Here is a thing you can do in about forty seconds, with no Sanskrit and no scholarship.
Open the Twelfth Skandha at its first chapter. It is a list of kings — dynasty after dynasty, father to son, all in the future tense because the man reciting it is describing what has not happened yet. Somewhere in the middle you get this:
"The brahmana will instate Chandragupta in the kingdom. Chandragupta's son will be Varisara and Varisara's son will be Ashokavardhana… Shatadhanva's son will be Brihadratha. O extender of the Kuru lineage! During kali yuga, the ten kings of the Mourya dynasty will enjoy the earth for more than one hundred and thirty-seven years."1
Count the names. Chandragupta, Varisara, Ashokavardhana, Suyaśa, Saṅgata, Śāliśuka, Somaśarmā, Śatadhanvā, Bṛhadratha.
Nine.
⚑ The sentence says ten and the list holds nine, and the two are three lines apart.
⚠ Everything on this page is bounded to this EPUB as ingested. No claim is made about the Sanskrit, about Debroy's manuscript, or about any print edition.
Keep reading. Two dynasties later:
"Thereafter, there will be Agnimitra and Agnimitra's son will be Sujyeshtha… Thereafter, there will be Bhagavata and Bhagavata's son will be Devabhuti. The ten kings of the Shunga dynasty will enjoy the earth for more than one hundred years."1
Count again. Agnimitra, Sujyeṣṭha, Vasumitra, Bhadraka, Pulinda, Ghoṣa, Vajramitra, Bhāgavata, Devabhūti. Nine.
⚑ Same shortfall, same magnitude, same page, one dynasty apart — and this one has a footnote:
"The name of Agnimitra's father, Pushyamitra, is missing in the text. Hence, the text mentions ten kings of the Shunga dynasty, but only names nine."2
⚠ The Mauryas get nothing. Not a note, not a bracket, not a query.
⚑ Two identical arithmetic failures, adjacent, one flagged and one silent. ⚠ And the flag is not stingy — it is a good note, naming the missing king and explaining his absence. The apparatus is capable of catching this. It caught it once.
The Āndhra list is longer and the gap is bigger.
Twenty-three men are named, Bali through Lomadhi, plus one unnumbered plural — "his sons will be the Bahus." Then:
"O descendant of the Kuru lineage! These thirty kings will enjoy the earth for four hundred and fifty-six years."1
⚑ And here the note does appear, and it does something the other one did not.
"The names listed actually fall short of thirty. But thirty is the number also cited in other Puranas."3
⚠ That second sentence changes the finding and must be carried, not dropped. The shortfall is not local damage. Other Purāṇas carry the same figure. ⚑ Which means the number thirty has been transmitted independently of the list it is supposed to count, and has survived in that condition across texts.
⚠ A number that outlives its own list is not functioning as a count.
The very last chapter does it with a total you can add up on paper.
Sūta lists eighteen Purāṇas with a verse-count for each. Brahmā ten thousand. Padma fifty thousand. Viṣṇu twenty-three thousand. Śiva twenty-four. Bhāgavata eighteen. Nārada twenty-five. Mārkaṇḍeya nine. Agni fifteen thousand four hundred. And so on down to Brahmāṇḍa at twelve thousand. Then:
"Thus, the total number of verses in all the Puranas is said to be four hundred thousand."4
Add the eighteen figures he just gave. 396,000.
⚑ A four-thousand-verse gap, in a paragraph whose only content is the numbers themselves, arithmetically checkable in under a minute. ⚠ No note.
⚑ And two sentences later he restates one of the eighteen for no reason — "Of these, Bhagavata is said to have eighteen thousand" — a figure he gave three sentences earlier. The paragraph is willing to repeat a number and unwilling to add them.
The Veda-branch genealogy in 12(6) does the same thing at small scale, which is where it is easiest to see.
"Shakalya's disciple, the sage Jatukarna, divided his samhita into three parts and adding nirukta, gave it to Balaka, Paija, Vaitala and Viraja."5
Three parts. Four names. ⚠ Unflagged.
⚑ And the same chapter, two paragraphs on, sets up an arithmetic that does close — but only if you do it yourself. Sukarmā divides the Sāma Veda into one thousand saṃhitās. Pauṣyañji's five disciples take a hundred each: five hundred. Hiraṇyanābha's disciple Kṛta takes twenty-four. "The rest were taught by Avantya."5
Four hundred and seventy-six, if you want it. ⚑ The text does not want it. ⚠ The one place the numbers reconcile is the one place no total is announced.
The sharpest instance is in 12(11), because the correct number is stated three separate ways before it fails.
Śaunaka asks the question using it. "Shuka spoke to him about the seven different categories that reside in each of the solar months."6
Sūta then delivers twelve months, seven names each. Dhātā, Kṛtasthalī, Heti, Vāsuki, Rathakṛt, Pulastya, Tumburu — and eleven more sets built the same way.
And the apparatus confirms it. *"Dhata is the form of Surya, Kritasthali is the apsara, Heti is the rakshasa, Vasuki is the naga, Rathakrit is the yaksha, Pulastya is the rishi and Tumburu is the gandharava. For the other eleven months, this exact order is followed."*7 Seven roles, named.
Then, immediately after the twelfth list:
"Through each of the twelve months, the divinity travels along with six sets of associates."6
⚑ Six. ⚠ Against the questioner's seven, the body's twelve-times-seven, and the note's seven roles. ⚑ Unflagged.
You are reading something long and it gives you a number. Stop and count. It takes forty seconds and it is the single cheapest audit available to a reader, and almost nobody performs it.
You find a gap. Now the interesting part, which is not the gap. Ask who else has already had the chance to find it — an editor, a translator, a commentator, four hundred years of scribes — and what it means that they did not, or did and said nothing.
You are the one producing the list. Write the number last, from the list, not first, from your idea of how long the list is going to be. The number written first is a plan; the number written last is a count. Nearly all of these failures are a plan that survived into print.
Somebody hands you a total. Look at whether the addends are on the same page. If they are, add them — a total sitting beside its own components is not asking to be trusted, it is asking to be checked, and it has just told you nobody has.
Six failures in one Skandha, and the interesting question is not how did this happen. Scribal transmission over two thousand years is a sufficient answer and a boring one.
⚑ The interesting question is why nobody minded.
Because these are not hard. Nine names under a heading that says ten. Eighteen figures that fall four thousand short of their own stated sum. Twelve lists of seven followed by the word six. No manuscript comparison, no philology, no second text — the disproof is printed beside the claim in every case.
⚠ And it survived anyway. Copied, recited, memorised, commented on, printed, translated, and reprinted.
⚑ The only reading that fits all six is that the number is not a claim about quantity. Ten kings is doing what once upon a time does. It marks the passage as a dynastic list, the way four hundred thousand marks a corpus as vast and thirty marks a dynasty as long. A reader who counts has misunderstood the sentence's job — which is why, for two millennia, essentially nobody did.
⚑ And the Āndhra note is the evidence for that reading rather than a complication of it. Thirty travels between Purāṇas while the list it heads does not. ⚠ A quantity that migrates without its contents is a label.
The strongest support is the 400,000 total, because it is the only instance with no possible transmission excuse. ⚑ The eighteen addends and the sum are in one paragraph. Damage that removed a king could not produce a four-thousand-verse gap without also removing a whole named Purāṇa, and none is missing.
A second support is the 12(11) case, where three independent statements of seven — question, body, and note — precede the word six. ⚑ The failure is not in a list. It is in the summary of a list the reader has just been given twelve times.
The tension, preserved. ⚠ The label reading does not cover everything. If the numbers are genre markers, why does the apparatus flag two of them? ⚑ Debroy treats the Śuṅga and Āndhra figures as errors requiring explanation, which means at least one modern reader takes them as counts. The page does not resolve this: the same figure can be a marker inside the tradition and an error outside it, and both readers are in this volume — one in the body, one in the notes.
A second tension. ⚠ Selective flagging has an available explanation and it is not flattering to the finding. The two flagged cases are the two where an external king-list exists to check against. ⚑ On that reading the apparatus is not auditing the text at all — it is reconciling the text with history, and the internally checkable failures go unflagged because nothing outside the book contradicts them. [n=2 OBSERVED] — two flags is not enough to characterise a method, and this page does not.
Open question. ⚑ Do Volumes 1 and 2 carry the same signature at the same rate? ⚠ Unchecked. Not found in the range read — and until it is checked, nothing here is a claim about Debroy's edition as a whole.
Set beside A Guṇa Is the Outcome of Ten Things, and Nine Are Listed, this is the same failure in doctrine rather than dynasty. ⚑ There the shortfall sits inside a teaching Kṛṣṇa is delivering to Uddhava about how conduct is produced — so the pattern is not confined to the genealogical material, where a reader might excuse it as archive-keeping.
Set beside Three of the Twenty-Four Are Named and Never Taught, the shape gets stranger, because there the count fails in both directions at once — twenty-four announced, three of them never taught, and the total then exceeded by a mountain, a tree and a body. ⚑ A list can be short and long simultaneously and nobody notices either. ⚠ With that page, 11(13), and the six instances here, the finding is [n≥3 PATTERN] on this volume's counts failing internally — and it is stated at that strength and no higher.
Set beside The Text Argues With Itself, the contrast is what makes this page worth having. ⚑ At 10(77) the narrator interrupts his own story to attack the verse he has just delivered — "they do not remember what they themselves have said." The text is capable of catching itself, out loud, mid-sentence. ⚠ It does that for a doctrinal problem about whether Kṛṣṇa can grieve, and never once for a number.
Hallucination and Confidence in Language Models — a system trained to produce answer-shaped output produces it regardless of whether it has the answer, because the objective rewards something over nothing and never rewards abstention. The false material arrives well-formed, confident, and specific.
⚑ A dynastic list has the same shape pressure. The form comes with a total attached; ten kings is what the sentence expects, and supplying it costs nothing at the moment of composition. The number is answer-shaped output.
⚑ What neither field gives alone is the survival question. The machine case is diagnosed as a defect and measured and worked on. ⚠ These numbers were not corrected because they were never treated as answers. They passed through two thousand years of people who could count, in a tradition with an obsessive relationship to precise enumeration, and were reproduced unchanged. A hallucination that nobody bothers to fix, in a culture that fixes things, is not a hallucination — it has been reclassified by its readers into something that cannot be wrong.
⚑ That reclassification is the mechanism worth having, and it inverts the engineering intuition: confident false specifics are a problem only where readers are treating specifics as claims. Change the reading contract and the same output stops being an error.
Messenger Effect & Source Credibility — identical content persuades differently depending on who is understood to be delivering it; the source does work the message does not. 🚩 REPLICATION RISK — Hovland & Weiss is 1951 and pre-dates the replication era.
⚑ The effect predicts belief without checking, and that is nearly what happens here — but not quite, and the gap is the point. The messenger effect describes deference to an authority you cannot verify. ⚠ Here the verification material is in your hands. The nine names are printed above the word ten. The eighteen figures are printed above their own wrong total.
⚑ So this is a harder case than credibility-transfer covers: not trust in place of evidence, but evidence declined while held. The reader is not failing to check. The reader has understood that this is not the kind of sentence you check — which is a different cognitive act, and the one that actually preserved these figures.
Sharpest implication. ⚑ The Twelfth Skandha fails its own arithmetic six times in twelve chapters, and every failure is checkable against material on the same page. Nine Mauryas under a heading that says ten. Nine Śuṅgas the same. Twenty-three Āndhras under thirty. Eighteen Purāṇa figures that come to 396,000 under a stated 400,000. Three parts to four recipients. Twelve lists of seven, summarised as six. ⚠ Two are flagged; four are not; all six survived transmission intact. ⚑ *The only reading that covers all six is that these numbers were never doing arithmetic — they are genre markers, and the note conceding that thirty travels between Purāṇas without its list is the evidence.
Generative questions.