You write down a reference. Chapter 388. It is the chapter where the text publishes the size of every one of its own sections — a table, twelve rows, summing to exactly a hundred thousand. Good, solid, citable.
Six weeks later somebody follows the reference and lands in a completely different chapter, in a different Saṃhitā, about the succession of teachers.
Both are chapter 388. One is the last chapter of the Kailāsa Saṃhitā and one is the first chapter of the Vāyavīya. They are printed thirteen lines apart and the number does not distinguish them.
There is a second pair, and a gap, and the same problem in the endnotes. None of it is stated anywhere in the edition, in any of its 2,124 notes.
The duplication was not noticed by reading. It came out of a coverage check — a scripted census of every chapter heading in the body, run to prove that no chapter had been silently skipped.1
160 headings, across the range 305–463, which is 159 slots.
362-6(48) (Manifestation of Sarasvati) and 362-6(49) (Uma's Manifestation).388-7(23), the Kailāsa's closing chapter, and 388-8.1(1), the Vāyavīya's opening one.158 distinct numbers plus 2 duplicates = 160 headings. The arithmetic closes exactly.1
⚑ And the totals confirm the diagnosis rather than contradicting it. The Introduction states 464 chapters; Volumes 1 and 2 cover 1–304; this volume has 160; 304 + 160 = 464 ✓. The highest number reached is 463 — short by exactly one, which is what two duplicates and one gap produce. The count is right. Only the numbering is broken.
This matters, and it is stated plainly in the front matter:
"A word about the numbering we have followed for the chapters, such as Chapter 423-8.2(1). The first number is a consecutive numbering of chapters of the entire Shiva Purana."2
The running number is an editorial device. The text does not carry it. What the text carries is the second element — the Saṃhitā-internal code — and that is exact throughout. Umā runs 6(1) to 6(51), which is fifty-one, matching the Introduction's own statement. Kailāsa runs 7(1) to 7(23), twenty-three, likewise.1
⚑ So the failure is confined to the layer the translator added, and it is invisible from inside either Saṃhitā. A copyist working through the Kailāsa would see 7(23) end and 8.1(1) begin, correctly. Only somebody counting across the whole book from chapter one would find the collision — which is exactly what nobody has to do to read it, and exactly what a coverage census does.
The same hazard runs a second time, in the apparatus, on a different axis.
Endnote numbers restart per section. "Note 12" is two different notes in this volume: the Hazra bibliography note attached to the Introduction, and the note on the thousand-name numbering attached to the Koṭi Rudra Saṃhitā — "The text naturally has no numbering. We have added the numbering."3
A bare note number does not identify a note in this edition.
⚑ This is worse than the chapter case, because chapter numbers at least appear in a heading a reader can see. Note numbers appear as small raised digits in a body four thousand lines away from their targets. There is no signal at either end that the numbering has reset, and a vault citing "note 12" has cited nothing.
⛔ Both hazards are the translator's own apparatus, and both are unflagged in all 2,124 notes.4
The demonstration is not hypothetical, and it happened to the build that produced this page.
Two separate findings were logged against "ch 388." One was the shloka budget — the twelve-section table summing to 100,000, and the Umā Saṃhitā running forty-six per cent over its published weight. The other was the guru-succession loop: "Earlier, Vyasa… spoke about this to lords among sages in Naimisharanya. Therefore, he is the first guru", in a chain that runs Vyāsa → Vāmadeva → Sanatkumāra → Vyāsa.5
They are in different chapters, in different Saṃhitās, written by whatever hands wrote each, and the build had filed them under one number. The budget is 388-8.1(1), the Vāyavīya Pūrva's opening chapter. The guru-loop is 388-7(23), the Kailāsa's closing one. Nothing in the notes, the headings or the running text distinguishes them.
⚑ The error was caught by the coverage census and not by anything in the reading, which is the point of the case. A citation collision does not announce itself; it produces two claims that agree about their address and cannot both be checked. The correction cost an hour. It would have cost far more downstream, because the two chapters belong to the two Saṃhitās this build has separately argued are compositionally independent — so a claim attributed to the wrong one would have been evidence in an argument about which section says what.
Worth stating precisely, because the temptation is to over-generalise from a real hazard.
What breaks. Any citation of the form "ch 388" or "note 12." Any argument in which the location of a passage is doing work — which section it sits in, which narrator speaks it, which Saṃhitā's doctrine it belongs to.
What does not break. The text. The chapter count. The Saṃhitā-internal numbering. Retrieval, mostly — a reader who follows a bad reference will usually notice that the chapter is about something else. And the substance of any claim that quotes its passage: a quotation carries its own identity in a way a number does not.
⚑ The practical upshot is a citation form rather than an alarm. Cite this edition as 388-8.1(1), and cite every endnote with its line reference. That is the whole remedy, and it costs nothing to apply.
One more thing about this numbering, and it is the oldest layer of it.
The Vāyavīya is numbered 8 in this edition's codes — 8.1, 8.2 — and that is the scheme this vault has used across three volumes and some three hundred pages. Debroy explains where the eight comes from:
"In the Sanskrit text of the Shiva Purana used by us, seven chapters and 373 shlokas from the Skanda Purana have also been included, since they concern the Shiva Purana… (With the Skanda Purana segment included, this numbering extends to eight, not seven.)"6
The abridged Purāṇa has seven Saṃhitās. The eighth number is an artefact of an interpolation, and the vault's own citation scheme encodes it.
⚑ Harmless for retrieval, and worth recording anyway — not because anything is wrong, but because it is a second case of the same species: a numbering the vault treats as belonging to the text, which belongs to the edition.
You are building on a body of material with identifiers you did not create, and you want to know whether the identifiers identify.
Test whether the key is unique before you use it as a key. Not by reading — by counting. Extract every identifier in the corpus, count them, count the distinct ones, and compare. Here that is one line of script and it took forty seconds; it found a collision that a hundred and sixty chapters of careful reading had walked straight past twice. A duplicate identifier is invisible to a reader and obvious to a counter, because reading follows the reference forward and counting looks at the set.
Ask whose numbering it is. Every corpus has layers of identifier: the ones the object came with, and the ones somebody added to make it referenceable. They fail differently. The ones the object came with fail by corruption and are usually flagged. The ones an editor added fail by arithmetic, silently, because nobody's job was to check them and there is no authority to check them against.
Prefer the compound key over the convenient one. 388-8.1(1) is uglier than 388 and it is unique. When both a local and a global identifier exist, the global one is almost always the one that collides — it is doing more work across more material with fewer constraints — and it is always the one people quote, because it is shorter.
And when you find a collision, go back and re-check the claims that used it, not just the citation. The citation is the cheap part. What matters is whether anything you concluded depended on the two things being the same thing.
Evidence. Scripted census of every chapter heading in the body, lines 129–1599: 160 headings, 158 distinct numbers, the two duplicates and the gap located to their lines, and per-chapter word counts summing to within 0.5% of the independently measured body total.1 The Introduction's statement of the numbering scheme read in place.2 Both notes numbered 12 read at their own lines.3 The complete endnote census confirming neither hazard is mentioned anywhere.4
Tension 1 — this is an artefact, and artefacts are not findings about the text. ⚑ The defence is that the vault's exposure is to the edition, not to the Sanskrit: three hundred pages cite chapter numbers, and the chapter numbers are the edition's. A fact about the instrument is a fact about every measurement taken with it.
Tension 2 — the missing 365 has no explanation here. A skipped number in a running sequence could be a chapter dropped in composition, a numbering slip, or a chapter merged into a neighbour. [n=1 OBSERVED] and nothing more; this page does not know which, and the arithmetic closing at 160 headings means nothing is missing from the book, only from the sequence.
Open question. Do Volumes 1 and 2 carry the same defect? The census is trivially runnable on both files and has not been run. If the running number collides there too, every bare "ch N" citation in three hundred vault pages is suspect rather than merely unwise.
Against The Apparatus Is the Translator's: the same thesis in its least controversial form. That page's twenty-three editorial acts each require somebody to accept that a disclosed change is still a change. This one requires nothing — an identifier that names two objects is broken by any standard, the translator would say so himself, and it is the cleanest available demonstration that the apparatus is a made thing with its own error modes.
Against The Saṃhitā That Grew: the collision lands directly on that page's evidence. Its central passage is 388-8.1(1), and the other 388 is in the very Saṃhitā the shloka table is used to distinguish from the Umā. ⚑ A finding about which section material sits in was filed under a number shared by two sections. No conclusion changed; the address did.
Against The Pāśupata Charter: both pages are cases of a number being trusted because it was printed. There the trusted number was an endnote's multiplication standing in for a count; here it is a running index standing in for an identity. In both cases the vault took an editorial convenience for a property of the text.
Zero handshakes, and the reason is worth one line: this is a defect in a citation apparatus, and the general lesson — check key uniqueness before relying on a key — is already fully carried by the Implementation Workflow above without needing a second domain to license it. Manufacturing a cross-domain connection here would produce a decorative one.
Sharpest implication. A coverage census of all 160 chapter headings shows that this edition's running chapter number, which is the translator's own device and not the text's, fails three times in one volume: 362 names two chapters, 388 names two — one the Kailāsa's last and one the Vāyavīya's first, printed thirteen lines apart — and 365 names none. The arithmetic closes exactly at 160 headings and the total chapter count is right; only the index is broken, short by exactly one, which is what two duplicates and a gap produce. The same hazard runs in the apparatus, where endnote numbers restart per section and "note 12" is two notes. Neither fact is stated anywhere in the edition, in any of its 2,124 notes. The Saṃhitā-internal numbering is exact throughout, so the remedy is a citation form and not an alarm: 388-8.1(1), never "ch 388," and every endnote with its line reference. And the build that found it had already filed two unrelated findings under the same bare number and had to correct both.
Generative questions.