Volume 2 tells the same story twice, four times over, without ever saying so.
That would be unremarkable in a compiled text — long compilations repeat themselves constantly, and nobody expects a Purāṇa to run a consistency check. What makes it worth a page is the other half of the record: this volume has working machinery for disclosing repetition, and it uses that machinery, out loud, at least five times.
So the finding is not the text repeats itself. It is that a text demonstrably equipped to say "I have told you this before, and here is why I am telling it again" declines to say it on the four occasions where the two versions do not agree.
The strongest of the four, because the translator's eye is demonstrably on the passage.
Chapter 152. Śiva flings the semen on the ground. "Urged by all the immortals, Agni assumed the form of a dove. Using his beak, he swallowed up all of Shambhu's semen." Pārvatī arrives, curses the gods and their wives with barrenness, and curses Agni separately. Then the gods, who consume oblations through fire, all become pregnant, are scorched, and are ordered to vomit it out. Agni is left holding it; Śiva instructs him to deposit it in a woman's womb; Nārada advises the Māgha bathing-place; the wives of the seven sages arrive.1
Chapter 154. Śiva convenes an assembly and puts the question under something close to oath — "You are constant witnesses to all karma in the universe. Did you conceal it, or do you know who did it?" — and the parties depose in turn. Brahmā: it fell on the ground. Kṣiti: "I flung it into the fire." Vahni: "in the form of a dove, I immediately flung it down on Mount Kailasa." The mountain: "I immediately hurled it into Ganga." Gaṅgā: "I… flung it into a clump of reeds." Vāyu: it turned into a child there. Then the Kṛttikās.2
⚑ These are not two summaries of one route. Ch 154's sworn deposition contains no gods' pregnancy, no vomiting, no Agni-to-womb instruction, no Māgha bathing, and no six wives — the entire ch 152 apparatus is missing. And ch 152 contains no mountain, no Gaṅgā and no reeds. The dove survives in both, and it is the only thing that does.
The formal setting makes it worse rather than better. Ch 154 is staged as testimony taken under threat, precisely so that the truth will be established — and what it establishes is a different sequence of events from the one the reader watched two chapters earlier.
And Debroy is looking straight at it. Endnote 57 sits in the middle of that deposition and reads, in full: "A few shlokas are repetitions of shlokas in the preceding chapter."3 He notices the repetition. He does not notice that the repetition contradicts what it repeats.
The most mechanically demonstrable, because you can lay the two chapters side by side and watch them track.
Ch 183 and ch 259 both run: Bṛhaspati and Indra travel to Kailāsa · Śiva blocks the road naked with matted hair to test them · Indra asks who he is, three times, and is met with silence · Indra threatens him with the vajra and raises it · Śiva paralyses the arm before the throw · Indra burns internally, "like a serpent, burnt down by a mantra" (183) / "like a valiant serpent countered by a mantra" (259) · Bṛhaspati recognises Hara and prostrates himself like a rod · he begs for Indra's life · Śiva answers with the identical simile — "A snake cannot don the skin it has cast off again" (183) / "How can a snake again don the skin that it has cast off?" (259) · Bṛhaspati presses · Śiva relents, renames him Jīva for having granted Indra his life, and flings the forehead-fire into the salty ocean, where it becomes Jalandhara.45
Every load-bearing element, in the same order, with the same two similes.
⚑ And ch 259 opens: "Hear about Paramesha's original avatara." Not again, not as I told you — original. The chapter presents itself as a first telling, seventy-six chapters after the first telling.
The note layer glosses Jīva on both occasions and derives it differently each time: note 327, "The root of the word Jiva means life, the principle of life"; note 892, "From the root jiv, meaning to revive, or come to life."3 Two annotations on the same etymology in the same volume, neither referring to the other.
The theologically largest, and the one that came out of a range that had been skipped and was recovered.
Both versions share a three-element setup: Hiraṇyakaśipu is torn open with talons; Prahlāda intercedes; the intercession fails and the rage persists. That shared frame is what makes them two answers to one question rather than two unrelated stories.
Ch 216 resolves it by curse and rehabilitation — Nṛsiṃha cursed by the Saptarṣis, swallowed, and restored at Badarī. Ch 241 resolves it by destruction — the Śarabha tears him apart, he is flayed, and Śiva wears the hide.
⚑ Two corrections belong on the record here, both made during this build.
The formulation "the two versions disagree about whether Viṣṇu dies" is withdrawn. Endnote 779 glosses "his life came to an end" with "In the form of Nrisimha."3 The translator narrows the death to the avatāra, not the god. The doublet holds; that particular sharpening of it does not.
And a competing reading must be carried rather than dismissed. The two sit in different narrative frames — Sanatkumāra to Vyāsa in the Yuddha Khaṇḍa, Nandīśvara to Sanatkumāra in the Śata Rudra Saṃhitā. That is a real compositional explanation for how both could be in one book. What it does not explain is the non-disclosure — because the Śata Rudra Saṃhitā is the very section that uses the back-reference device, and uses it two chapters earlier.
The weakest of the four, and it is tagged that way.
Ch 179. To make Tripura killable, Śiva stipulates: "The asuras can be killed only if the devas assume the forms of different animals and become Pashupatis, not otherwise." The gods balk — "They were suspicious of becoming animals" — and are sold the exit up front: "If a person practices my vow of Pashupata, he will be freed from the state of being an animal… twelve years, and three years after that." The chapter closes the loop: "Hence, all the excellent devas and asuras became the lord's animals."6
Ch 253. Suvarcā, Dadhīci's widow, learns what the gods did to her husband's bones: *"O gods! You are extremely wicked… You are avaricious and do not behave like gods. Therefore, from today, all of you, including Indra, will become animals."*7
Two accounts of how the devas came to be paśus. Neither refers to the other.
⚑ State the competing reading, because it is strong. Ch 179 is a theological status — paśu as the bound soul, releasable by Pāśupata discipline — while ch 253 is a punitive curse from an angry widow. Those may be two registers of one word rather than two origins of one condition. Tagged [n=1 OBSERVED] and deliberately not carried at the same confidence as the other three.
Now the other side of the ledger, which is what makes the four doublets a finding rather than an observation about compiled texts.
Mechanism 1 — kalpa-variance. The text has an explicit doctrine for accommodating incompatible versions. Ch 163, on Gaṇeśa: "Because of differences in kalpas, the story of Ganesha's birth is spoken about in different ways. I will now speak about Ganesha's origin during the kalpa known as Shveta."8 And it deploys it a second time at ch 298, on Gautama's two curses: "These differences occur because incidents have occurred in different kalpas." — that one raised by the listeners, with the narrator conceding "All of you have lovingly heard the earlier account too."9
Mechanism 2 — back-reference and decline. Ch 239, on Vīrabhadra: *"I have told you everything about his conduct when I spoke of Sati's conduct. Since you have heard it many times, I will not speak about in great detail."*10 The text can recognise its own prior telling and choose brevity on that basis.
Mechanism 3 — announced re-telling. Ch 286, on Kumāra: *"I have already told you about Kumara's divine conduct, but I will describe it again."*11 And ch 217, where the listener initiates it: *"I have heard about this briefly. Please tell me about it again, in detail."*12
Three mechanisms, five deployments, four of them in the narrator's own mouth.
⚑ So the volume can (a) explain a contradiction doctrinally, (b) flag a repetition and shorten it, and (c) announce a re-telling before delivering one. On the four doublets it does none of the three. Ch 259 goes further than silence and calls its retelling original.
That is the whole argument, and it is worth being precise about what it does and does not establish. It does not establish authorial deceit, and it does not establish that a single mind composed all of it. It establishes something narrower and more interesting: the disclosure conventions were available in the room and were not applied to the cases that most needed them. Whether that is inattention, layered composition that never got a final pass, or indifference to a kind of consistency modern readers care about and the compilers did not, the volume does not say.
Debroy is not an inattentive annotator — he flags repetition elsewhere in this volume without prompting, and endnote 501 is a model of the kind: "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."3
That is a translator openly telling you where a seam runs.
Which makes the pattern of what he does not flag the more legible. On the Kārttikeya doublet he sees repetition and files it as repetition (note 57). On the Avadhūteśvara doublet he glosses the same name twice with different derivations and connects neither gloss to the other (notes 327, 892). On the Nṛsiṃha doublet the surrounding notes 524–531 cover ch 215–216 and address none of it; note 779 addresses the death and narrows it.
⚑ The distinction that emerges is between local and cross-chapter annotation. Every note above is doing careful work inside its own passage. Not one of them reaches across chapters to say this is the second time. Which is a limit of the note apparatus rather than a failure of the annotator — but it means the doublets are exactly the class of thing this edition's paratext is structurally unable to surface, and a reader who trusts the notes to catch inconsistency will not find them.
And corroboration must be kept separate from discovery. Endnote 501 flags the Andhaka block as compositionally distinct; the build then found two independent confirmations in the English — the ch 214 attribution collapse, and an incomplete sentence in ch 216: "But since he had received a boon, even though his body as a result of the blow inflicted by the lord of gods, he did not give up the fight," where the subordinate clause has no verb.13 That is corroboration of a flag the translator raised first, and it belongs in a different column from the four doublets, which nobody flagged.
You are reading a long compiled document — a codebase's docs, a policy manual, a body of case law, a wiki that six people have edited over four years — and you want to know how much of it to trust.
Do not start by looking for contradictions. Start by looking for the disclosure conventions, because they tell you what the document's authors considered worth marking.
Read for the phrases that acknowledge a prior version: as noted above · this supersedes · see also · we have covered this, so briefly. Collect them. That inventory is your baseline — it tells you the document's authors had a mechanism and knew how to use it.
Then look for repetition, and sort every instance into two piles: disclosed and silent.
The disclosed ones are almost never your problem. Somebody was paying attention when they wrote those, and the versions usually agree.
⚑ The silent ones are where the seams are, and the ratio is the diagnostic. A document with no disclosure convention at all tells you very little — you cannot distinguish carelessness from an editorial style that simply doesn't mark these things. A document that marks repetition five times and stays silent on the four places where the versions conflict is telling you something specific: the convention existed, the attention existed, and it did not reach those four places. Those four are where separate material was joined, or where a revision landed without a pass over what it contradicted.
Two cautions from this material.
Watch for the version that calls itself first. Ch 259 announces an original account of an episode told seventy-six chapters earlier. In an ordinary document that is the signature of a section written independently and pasted in by somebody who did not read the whole.
And distinguish contradiction from register.
⚚ ENRICHED 2026-08-11 — Volume 3 reproduces both halves of this finding independently, and widens the gap.
The doublets. The concluding volume supplies at least five: the Sudarśana's origin told at ch 305 and again at ch 307 with three disagreements and a connective sentence saying "After this"; the Kṛṣṇa–Upamanyu episode at ch 315–317 and ch 423, disagreeing on the place, the initiation, the interval, the boon count and who the son is; the Kena yakṣa episode at ch 362 and ch 424 with the deity swapped; two incompatible etiologies for Sudyumna's sex at ch 308 and ch 350; and two origins for Kauśikī at ch 361 and ch 412.V1
The machinery. The same volume states an explicit doctrine of why accounts disagree — ch 400-8.1(13): "in different kalpas, Brahma, Vishnu and Maheshvara are born from each other… Based on what happened in one particular kalpa, maharshis speak about their relative power" — which would licence every one of them.V2
⛔ And it is applied to none of them. ⚚ The machinery is not merely available in that volume; it is in competent active use — manvantara-sequencing reconciling two deaths of Dakṣa, a correct back-reference-and-decline, an explicit identity-statement harmonising three names of one goddess, *"Shatakshi, Shakambhari and Durga are the same. All three are identical."*V3
⚚⚚ So the gap this page measures is wider than Volume 2 could show. ⛔ The volume harmonises three names of one goddess in a sentence and never harmonises its two hell systems. And the reading that follows is worth stating as a reading: a text this comfortable with announced multiplicity is probably not tolerating its doublets — it is not noticing them. The ch 179 / ch 253 pair looks like a doublet and may just be one word used theologically in one place and punitively in another. Before you file two passages as conflicting, ask whether they are answering the same question — the Nṛsiṃha pair shares a three-element setup and clearly is; the paśu pair may not be.
Evidence. All four doublets and all five disclosure deployments were re-read in full sequence for this page, and every cited endnote was read at its own line rather than quoted from notes: ch 152 and 154;12 ch 183 and 259;45 ch 179 and 253;67 ch 163, 298, 239, 286 and 217;89101112 endnotes 57, 327, 501, 779 and 892.3 The ch 216 / ch 241 pair is carried from the episode-level build with its two corrections attached.
Tension 1 — the competing compositional reading is real and unresolved. Different frames and different saṃhitās genuinely explain how contradictory material coexists in one compilation. That reading is carried, not dismissed. What it does not explain is why a text with an active disclosure convention applied it five times and not here.
Tension 2 — the paśu pair may be register rather than contradiction.67 Tagged [n=1 OBSERVED] and held below the other three.
Tension 3 — a withdrawn claim, retained. "The two versions disagree about whether Viṣṇu dies" was overturned by endnote 779, which narrows the death to the avatāra.3 The doublet survives the correction; the sharpening did not, and it is recorded here so that nobody re-derives it.
Open question. The disclosure convention runs five times in 154 chapters and the doublets run four. That is a small denominator on both sides. Whether the ratio means anything, or whether five deployments is simply what a compilation of this size throws off by accident, cannot be settled inside one volume — and the next volume is unread.
Against The Śarabha and the Hide: that page holds the ch 241 episode; this one holds its relationship to ch 216 and the two corrections that landed on the pair.
Against Kumāra, and the Son Who Left: the ch 152 / ch 154 doublet is the conception that page narrates, and this one is where the two incompatible routes are set side by side rather than harmonised.
Against The Two Gautama Accounts: the sharpest internal contrast in the build. Gautama's two versions are disclosed — raised by the listeners and explained by kalpa-variance at ch 298 — and the four here are not. Same book, same device, opposite treatment, and nothing distinguishes the cases except that somebody noticed one set.
Against Whole Narratives Annexed: that page finds the volume's citation habit pointed meticulously inward and never outward. This one qualifies it — the inward tracking is real but incomplete, and the four doublets are exactly where it fails.
The AI Second-Brain Retrieval Shift — that page argues that semantic retrieval dissolves the filing problem: you no longer have to have guessed, at the moment of saving, where a note would later be looked for, because recall now works on meaning rather than on location.
This volume supplies the strongest illustration of what that shift is actually buying, because it shows the pre-shift failure mode in a pure form.
Debroy's note apparatus is a location-bound index. Every note in it does careful work inside its own passage — note 57 correctly identifies repetition in the chapter it sits in, notes 327 and 892 each correctly derive Jīva in the chapter they sit in. Not one of them reaches across seventy-six chapters to say this is the second time, and the reason is architectural rather than a lapse of attention: an annotation anchored to a line has no way of knowing what another line says. The four doublets are precisely the class of finding that a positional index is structurally unable to surface, which is why they survived a careful annotator.
Meaning-based retrieval inverts that. Ch 183 and ch 259 share their similes, their sequence and their etymology; they are near-identical in meaning and separated by half a volume in position. A system that retrieves by meaning surfaces them as neighbours automatically, and a system that retrieves by position never will.
The refinement the pairing produces cuts against the shift's usual framing, though. What actually caught these was not retrieval at all — it was noticing that the text has a disclosure convention and counting where it was not used. Five acknowledged repetitions establish the norm; four silences become norm violations, and violations localise seams in a way raw similarity cannot. Semantic retrieval will hand you every pair of similar passages, including the hundreds that are similar for good reasons. It cannot tell you which similarity the author was obliged to mark and didn't. That judgement is still the reader's, and it is where the finding lives.
One handshake, not two. A second on translation-apparatus limits has no page-level counterpart in the vault, and manufacturing one would violate the standard.
Sharpest implication. This volume tells four stories twice without saying so, and it owns the tool for saying so. Kārttikeya's conception is deposed under oath in ch 154 along a route that ch 152 does not contain — no divine pregnancy, no vomiting, no six wives — and Debroy's note on that exact passage observes the repetition and not the contradiction. The Avadhūteśvara episode runs beat for beat across chapters seventy-six apart, down to two shared similes and the same etymology of Jīva, and the later telling calls itself the original. Nṛsiṃha's rage is resolved by curse-and-rehabilitation in one saṃhitā and by flaying in another, with the same three-element setup underneath. And against all of that, the text explains contradictions by kalpa-variance twice, flags and shortens a retelling once, and announces a retelling twice more — five deployments of a convention that never reaches the four cases that need it. The claim is not deceit and not even inconsistency, both of which are ordinary in a compilation. It is that the norm existed, was in active use, and did not extend to the seams — which makes the silences locate the joins.
Generative questions.