This volume says "there are four kinds of mukti" and then names five. It says "among these four" and names a thing that is not one of the four. It says "therefore, there are fourteen kinds of vidya" and calls them eighteen in the next sentence. It lists eighteen Purāṇas and describes seventeen. It promises a thousand names and prints 1,024. It announces 112 disciples and enumerates 106.
The temptation is to read that as a sloppy text.
It is not, and the evidence against that reading is in the same volume. Twenty-odd enumerations run clean, including an eight-tier cascade where each tier adds exactly eight and explicitly inherits the tier below, checked at every rung. The book can count. It counts well, at length, under load.
Which makes the seven failures localised defects rather than ambient noise — and that is what turns an announced number into a usable instrument. A text that miscounts everywhere tells you nothing when it miscounts. A text that miscounts seven times in a hundred and sixty chapters is telling you where to look.
Every one is checkable in minutes, and the fourth column is the finding.
| Where | Announced | Delivered | Flagged? |
|---|---|---|---|
| ch 306-5(35) | "one thousand names" | 1,024 numbered entries | ✅ note 12 |
| ch 309-5(38) | "Four methods are described for obtaining emancipation" → "Among these four, the strongest is the vrata on Shivaratri" | Śivarātri is not one of the four — they are worship of Śiva, japa of Rudra, fasting in Śiva's temple, death in Vārāṇasī | ✗ |
| ch 312-5(41) | "There are four kinds of mukti" | sārūpya, sālokya, sānnidhya, sāyujya — then "The fifth state, known as kaivalya" | ✗ |
| ch 348-6(34) | "Svayambhuva Manu had ten great-souled sons" | nine named | ✗ |
| ch 358-6(44) | eighteen Purāṇas rolled | seventeen described, with a duplicated clause where the missing one belongs | ✗ |
| ch 388-8.1(1) | "Therefore, there are fourteen kinds of vidya" — after four more have been added to a correct fourteen | "These eighteen kinds of vidya follow different paths", next sentence | ✗ |
| ch 431-8.2(9) | 28 Yogācāryas × "four shishyas" each = 112 | 106, no gaps, no duplicate indices | ✅ note 1948 |
1234567
⚑ Five of seven are unflagged in an apparatus of 2,124 notes that flags inconsistencies constantly elsewhere. ⛔ And ch 312's case is the sharpest version of that: note 1095 defines the fifth state — "Beyond sarupya, salokya, sannidhya and sayujya" — without remarking that the chapter opened by announcing four.3 The translator glossed the surplus item and did not see the surplus.
Ch 431-8.2(9) is the exception, and it is worth the space because of what the exception found.
The chapter names twenty-eight Yogācāryas — counted, 28 ✓, ending at Lakulīśvara — and states "Each of them had four shishyas." The enumerated list runs (1) to (106).
Note 1948: *"One for each dvapara yuga. 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 text does not have the numbering for shishyas. We have added this for convenience.… There should be 112 names for shishyas, but we only have 106, suggesting one shloka is missing from the text."*7
⚑ That is a new species of editorial act in this edition. Every other intervention records an inconsistency, an emendation or a difficulty. This one infers physical loss from arithmetic — and the inference is coherent, because six missing names is about one shloka's worth in a list of that density.
⭐ So one announce/deliver gap in this text has been chased to the end, and it turned out to be damage rather than carelessness.
⛔ Stated at exactly that strength and no further. This is one confirmed case. It is not a theory of the other five, and this page does not offer it as one. What it does is raise the prior: the unflagged gaps are now worth chasing rather than dismissing.
⚑ And note what else that single note does. It reclassifies a proper name as an adjective to make a count come out, and it imposes numbering the text does not have. Both are habits of this apparatus rather than one-offs.8
Ch 358-6(44) is the one that a reader could find without counting, because the damage left a mark.
The chapter rolls the eighteen: "Brahma, Padma, Vishnu, Shiva, Bhagavata, Bhavishya, Narada, Markandeya, Agni, Brahmavaivarta, Linga, Varaha, Vamana, Kurma, Matsya, Garuda, Skanda and Brahmanda" — eighteen ✓, with Śiva in and Vāyu out.
Nandikeśvara then describes them and delivers seventeen. Kūrma is named in the roll-call and never described. And the sentence at the point where it should sit says the same thing twice:
"The story of Vamana is in Vamana, the account of Matsya in Matsya and Matsya is described in Matsya."5
⚑ The same clause, duplicated, in the exact slot Kūrma vacated. ⛔ Unflagged — the notes in that range are frame clarifications.
⚑ And the internal ordinals hold right up to the gap: "the seventh" on Mārkaṇḍeya and "the twelfth" on Varāha are both correct for the roll-call order. The descriptions track the list until Kūrma drops out, which is what makes the duplicated clause look like a repair rather than a slip of style.
The negative control is what makes the whole page mean anything, so it is worth setting out in full.
Ch 460-8.2(38) builds the siddhis as a cumulative ladder. Each tier announces a running total, enumerates exactly eight new items, and states explicitly that it inherits everything below.
| Tier | Announced total | New enumerated | Inherited |
|---|---|---|---|
| earth (paiśāca) | 8 | 8 ✓ | — |
| water | 16 | 8 ✓ | + 8 |
| fire | 24 | 8 ✓ | + 16 |
| wind | 32 | 8 ✓ | + 24 |
| space (aindra) | 40 | 8 ✓ | + 32 |
| mind (cāndramasa) | 48 | 8 ✓ | + 40 |
| Prajāpati | 56 | 8 ✓ | + 48 |
| Brahmā | 64 | 8 ✓ | + 56 |
Every tier checks.9
And the surrounding chapters pass a further dozen: five yogas ✓, ten prāṇas ✓, five yamas ✓, five niyamas ✓, eight āsanas ✓, three prāṇāyāmas ✓, ten impediments ✓, six upasargas ✓.9 Elsewhere, ch 404's genealogy resolves Dakṣa's twenty-four daughters into thirteen and eleven, names the eleven, matches them to eleven recipients, and checks on both sides.10
⚑⚑ Roughly twenty consecutive clean enumerations, including an eight-rung cumulative progression with explicit inheritance clauses. That is not a text that cannot count. It is a text that counts reliably and fails in specific places — which is the precondition for treating a failure as a signal at all.
Put the two halves together and the interesting number is not seven.
Seven gaps. Two flagged. Five not. In an apparatus that intervenes twenty-odd times on other grounds, reads inconsistencies closely, and in one case reconstructs physical damage from a shortfall of six names.
⚑ So the instrument works and the apparatus is running it at about twenty-nine per cent. ⛔ That is a rate on seven instances, which is a small denominator, and the phrasing here is deliberately about rather than a decimal.
⚑ The corpus figure is worse and it is worth carrying: with Volume 2's 108-announced-100-delivered and Volume 1's Vīrabhadra losing a thousand arms between two chapters, the corpus stands at nine announce/deliver failures across three volumes, of which the translator caught two.11
⛔ What is claimed is the detection record. [n≥3 PATTERN] on the phenomenon's recurrence. [n=1 OBSERVED] on the one diagnosis that resolved. No claim at all about what the unflagged five are — damage, authorial slip, transmission, or an artefact of translating a list.
⛑ This page would be dishonest without the failure that produced its own governing rule.
Earlier in this build, a script-verified arithmetic finding about ch 322's hell-grid was filed and then retracted. The discrepancy was entirely an artefact of an editorial note — the grid's structure is partly imposed, with the translator recording that "it is impossible to reconcile the number without being subjective."12 The finding was caught only because the endnote was read before the entry was written.
⚑ The rule that came out of it is the one every count on this page was run under: never run an arithmetic check on an enumeration without reading its endnotes first. Three of the seven above have notes attaching, and all three were read in full before the row was written.
⛔ And this is the second reason the ch 431 case matters. It is the one place where the arithmetic and the apparatus agree, which is the only configuration in which a count finding is safe on its own.
You are checking a document that states its own totals.
Treat every announced number as a checksum the author volunteered. ⚑ "There are four kinds of X" followed by a list is a claim you can test in seconds, and the author supplied both halves. Most documents contain more of these than anyone runs.
Establish the baseline before you interpret a failure. ⛔ This is the step people skip and it is the one that carries the argument. Seven gaps mean something here only because twenty other enumerations are exact. If the same document had failed a third of its counts, seven failures would be weather. Count the clean ones first; they are what makes the dirty ones evidence.
Read the annotations before you claim the discrepancy. ⚑ At least one apparent error in this volume was created by the editor and one was diagnosed by him. An unannotated gap and an annotated one are different objects, and you cannot tell them apart from the body text — which is precisely the condition an apparatus that fixes silently produces.
Look for the mark the repair left. ⚑ Ch 358's missing Purāṇa has a duplicated clause sitting in its slot. Damage is often tidier at the seam than the surrounding prose, and a sentence that says the same thing twice in the place where an item should be is worth more than the count that led you there.
And notice the detection rate, not just the errors. ⛔ Five of seven unflagged, by an editor who flags constantly, tells you something about the review process rather than about the text — and if you are relying on that review, that is the number that should concern you.
Evidence. All seven read in place on the source file and all counts re-run: ch 306-5(35) line 145ff, numbered entries counted to a maximum index of 1,024;1 ch 309-5(38) lines 191 and 193, the four methods and the "among these four" clause read in sequence;2 ch 312-5(41) line 225, with note 1095 read in full;3 ch 348-6(34) line 543, nine names counted against an announced ten;4 ch 358-6(44) lines 659–679, the roll-call, the seventeen descriptions and the duplicated clause;5 ch 388-8.1(1) line 995, the three-sentence self-correcting seam;6 ch 431-8.2(9) lines 1357–1362, indices counted, with note 1948 read in full.7 The negative control at ch 460-8.2(38), all eight tiers re-added.9
⛔ Cited in the Saṃhitā-internal form throughout, and note numbers are given with line references — both hazards are undisclosed in this edition, and note "12" is two different notes.13
Tension 1 — the seven are not one kind of failure. ⚑ Ch 312's fifth mukti and ch 309's fifth vrata look like a doctrine outgrowing a fixed formula. Ch 431's is diagnosed as physical loss. Ch 388's is a three-sentence seam that corrects itself and then does not. Grouping them as "announce/deliver" is a description of the detection method, not a claim that they share a cause.
Tension 2 — two of the seven may be translation artefacts. ⛔ The thousand-names count is the translator's own numbering, and his note says as much: "There is always a problem distinguishing an adjective from a proper name, probably explaining why the actual number exceeds 1000."1 ⚑ So ch 306's gap may be a property of rendering Sanskrit compounds into English list items rather than of the text. Recorded in the table because the note discloses it; not counted as evidence about the Sanskrit.
Tension 3 — the baseline is measured on chapters that were read. ⚑ The twenty clean enumerations are real and they are not a random sample; they cluster in the yoga and genealogy material. A hostile reading is that the counting is clean where the material is procedural and fails where it is doctrinal — which would be a finding rather than a control. [n=1 OBSERVED] on that suggestion; not tested.
Open question. The one gap that was chased turned out to be a missing shloka. Would the other five resolve the same way against a Sanskrit text — and is there any of the five where the surrounding prose carries a mark like ch 358's duplicated clause?
Against The Apparatus Is the Translator's: this page supplies that one's arithmetic wing. ⚑ Note 1948 does three things in one note — reclassifies a name as an adjective to fix a count, imposes numbering the text lacks, and infers a lost shloka. All three are that page's territory and they arrive together, which is the closest this edition comes to showing its whole method in a single place.
Against The Saṃhitā That Grew: ch 431 gives that page its first concrete mechanism. Its finding is that the printed Saṃhitās diverge from the shloka budget the text publishes for itself, with Vidyeśvara running short. ⚑ Here is a demonstrated instance of a shloka missing from this edition, found by counting. ⛔ It does not explain Vidyeśvara — this loss is in the Vāyavīya, which runs over budget. What it establishes is that shlokas are demonstrably missing and that arithmetic finds them.
Against Two Hells, Two Shadows, Two Tattva Counts: the two failure modes are different and should not be conflated. ⚑ That page's material is two coherent systems that disagree; this page's is one system that does not deliver what it announced. Two authors versus one gap. ⛔ The hell-grid retraction sits between them as a warning: an apparent count error there was an editorial artefact of exactly the kind that would have merged the two categories wrongly.
Against Only What a Character Asks: the explanation for why none of the seven is ever noticed inside the text. ⚑ Detection here runs through a listener raising an objection, and enumerations have no listener — so the announced total and the delivered list can disagree in one paragraph with nothing available to fire.
Against The Five Lost Saṃhitās: the same arithmetic instrument on the largest object it has. ⚑ That page holds the text's own account of what was cut from it; this one holds the smaller losses that nobody announced.
Disguising Numbers and Direction — that page's claim is that a trail is a confession, and the confession can be forged: leave a footprint that reports fewer men, a lighter body, the wrong direction.
The Purāṇic material is the same object in the configuration where nobody was writing to be read that way, and the contrast produces the rule.
An announced total is a footprint. "There are four kinds of mukti" commits the author to something a reader can check against the ground. ⚑ And in this text those commitments were left where they fell — because the announcements were doctrinal furniture, not claims anyone expected to be audited. Which is exactly why the instrument works: the checksum is trustworthy in proportion to how little anyone cared about it.
Now invert it. In any setting where a stated total is known to be the thing that gets checked — a headcount, an inventory, a set of accounts, a citation list — the announced number is the first thing an operator adjusts, and adjusting it is cheaper than adjusting the underlying list. ⛔ The forged version of ch 312 does not say five and announce four. It says four and prints four, and the fifth item is simply gone.
⚑ So the diagnostic rule that falls out is uncomfortable and useful: a self-supplied checksum detects error, not deception, and it detects error best where nobody was defending anything. The moment a document is written under audit, the announced total stops being evidence about the contents and becomes evidence about what the author wanted the contents to look like. Where the two disagree, you have found carelessness. Where they agree, you have found nothing — and this text's twenty clean enumerations mean something only because nothing in it was under scrutiny.
Running back the other way, the corpus supplies the counter-tracking detail the influence framing tends to miss. Ch 358's missing Purāṇa left a duplicated clause behind. A footprint can be forged, but the forging is itself an act that leaves marks at the seam — and the repair is often more legible than the thing repaired.
Sharpest implication. "There are four kinds of mukti" — sārūpya, sālokya, sānnidhya, sāyujya — and then, in the same speech, "The fifth state, known as kaivalya." The endnote defines the fifth and says nothing about the four. "Four methods are described for obtaining emancipation", then "Among these four, the strongest is the vrata on Shivaratri", which is not one of them. "Therefore, there are fourteen kinds of vidya" — after four have been added to a correct fourteen — and one sentence later "These eighteen kinds of vidya follow different paths." Eighteen Purāṇas rolled and seventeen described, with "the account of Matsya in Matsya and Matsya is described in Matsya" sitting in the slot Kūrma vacated. Ten sons announced and nine named. A thousand names promised and 1,024 printed. Seven gaps. Two flagged. ⚑⚑ And the one the translator chased resolved into something better than an error: "There should be 112 names for shishyas, but we only have 106, suggesting one shloka is missing from the text." Physical damage, inferred from arithmetic, and correct. Which is only legible because the same volume runs an eight-tier siddhi cascade — 8, 16, 24, 32, 40, 48, 56, 64, each tier adding exactly eight and naming what it inherits — that checks at every rung. The book can count. It is the twenty clean enumerations that make the seven dirty ones evidence, and the detection rate — five of seven unread by an apparatus that intervenes constantly — that makes them a finding about the edition rather than about the text.
Generative questions.