A reading instrument was pointed at this corpus before the reading started. It was built to catch frame violations — moments where a hymn addressed to one god slips and uses the vocabulary of another, on the theory that a Śaiva compiler working over Vaiṣṇava material would leave fingerprints in the vocatives.
It ran across 304 chapters, two volumes, and it found one thing.
That number is the page. Not because one find is a bad yield — one real find in 304 chapters is a perfectly respectable outcome for a cheap instrument — but because of what almost happened next, which is that a single hit got treated as a validated method. It was not. Everything else this build produced came from instruments nobody had planned, and the honest record of which tool found what is more useful than any of the individual findings.
Chapter 186. The devas have fled Jalandhara and gone to Vaikuṇṭha, and they praise Viṣṇu — sustained, formal, and unambiguously his: Matsya recovering the Vedas, Kūrma under Mandara, Sūkara raising the earth, Vāmana deceiving Bali, Paraśurāma, Rāma, Kṛṣṇa with Rādhā, Kapila teaching Devahūti, Vedavyāsa collating the Vedas. It closes "O Hari! You are attired in a yellow garment. Your mount is Tarkshya."1
And inside it, in the avatāra sequence between Kṛṣṇa and Kalki, sits this:
"You were the acharya who taught yoga to Jainas. O Uma's lord! You assumed the form of Buddha."1
Umā's lord is Śiva. It is a Śaiva vocative, in the middle of a Vaiṣṇava hymn, addressed to Viṣṇu.
The note layer brackets the line and does not touch it. Endnote 353 sits on Rādhā four clauses earlier — "One of the rare instances of Radha being mentioned in Puranas other than Brahma Vaivarta Purana" — and endnote 354 sits on Devahūti four clauses later, "Kapila's mother, to whom, he taught samkhya yoga."2 The annotator is attentive on both sides of it and says nothing about the vocative between them.
⚑ Tagged [n=1 OBSERVED], and the tag is doing real work. One instance is one instance. It is not a pattern, it does not support "the compiler's hand shows," and it never will on its own.
And here is why the tag has to stay, because the counter-reading is strong and comes from fourteen chapters earlier in the same section.
Chapter 172 is a hymn to Śiva, and it absorbs Vaiṣṇava material wholesale, at length, without apology:
"I prostrate myself before Pradyumna, Aniruddha, Vasudeva and Samkarshana… I prostrate myself before the one who destroyed Kamsa. You are the one who crushed Chanura. You are Damodara, the one who swallowed poison. You are Hrishikesha Achyuta… You are Adhokshaja… You are the divinity, Narayana. You are beyond Narayana. Narayana is your own form. You emerged from Narayana's body."3
That is not a slip. It is a sustained argument, and the last clause states the theology explicitly — the absorption is the point being made.
The same chapter goes further and files Viṣṇu inside a list of Śiva's superlatives: "Among those devoted to Shiva, you are Achyuta."
⚑ So the volume demonstrably does this deliberately, in the reverse direction, four chapters into the same saṃhitā. Which means "O Umā's lord!" in a Viṣṇu hymn may not be a seam at all. It may be the house style running the other way — a text that treats the two vocabularies as interchangeable because it holds that the two gods are, with Śiva senior.
The find survives as an observation. It does not survive as evidence of compilation, and stating it as such would be exactly the error the tag exists to prevent.
The planned instrument returned one ambiguous hit. Four unplanned ones returned the build's substantive findings, and none of them was designed in advance.
Arithmetic. Chapter 219 announces "one hundred and eight forms of the paramatman" and then enumerates them, numbered on the page, from (1) Mahadeva to (100) Shiva.4 Eight short. Nobody was looking for this; it emerged from counting a list that happened to be numbered.
Attribution. In chapter 214 the speech-tag "The ministers said" appears four times in one stretch, and one of those four is the ministers quoting Śiva verbatim in the first person with no marker distinguishing it. The translator has to intervene twice to keep the reader oriented — endnote 508, "Reporting Shiva's words. The words are full of sarcasm", and endnote 510, "Earlier, the ministers reported Shiva's words. These are their own words."2 That is a seam visible only if you are tracking who is speaking across a long stretch, which no plan called for.
Doublet detection. Four unflagged repetitions, none of which any lens was pointed at, all found by noticing that a chapter felt familiar and going back to check.
Whole-body construction sweeps. Counting the volume's own formulaic phrases — will be known as, came to be known as, that is the reason … known — produced the naming corpus that corrected an earlier claim about etiologies.
⚑ Four instruments, none of them planned, all of them cheap, and all of them working the same way: they count something the text does mechanically, and they notice when the count comes out wrong. The frame lens was different in kind — it was looking for a quality (does this vocative feel out of place), and qualities do not come out wrong. They come out arguable, which is what happened.
That is the transferable lesson, and it is worth stating flatly: the instruments that worked measured things the text was not choosing. A compiler chooses vocatives. A compiler does not choose whether the list he copied has 100 items or 108, or whether a speech-tag he inherited covers two voices. The unchosen features are where the seams show.
The arithmetic instrument was re-run deliberately, across every announced count in the volume — 108, 101, 28, 112, 16,000, 18 padmas, 86 divisions, 300,000 akṣauhiṇīs, 54, 505. The 28-Vyāsa / 112-disciple figure checks out exactly (28 × 4).
It found nothing else. The ch 219 slip stands alone.
⚑ Record that, because it is what makes the one discrepancy credible. An instrument that returns a hit every time it runs is measuring the operator. An instrument run twice across ten counts that returns one anomaly and nine clean results has demonstrated that it can come back empty — which is the only thing that licenses believing it when it does not.
The same applies to the frame lens, and it is the reason this page exists rather than a page announcing a discovery. The lens ran 304 chapters. Its honest headline from Volume 1 was zero unaided frame violations in 150 chapters. Volume 2 gave it one, contested. That is a null result with a footnote, and it should be reported as one.
Three failure modes, each caught during the build rather than in review, and each specific enough to be reusable.
Inline note markers break phrase matching. A sweep for the precedence formula missed a genuine instance at ch 180 because the phrase reads "Until Vinayaka313 is worshipped" — the footnote marker sits inside the phrase and no string search for "Until Vinayaka is worshipped" will ever return it. On a text with 1,000+ inline markers, any phrase-level sweep is silently lossy, and the loss is invisible.
Sampling on the dependent variable returns the answer it started with. An early claim held that nearly every name-origin in the volume derives from an injury, on the strength of nine verified out of nine. But the nine had been collected because they were injury-derived. A proper sweep of the volume's own naming constructions returned 32 instances across 26 chapters, of which roughly eleven are injury-derived — the volume also names by quality (Kālabhairava, Kālāgni), by place, and by lineage catalogue, and those are the majority. The corrected claim holds for narrative name-origins only.
⚑ And sequence claims cannot be built from notes. The build's own scoring came out at quotation-level claims 7 for 7, sequence-level claims 0 for 3. Every claim of the form this passage says P held. Every claim of the form X caused Y or this is the only instance of Z failed, because each had been assembled from two passages read separately and never re-read in order.
That is the most useful number on this page. It is not a general truth about reading; it is a measured property of one build, on one text, with one set of notes. But within that scope it is exact, and it converted directly into a rule: any page asserting causation or uniqueness gets a fresh sequential re-read before drafting, and pages asserting only what a passage says can be built from the log.
Not nothing, and the distinction matters.
"O Umā's lord!" in a Viṣṇu hymn is a genuine observation about the text, verified in place, with the note layer's silence on it verified too. What it cannot do is bear the weight it was built to bear.
Three things would change that, and none is available inside this volume. A second instance would move it from [n=1 OBSERVED] to [n=2 SUGGESTIVE]. A directional asymmetry — Śaiva vocatives appearing in Vaiṣṇava hymns but not the reverse — would make the compilation reading much stronger, and ch 172 is decisive evidence against that asymmetry existing. And a manuscript variant at that line would settle it outright, which is outside what a reading of one English translation can reach.
⚑ So the correct disposition is to hold it, tagged, with the counter-reading attached, and to say plainly that the instrument that produced it has not earned promotion. A lens that runs 304 chapters and returns one contested hit has told you something real — that the thing it was looking for is rare or absent — and that is a finding, not a failure. The failure would be reporting the hit and not the denominator.
You are about to read something long and you want to build an instrument for it. Here is what this build learned, in the order it learned it.
Pick features the author was not choosing. Vocatives are chosen; counts are not. Speech-tags are chosen; whether a speech-tag covers two voices is not. Word choice is chosen; whether an announced list matches its enumeration is not. Anything the author was deciding will yield arguable results, and arguable results cannot be counted. Aim at the mechanical residue.
Write down what the instrument found and what it did not. Before you report the anomaly, report the denominator: ten counts checked, one discrepancy. That sentence is what converts an interesting observation into evidence, and leaving it out is the single most common way an honest reader misleads.
Do not upgrade a method on one hit. The temptation after a lens finds something is to say the lens works and read everything through it. One find tells you the lens fired once. It does not tell you the lens is calibrated, and it definitely does not license going back to reinterpret earlier material through it.
Look for the counter-instance in the same neighbourhood. Before publishing "this hymn slips into the other tradition's vocabulary," check whether the text does the reverse deliberately somewhere nearby. Fourteen chapters away, this one does, at length, with an argument attached. That check takes ten minutes and it is the difference between a finding and an embarrassment.
And know which of your claims are sequence claims. This passage says P and P caused Q are different species with different failure rates, and you can tell them apart before you write. Sort your claims into the two piles and give the second pile a fresh read in order. On this build the first pile went seven for seven and the second went nought for three — and every one of those three would have shipped, because all of them were assembled from notes that were individually accurate.
Evidence. Ch 186 read in full and the vocative located in place;1 endnotes 353 and 354 read at their own lines to confirm the bracketing silence;2 ch 172 read in full for the counter-reading;3 ch 219's enumeration counted directly on the page from (1) to (100) against the announced 108;4 ch 214's four speech-tags located and endnotes 508 and 510 read.2
Tension 1 — the find may not be a find. Ch 172 does the reverse move deliberately and argues it.3 The build reports the observation and refuses the inference, and this is not a resolution.
Tension 2 — the strongest instruments were the unplanned ones. Which raises an uncomfortable question about planning: if arithmetic, attribution, doublets and construction sweeps all emerged from reading rather than from design, the value of pre-specifying a lens at all is unclear on this evidence.
Tension 3 — every failure mode was caught by the build's own rules and not by review. That is reassuring about the rules and says nothing about what the rules do not cover, which is unknown by construction.
Open question. The frame lens returned zero in Volume 1 and one contested hit in Volume 2. Volume 3 is unread, and no claim is made about what it contains. Whether a second instance exists is the single thing that would move this finding, and it cannot be settled here.
Against The Four Doublets and the Four Disclosures: the two method pages divide the same territory. That one holds a finding the annotation layer was structurally unable to see; this one holds a finding the annotation layer sat directly beside and did not mention. Different failure modes, and only the first is architectural.
Against Heterodoxy as a Manufactured Weapon: ch 186's Buddha-avatāra line belongs to that page as content and to this one as evidence. The vocative is a method artefact; the polemic it sits inside is a substantive finding, and conflating them would be exactly the error this page is about.
Against Damage Becomes Institution: that page carries the corrected etiology claim. This one carries why the original was wrong — nine confirmations collected because they confirmed.
Against Andhaka, and the Clause He Wrote Himself: ch 214's attribution collapse sits in the block endnote 501 already flags as compositionally distinct, which makes it corroboration rather than discovery — a distinction this page insists on keeping.
Levels of Evolutionary Analysis and Hypothesis Generation — that page treats the discipline of deriving a testable prediction from a theory before you go looking, so that the evidence can come back negative.
This build is a small natural experiment in exactly that, and it came out lopsided in an instructive way.
The pre-specified instrument behaved impeccably by that page's standards. It was derived from a hypothesis (a Śaiva compiler working over Vaiṣṇava material leaves fingerprints in the vocatives), it made a falsifiable prediction, it ran across a large corpus, and it returned almost nothing — one contested instance in 304 chapters, with a strong counter-reading available fourteen chapters from the hit. That is a clean, informative null, and it is worth more than it feels like it is worth, because it constrains: whatever this compiler did, he did not do that, or did it once.
The post hoc instruments produced everything substantive. Which looks, at first glance, like an argument against pre-specification.
⚑ It is not, and the reason sharpens both pages. The unplanned instruments succeeded because they measured features the author was not choosing — list arithmetic, speech-tag coverage, formulaic phrase counts. Those admit of a right answer, so a post hoc look at them cannot be talked into a result. The pre-specified lens aimed at a chosen feature, and chosen features are interpretable, which is precisely why it needed pre-specification to be trustworthy at all.
So the operative variable is not when you formed the hypothesis. It is whether the feature you are measuring can come out wrong independently of you. Where it can, discovery-mode looking is safe. Where it cannot, pre-specification is not a virtue but a minimum requirement — and even then, as here, it may buy you nothing but an honest zero.
One handshake, not two. The sampling-on-the-dependent-variable error would pair with methodology material the vault does not carry at page level.
Sharpest implication. An instrument built before the reading started ran 304 chapters and found one thing: a Śaiva vocative, "O Umā's lord!", sitting inside a sustained Vaiṣṇava hymn, with the translator's notes bracketing it on both sides and saying nothing. And fourteen chapters earlier the same volume does the reverse deliberately and at length, absorbing Pradyumna, Vāsudeva, Dāmodara and Nārāyaṇa into a hymn to Śiva and arguing the theology out loud — which means the one find may be house style rather than a seam. Meanwhile everything substantive this build produced came from instruments nobody had planned: a list announced at 108 and enumerated to 100; a speech-tag covering two voices that the translator had to gloss twice; four unflagged doublets; a phrase sweep that produced a naming corpus. The pattern is that the working instruments all measured things the author was not choosing — counts, tags, formulae — while the lens measured a choice, and choices come out arguable rather than wrong. And the build's own scoring says the rest: quotation claims seven for seven, sequence claims nought for three, every failure assembled from notes that were each individually accurate.
Generative questions.