑ One thing the census adds that the six could not. Five of Volume 3's disclosures are unlocated liberties — "we have taken minor liberties", "these sentences are difficult to understand, and we have taken liberties", and, on ch 418, "To impart clarity, we have taken liberties with some of the shlokas in this chapter."V3d A whole chapter, disclosed, with no indication of which shlokas. That is a third category this page did not have: not flagged, not fixed-at-a-named-spot, but fixed somewhere inside a stated range — which leaves the reader knowing that a repair happened and unable to locate it. Of the three, it is the only one that damages a whole chapter's evidential value at once.
You are working from a cleaned dataset, a translated document, a transcript, or somebody's tidied notes, and you are about to make an argument from an anomaly in it.
Three questions, in order, and the third is the one nobody asks.
Did the cleaner tell you what they cleaned? If there is no changelog, anomaly-based arguments are not available to you at all — you cannot distinguish a feature of the source from a survival of the cleaning. Debroy's six notes are what make this corpus usable for this kind of work; most sources have none.
At this specific spot, did they flag or fix? These are opposite operations with opposite residues. A flag preserves evidence; a fix destroys it and leaves a receipt. You need to check per-location, not per-document, because the same person does both — Debroy flags in chapter 77 and fixes in chapter 73.
And what class of change would they not have thought worth logging? This is the real limit. Nobody logs the changes that felt like translation rather than intervention. Expansion of a terse line is the example here: "we have expanded it a bit" got a note only because he happened to think it worth one, and no reader can tell from the English how often that happened silently.
The rule that falls out: an absence of anomalies in a cleaned source is not evidence of a clean original. It is evidence about the cleaner. And the corresponding discipline is to state your findings in the form the evidence supports — visible in this edition, at a spot with no disclosed intervention — which sounds pedantic until the day somebody checks.
The fifth disclosure is the odd one out and it is the most useful to this build.
"This sentence is unnecessary repetition."
It sits on the gaṇa census in chapter 85, an enormous list of named leaders with crore-counts. And what it flags is a duplication: Kāṣṭhagūḍa, Sukeśa and Vṛṣabha are given sixty crores in one line and sixty-four crores a few lines later.2
Note what that single note contains.
A repeated block — the same three names, twice, in one chapter.
And a numerical discrepancy — sixty against sixty-four, for the same beings.
Two of the build's three seam-signals, at one location, flagged by the translator in five words. And there is no vocative anywhere near it.
That matters because this build once had a compound signature — a flagged numerical failure sitting beside a flagged frame violation — which it called "as mechanical as this kind of evidence gets" on a sample of two, and which collapsed to two out of five as more instances arrived.6
Endnote 859 is the final word on it. The three signals have now been observed in every pairing except all three together, which is what you would expect if they are independent kinds of damage rather than symptoms of one cause. The compound signature is not merely unproven; it is being actively contradicted by the distribution.
The last thing worth stating is the counterfactual, because it is what makes this page a governance item rather than a curiosity.
Imagine the same translation with the notes stripped out.
The nadi/Nandi passage reads as clean Nandi text. The terse śloka reads as a normally-proportioned sentence. The Vaiṣṇava typo reads as a deliberate reference. The gaṇa census reads as a slightly long list. Every one of the six disclosures becomes invisible, and nothing in the English signals that anything happened.
And this build would have produced the same findings, at the same confidence, with no way to know which of them were artefacts.
That is the argument for the whole apparatus. The endnote layer is not commentary on the text; for this vault's purposes it is the instrument log — and this build established long ago that the notes carry the definitions the body assumes and never supplies, from the three bhaktis and the six vices to the four classes of officiating priest.8 This page is the same discovery applied to method rather than to content.
It also sets a standing requirement, and it is cheap. Any page in this vault arguing from a textual anomaly in a translated source must name the edition and state whether the translator disclosed an intervention at that spot. Where the answer is unknown, say so.
Debroy's own governance note is the model. He records that Hazra did not treat this as a Purāṇa at all and that Wilson thought it might be the Vāyu, and he refuses to adjudicate.9 A translator who declines to settle a live dispute is the same person who tells you when he has quietly fixed a river into a bull.
⚑ ENRICHED 2026-08-10 — the count was six because six was as far as the reading had gone. It is now eleven, and the title is being kept deliberately.
The Pārvatī Khaṇḍa's endnote layer (notes 906–1067) was mined at the opening of that range, and it discloses five further interventions — which settles the "six is a floor, not a ceiling" claim this page has been making on the strength of an argument rather than a count.
Three more fixes. Note 914: "The text states that they could see Shivaa. In view of what follows, there is a negative missing" — he supplies a negation the text does not have. Note 1031: karanda taken as karandava, duck, with the alternative he thinks likelier printed only in the note — "It is likely this is a typo for eranda, in which case, she is having castor oil." He translated against his own best guess and told you. And note 962, which is the most explicit deviation anywhere in the volume: "We have deviated from the text. The text should be translated as 'the mantra from the Narada tantra'… this unnecessary complication is the result of some distortion in the text at some point. A simpler statement will suffice, as in our translation."
Two more flags. Note 929 — *"These sentences don't really belong here."* A bare declaration of displacement with no repair attempted, attached to the compressed pre-telling at the end of chapter 104. Note 1059 — *"cannot be translated any other way. But it is possible that this is a typo."*
And two behaviours this page did not previously have a category for. Note 1015 is a pre-emptive defence — "There is no inconsistency since there was more than one leader with the same name" — answering a seam-reading nobody made. The apparatus has started arguing with its reader. Note 1064 is an admission of non-comprehension: "The allusion to varaha is unclear." Neither is a repair. Both are the translator's attention becoming visible in a new way.
What this range adds that the earlier one could not: flagged-versus-fixed is now demonstrable inside a single Khaṇḍa. Note 929 preserves the evidence and repairs nothing; note 962 repairs and files the original in the back matter. Thirty-three notes apart, same translator, same stretch of text. The distinction this page argues for is no longer assembled across the volume — it is visible in one place.
On the title. Six Disclosed Repairs now understates its own subject, and it is being left alone rather than reslugged. Renaming would break every inbound link in the corpus for a gain that is cosmetic, and this build's hard rule is that a link target must be a verified file. Read the title as the count at first assembly, not as a total. The current figure is eleven, and it remains a floor — the changes a translator does not experience as changes are, by construction, unlogged.
⚑ ENRICHED 2026-08-11 — Volume 2 raises the disclosed-repair floor and adds a category this page did not have.
Volume 1's count of six was recorded as a floor. Volume 2 clears it, and the new instances are worth separating by kind.
Disclosed at block scale, not line scale. Endnote 501 sits on a chapter heading — ch 214 — and covers a whole section: "The composition of these Andhaka chapters is quite different. The sentence constructions make comprehension difficult. Sometimes, the sentences are incomplete. This makes it difficult to understand and translate and some liberties have been taken."D1 That is a larger admission than any single-line repair, and it is the only one in either volume anchored to a boundary rather than to a sentence.
Inconsistencies raised and left open. Endnote 504: "There is an inconsistency. Having earlier used the dual, the text now mentions three ministers." Endnote 719: *"Eight forms have been mentioned and the adjective 'eighth' doesn't belong here."*D1
Inconsistencies raised and explained away. Endnote 500 notices that Hiraṇyakaśipu was earlier the elder brother and now the younger, and resolves it — "There is no inconsistency, because, for twins, there are two notions of who is older."D1
Usage flagged as unusual. Endnote 507: "The use of the word rakshasa is unusual, since rakshasas are different from daityas/danavas."D1
Meaning conceded as unrecoverable. Endnote 511, in full: "Meaning difficult to understand, since they were sarcastic."D1
Content supplied that the text omits. Endnote 532: "Thought not explicitly stated, Shiva's vyuha was in the form of a mountain."D1
⚑ And a category this page did not anticipate: silent repair. Endnotes 716 and 717 supply the complete five-face and eight-form mappings — organ by organ, element by element, with the directions each face turns on a five-faced liṅgam — for chapters whose bodies deliver those schemas incompletely.D2 Nothing marks the gap. A reader working from the notes comes away with a tidy system the text never contains.
Which is the finding this page should carry forward. Debroy discloses generously and consistently when he changes a word. He does not disclose when he completes a structure, and completing a structure is the more consequential intervention — because a changed word is visible as a change, and a supplied table is invisible as an addition.
Strongest support is the six notes themselves, quoted, spanning twenty-one chapters, in the translator's own words, none of them referring to any other. Five further disclosures from notes 906–1067 are recorded in the enrichment block above, taking the disclosed total to eleven.
Tension one, and it is the page's own. Six is a floor. The class of change a translator does not think worth logging is by construction undisclosed, and expansion is the demonstrated example.
Tension two. The vault's lenses work on flagged anomalies and are silently defeated by fixed ones. The same translator does both, in the same volume, twelve chapters apart.
Tension three. Endnote 788's expansion degrades the one measurement the repeated-block lens depends on — verbatim similarity — and does so invisibly.
Tension four. Endnote 859 puts a repeated block and a numerical failure at one site with no vocative, which contradicts the compound-signature reading rather than merely failing to support it.
Open questions. Are there disclosed interventions in the endnote ranges this build has not yet opened — ch 96–150? And is there any spot where Debroy flags an anomaly and records a fix, which would be the cleanest possible calibration of the two behaviours against each other?
The Vocative That Does Not Fit is the lens most exposed by this page. Its five violations are all flagged instances, which is what makes them usable — and its own honest admission, that it has never found a violation the translator did not catch first, now reads differently. The lens is not weak. It is measuring the translator's disclosure as much as the text's damage.
The Question Asked Twice holds the repeated-block finding and is the page most affected by endnote 788. Its comparison is made in English, and expansion is invisible in English. The finding stands; its precision does not survive as stated.
Give Me a Task is the build's one finding produced with no help from the translator at all — six occurrences of a formula that Debroy annotates nowhere. It is therefore the only one of the three lenses whose evidence is not mediated by his editorial habits.
The Disputed Status of This Text holds his refusal to adjudicate Hazra and Wilson. The two pages describe one disposition: a translator who declines to settle what he cannot settle, and tells you when he has settled something small.
Barnum and Planned Partial Disclosure (business). The case is a showman who published his own tricks — against every instinct in the surrounding material, which says hide the mechanism — and profited by it, because disclosure converted suspicion into participation.
Debroy's six notes are the scholarly form of the same move, and the payoff is structurally identical: admitting the repairs is what makes the unrepaired text credible.
The insight neither produces alone: partial disclosure buys trust in proportion to how costly the disclosed items are — and the costly ones here are precisely the ones a reader would never have caught. Nobody checking the English could have detected the nadi/Nandi substitution or the expanded śloka; he volunteered both. Barnum's published tricks work the same way, and the pair identifies the failure mode as well as the mechanism: disclosure earns trust only for the class of thing disclosed. Six repairs establish that Debroy discloses repairs. They establish nothing about the changes he did not experience as changes, which is the category that actually threatens the argument — and a reader who reads six admissions as evidence of general transparency has been persuaded by exactly the structure Barnum was selling.
Site Selection as Editorial Act (history). Olivelle's argument is that an inscription's location is not incidental — which text went where, and in what order a visitor encountered it, was itself an editorial decision as consequential as the wording.
The parallel is not location but layer. Debroy's interventions live in the endnotes; his translation lives in the body. A reader who reads only the body reads a repaired text as an original one, and the repair notice is a hundred pages away in a section most readers never open.
What the pair produces: where a disclosure is placed determines whether it functions as a disclosure at all. Olivelle's point is that placement carries meaning; this adds that placement carries effective visibility, and the two come apart. A note at the back is formally complete and practically invisible — which is why this build's method finding, that the endnotes carry what the body assumes, keeps recurring. For a corpus like this, the apparatus is not secondary material; it is where the text tells you what kind of object you are holding, and any vault page citing this source from the body alone is reading half an edition.
Sharpest implication. Debroy discloses six interventions across Volume 1 — not six flagged errors, six places where he changed or added and said so. "The text repeatedly says nadi (river). We have corrected it to Nandi." "Since this shloka is terse, we have expanded it a bit." "This is a typo." "Some liberties have been taken." "Should probably read Pūṣti." "This sentence is unnecessary repetition." Every seam argument this vault makes rests on a distinction those six force into the open: flagged versus fixed. A flag leaves the anomaly in the English and preserves the evidence; a fix removes it and leaves a receipt at the back of the book. The vault's lenses work only on the first, and the same translator does both, twelve chapters apart. Endnote 788 is worse than it looks: expansion of a terse line is invisible in the output and degrades precisely the measurement the repeated-block lens depends on, which is verbatim similarity in English. And endnote 859 quietly settles a live methodological question — it puts a repeated block and a numerical discrepancy at one site with no vocative anywhere near it, so the three seam-signals have now been seen in every pairing except all three together, which is what independent damage looks like rather than one cause. Six is a floor, not a ceiling. The changes a translator does not experience as changes are by construction unlogged, so the honest form of every finding in this build is narrower than it has been stated: visible in Debroy's English, at a spot with no disclosed repair.
Generative questions.