⛑ REWRITTEN 2026-08-11, and the thesis has changed. This page ran for two volumes as "a lens that never collapsed — it simply never did anything." That verdict was correct about what the instrument detects and wrong about what it is for. Volume 3 shows the flagged violations are not scattered: they cluster, and every cluster sits on material carrying a foreign narrative frame. It is not a detector. It is a map. The old page's evidence is preserved below in full, including the finding that failed.
Śiva is speaking. Brahmā and Viṣṇu are in front of him. He is explaining, at length and with three analogies, that he and Rudra are the same being and that nobody should think otherwise.
And then, mid-speech, he says this:
O great sage! One should not seek to make a difference between him and me.1
There is no sage present.
Brahmā is there. Viṣṇu is there. Neither is a sage, and neither has been addressed as one anywhere in the conversation — they have been O Hari, O Vishnu, O Brahma, O best among gods, O one born from a lotus, for two chapters.
Debroy stops and puts a note on it:
While Brahma can be addressed as a great sage, this is probably an inconsistency and is meant for Narada.2
Nārada is not in the scene. Nārada is two frames out.
The denominator goes first, because a page that reports arrivals without the count drifts upward without anyone deciding to inflate it.
| Chapters run | Instances | Found unaided | |
|---|---|---|---|
| Volume 1 | 150 | 5 | 0 |
| Volume 2 | 154 | 1 | 1 (contested — see below) |
| Volume 3 | 160 | 11 | 0 |
| Corpus | 464 | 17 | 1 |
Across four hundred and sixty-four chapters this instrument has found one thing the translator had not already flagged, and that one is arguable.3
⛔ No -lens language. No "mechanically checkable." The ratio is one in seventeen and it is the honest headline.
⚑ And Volume 3's rate going up does not make the instrument better. It makes the translator's flagging denser in this volume — eleven flags in one hundred and sixty chapters against five in Volume 1's hundred and fifty. Every one of the eleven is his.
The single contested unaided find, for completeness. Volume 2 ch 186 is a sustained hymn to Viṣṇu, and inside the avatāra sequence it says "You were the acharya who taught yoga to Jainas. O Uma's lord! You assumed the form of Buddha."E1 Umā's lord is Śiva; the notes bracket the line and do not touch it.E2 And it should be weighted almost not at all, because fourteen chapters earlier the same Saṃhitā runs the identical move in reverse, deliberately and at length: a hymn to Śiva absorbing Vaiṣṇava vocatives wholesale — "You are the divinity, Narayana. You are beyond Narayana."E3 House style running both directions is not a seam. [n=1 OBSERVED], held.
Here is what two volumes of null results were concealing.
The flagged violations do not scatter. They cluster — and every cluster sits on material with independent evidence of being imported from another text.4
| Cluster | The flags | The foreign frame | Independent evidence of import |
|---|---|---|---|
| ch 338-6(24) | — | Nārada ‹ the Mahābhārata | Note 1299 announces it in advance: "the next chapter reproduces those sections from the Mahabharata, almost verbatim" |
| ch 352-6(38) | 1411, 1412, 1414 | Vaiśampāyana → Janamejaya | Note 1411 names it outright — and the body states it too |
| ch 354–356-6(40–42) | 1432, 1439, 1440 | Mārkaṇḍeya → Bhīṣma → Yudhiṣṭhira | Note 1431 flags a roster mismatch across the same two chapters; note 1434 a name from nowhere |
Note 1411 is the decisive one. The body reads "O brahmanas!", and Debroy explains: "This is an inconsistency, because Vaishampayana is speaking to Janamejaya, though the conversation is being repeated to the brahmanas."5
That is the Mahābhārata's own outermost frame, running inside the Shiva Purāṇa, with the Purāṇa's Suta-to-sages vocative bleeding into it three times in one chapter. And the body names the frame itself: "This is exactly what Vaishampayana was asked by Parikshit's son."
⭐ The mechanism, stated precisely. When this text absorbs a block that arrives with its own speaker and listener, the host's vocatives leak into the graft — because the reciter is speaking inside the outer frame while transmitting the inner one. The leak is not a scribal slip. It is a join. [n≥3 PATTERN] — three clusters, each with independent import evidence.
⚑⚑ And the reframe has a second instrument agreeing with it, built for something else entirely. All three clusters are in the Umā Saṃhitā — the one Saṃhitā that runs forty-six per cent over the shloka budget the text publishes for itself, in a table whose other five rows sit within a few per cent.6 Two wholly independent measurements, one arithmetic and one grammatical, name the same section as the one that grew.
Two of Volume 3's clusters do not fit the import reading, and the page is worth less without them.
Ch 341-6(27), note 1317. Mid-speech from Śiva to Devī, the body reads *"Engaged in yoga, there must be proper dharana first. Dhyana follows. O sage!"* Debroy's explanation is not import: *"This inconsistency has crept in because the entire conversation is being repeated to Vyasa."*7
Ch 380-7(15), notes 1687–1688. Again Śiva→Pārvatī, "O sage!" and "O lord among sages!", and again the explanation is re-narration depth rather than a graft.8
⛔ So the pattern is clustering on imports, not all violations are imports. Two of five Volume 3 clusters are ordinary frame-slippage in re-narrated material. Stated at that strength and no stronger.
⚑ And note 1317 costs the page its oldest structural claim. Volume 1's finding was that all five violations sat at the innermost level of quotation, five for five, and that the innermost layer is both the easiest to lose track of and the most portable. Ch 341's frame is four deep — Suta ‹ Sanatkumāra ‹ Vyāsa-as-listener ‹ Śiva→Devī — and the leak comes from the second layer, not the innermost. Debroy's own explanation names the outer layer. The depth correlation is qualified, not confirmed, and it does not extend past Volume 1.
The frame count is what makes any of this legible, and this literature nests deep.9
Level 1. Suta is reciting to the sages at Naimiṣāraṇya. Level 2. Suta is reporting what Vyāsa said. Level 3. Inside that, Brahmā is speaking to Nārada — where the whole Rudra Saṃhitā cosmogony lives. Level 4. And inside that, Brahmā quotes Śiva speaking to Brahmā and Viṣṇu.
The punctuation in the translation carries all four: '"Brahma said, '…Shri Mahesha said, "…"'"'
So the vocative appears at level 4 and it belongs at level 3 — where O great sage is not merely acceptable but the correct and expected form, because Brahmā has been addressing Nārada that way throughout.
One of those has fallen through the floor of its own frame and landed one level down.
Volume 1's five all did this. Volume 3 shows the fall can also run the other way, and it adds a species Volume 1 did not have: an unmarked interlocutor switch. Inside seven lines at ch 358-6(44) the speaker flips twice with nothing in the English to mark it, and Debroy supplies both corrections — "In this part, Suta is not talking to the sages, but is asking his teacher, Vedavyasa" and, four clauses later, *"Suta is now speaking to the sages."*10
⚑ That is a different defect from a misfitting address: not a wrong word inside a correct frame, but a frame change with no signal at all. It goes in the tally and not in the depth correlation.
One thing on this page is not about the text at all, and it survived both structural collapses.
Debroy's language changes as he goes, across five Volume 1 notes that never refer to one another:
| Ch | What he writes |
|---|---|
| 41 | "probably an inconsistency and is meant for Narada" |
| 61 | "this is an inconsistency" |
| 77 | "this is clearly an inconsistency" |
| 95 | "it is likely that this is meant for Narada and has crept in" |
⚑ He is not merely getting more certain. He has moved from naming a defect to naming a process — crept in is a claim about how the sentence arrived, not about whether it belongs. And he does it without ever assembling the five, because his notes are local and none of them cites another.
Volume 3 confirms the verb. Note 1317 uses the identical construction — "This inconsistency has crept in because the entire conversation is being repeated to Vyasa" — and supplies the mechanism outright.7
⚑ Which is worth stating as what it is: a translator converging on a theory of the text's composition that he never states as one, in the same apparatus that elsewhere performs twenty-three undisclosed-in-the-body editorial acts. The reader who follows every note gets the theory. The reader of the body gets a smooth Purāṇa.
For a while this page carried a second correspondence, and it was the strongest thing on it. It is now the most instructive thing on it, because it failed.
At chapter 32 Debroy flags two things within a few lines: endnote 502, the third bead-array's total "falls slightly short of one thousand"; and endnote 503, "O lord of sages!" in a speech to Pārvatī, "this is an inconsistency."11 At chapters 40–41, two more: endnote 576, of the thirty-eight aksharas, "One has to guess how this number of 38 is arrived at";12 and endnote 580, the "O great sage!" above.
Both sites paired a numerical failure with a frame violation, the pairing was the vault's own, and the decisive test was stated plainly: find a third site with both.
Three more sites arrived and none of them had both. Chapter 77 has the violation and no flagged numerical failure in chapters 76–78. Chapter 61 has the violation and none either — the notes around 746 are botanical, etymological and geographical.13 Chapter 95, likewise.14
The compound signature stands at two out of five and the numerical pairing was coincidence.
⚑ That is worth recording rather than quietly dropping, because the depth correlation went the other way over exactly the same three instances — from 2/2 to 5/5. One regularity strengthened and one collapsed on identical evidence, which is what a real test looks like.
⛔ And the page's own language was the failure. The compound signature was written up here as "as mechanical as this kind of evidence gets." It was two out of two at the time. The overconfidence was not in the observation; it was in the register. ⚑ Volume 3 has now added the second half of the lesson: the surviving regularity, the depth correlation, did not survive either. Note 1317's second-layer leak breaks it. Both of the page's structural claims have now failed on later evidence, and what has replaced them — the import-clustering — is the first one supported by an instrument other than itself.
You are reading a long report that four people contributed to, and it has been through an editor who smoothed it.
You will not catch the joins from the prose. Everyone writes competently and the editor has made the paragraphs match. So do not read for style. Read for who the document thinks it is talking to.
Somewhere on page eleven a sentence says as we discussed on Tuesday. And you were not at a meeting on Tuesday, and the report is addressed to the board. That sentence came from an email. It was pasted in whole, and the paste brought a piece of a different conversation with it — one that had a different sender, a different recipient, and a shared Tuesday.
Nobody caught it because catching it requires you to notice an address, and readers do not attend to addresses. They attend to content. As we discussed on Tuesday reads as filler; the eye slides over it exactly the way it slides over O great sage.
And then — this is the part two volumes of this page got wrong — do not expect the address check to be your detector. Run it and it will find almost nothing you were not already told. What it will do, and what nothing else does as cheaply, is tell you where. Take every misfitting address anyone has ever flagged in the document, plot them, and look at the distribution. If they scatter, you have sloppiness. If they cluster, each cluster is a paste, and you now have a map of which sections came from outside — which is a far more useful thing than a list of errors, and it is available even when every individual flag was somebody else's work.
Then find a second instrument that does not know about the first. Word counts against a stated budget. Timestamps. Formatting inherited from another tool. Anything measuring a different property. Two instruments agreeing on a boundary is worth more than either at ten times the sensitivity, because the failure modes are unrelated.
Evidence. All five Volume 1 flags read at their own endnote lines;2111314 the Volume 2 hymn and its bracketing notes read in full;E1E2E3 all eleven Volume 3 flags recovered by complete scripted census of 2,124 endnotes and read in full at their own lines;5781015 the three import clusters cross-checked against independent evidence — a translator-announced verbatim reproduction, a named external frame stated in both note and body, and a flagged roster mismatch;515 the Umā Saṃhitā's shloka overrun taken from the text's own published budget.6
Tension 1 — Debroy's hedge, preserved. He allows that "Brahma can be addressed as a great sage" before rejecting it. The innocent reading is available and he does not close it off.
Tension 2 — the instrument has found nothing. Its demonstrated value is confirmation, false-positive elimination, and now localisation. Not discovery. ⚑ The reframe does not repair the detection record and must not be written as though it does: the detection was always Debroy's, and the pattern across his flags is the vault's contribution.
Tension 3 — clustering is not proof of import. Two of Volume 3's five clusters are re-narration depth by the translator's own account. What licenses the reading is that the three with independent import evidence are also the three that cluster densely, and that a second instrument names the same Saṃhitā.
Tension 4 — note 1439 flags and half-retracts in nine words. "An inconsistency, though Bhishma can be addressed as sage."15 Volume 1 catalogued a pre-emptive defence — a note answering a seam-reading nobody had made. This is its mirror: a retreat inside the flag itself. [n=1 OBSERVED], and it means the flag-set is not a clean population.
Tension 5 — the depth correlation is now qualified and the page kept it anyway. Five for five in Volume 1; broken by note 1317 in Volume 3. It is retained above because the Volume 1 five are still five, and because a correlation that holds in one volume and not another is information about the volumes.
Open questions. Volume 1 contains the Vidyeśvara Saṃhitā, which the shloka budget shows came up twenty-three per cent short. On the absorption reading it should carry no imported frames at all. That is a real falsification condition, it is runnable in an afternoon, and it has not been run. And: would this instrument detect a ritual graft, where there are no speakers to leak?
Against The Saṃhitā That Grew: the convergence that produced this rewrite. That page arrives at the Umā Saṃhitā by arithmetic on a published table; this one arrives there by a census of misfitting addresses. Neither instrument was built to find the other and they name the same fifty-one chapters. ⚑ That is the strongest structural result in three volumes of work on this text, and it is worth noting that it required both pages to be wrong about themselves first — one thought it was measuring scale, the other thought it was a detector.
Against The Question Asked Twice: that page documents the volume's other seam-detector — a twelve-item question repeated verbatim across frames, in two passages whose vocatives were impeccable — and argues for its superiority. This page does not overturn that; it narrows it. The duplication lens finds joins where whole blocks were reused. This one finds joins where a fragment carried its context with it. Different sizes of damage, and a compiler who copies a large block is likely to fix the address while one inheriting a short doctrinal passage may not.
Against The Nested Narrators: that page argued this frame structure "permits accretion and forbids revision." ⚑ Volume 3 supplies the sharpest possible instance — an entire foreign outermost frame, the Mahābhārata's own, running inside the Purāṇa with the host's vocatives leaking into it, and no position anywhere in the book from which anyone could have noticed.
Against The Apparatus Is the Translator's: the uncomfortable dependency, stated. Sixteen of seventeen instances on this page are the translator's flags. What the vault contributes is the distribution, which no individual note could contain — but the population being distributed is entirely his, and if his flagging is denser in some Saṃhitās than others for reasons unrelated to the text, the clustering is an artefact of his attention. That possibility is not excluded here. The Umā convergence is what makes it unlikely, not impossible.
Literacy Rupture: The Shift from Oral to Written Knowledge Organization — the argument there is that writing does not neutrally transmit oral knowledge but reorganises it: knowledge leaves the keeper's body and enters an object, becoming portable, duplicable, and usable by someone who does not understand it.
That last clause is where the vocative lives.
A vocative is a trace of a speaking situation. In live recitation it cannot go wrong for long — the performer can see the room, the audience knows who is being addressed, and a wrong address is audible immediately to everyone present. The error has nowhere to hide. The moment the material becomes an object that can be copied, the address travels with the text into rooms it was never spoken in, and nobody is present to be confused by it, because the reader is not being addressed at all.
The insight neither domain gives alone: stray vocatives are a literacy artefact, and they are only possible in a tradition that writes. Which makes them better than a curiosity. They are evidence not merely that a text is composite but that the compositing happened after the material was written down, because in a purely oral tradition this error class self-corrects at every performance. For a literature with almost no external dating, that is a rough ordering instrument: a frame violation is a lower bound on literacy at the point of that join.
⚑ And the reframe sharpens it into something the literacy page can use directly. The leak is directional. It is always the host's address appearing in the guest's material, never the reverse — because the person doing the joining was reciting in his own idiom while carrying somebody else's text. So a frame violation does not merely date a join; it tells you which side of the join was the room and which was the page. That is a distinction an oral tradition cannot produce at all, and it is precisely the moment the literacy-rupture page describes: the instant when one conversation is being performed and another is being read out inside it.
One handshake. The forensic-inference material once carried here as a second connection is not reached for: at one unaided find in seventeen, a page about detecting concealment has nothing to detect.
Sharpest implication. Mid-speech from Śiva to Brahmā and Viṣṇu, the text says "O great sage!" and there is no sage in the room; Debroy flags it and names its owner — "meant for Narada" — who is two frames out. The corpus does this seventeen times across four hundred and sixty-four chapters, and sixteen of the seventeen were found by the translator, not by this vault. For two volumes that verdict stood as a lens that never collapsed and never did anything. Volume 3 shows it was measuring the wrong thing. The flags do not scatter: they cluster, and each cluster sits on material with independent evidence of being imported — an announced near-verbatim Mahābhārata reproduction, the Mahābhārata's own Vaiśampāyana-to-Janamejaya frame running inside the Purāṇa with the host's vocatives bleeding into it three times in one chapter, and a five-deep Mārkaṇḍeya frame that breaks three times beside a flagged roster mismatch. The leak is not a slip; it is a join — the reciter speaking inside the outer frame while transmitting the inner one. And all three clusters sit in the one Saṃhitā that runs forty-six per cent over the shloka budget the text publishes for itself. Two instruments, arithmetic and grammatical, neither built to find the other, naming the same fifty-one chapters. Against which: two further clusters are ordinary re-narration depth, and one of them breaks the page's own long-standing innermost-frame correlation. Both of this page's structural claims have now failed on later evidence, and the one that replaced them is the first supported by something other than itself.
Generative questions.