Volume 2 puts its best sentences in the mouths of people it has already discredited.
Eight superlative dharma-claims run through the volume — X is the supreme dharma — and four of them are spoken by a character who is actively deceiving somebody at that moment. The two most humane are Arihat's, and Arihat is a manufactured heretic, built to order by Viṣṇu with a written brief:
"All those with bodies are equal"… "there is no dharma that is equal to compassion towards beings"… "How can four sons born simultaneously from the same body have different varnas? Hence, the division into varnas and sub-varnas is not convincing."1
Nobody answers a word of it. Not in that chapter, not in the fifteen that follow. The equality argument stands. The varṇa objection stands. Three cities get burned instead.
And the same volume, at ch 159, prints Tāraka's itemised indictment of Viṣṇu and Indra — Bali bound by deception, Madhu and Kaitabha beheaded by crookedness, Gautama's wife violated, Vṛtra killed — most of which the corpus narrates elsewhere approvingly. The only reply is the narrator observing that making the accusation destroyed the accuser's merit.2
The other side of the tension is this vault's own practice.
Every claim on every page I built carries an inline footnote. This is a rule. No claim enters ARCHIVES without a source stub, a chapter, a line number, and increasingly — after endnote 511 turned out to say something other than the notes recorded — a verbatim quote read at its own line.
The volume demonstrates, at length and without noticing, that attribution is a disposal mechanism. Assign a sentence to a discredited speaker and nobody has to weigh it. It is more efficient than refutation, it costs nothing, and it works on true sentences exactly as well as false ones — which is why the author could afford to hand his straw man the best arguments in the book.
And I have spent this entire build constructing, page by page, an apparatus that makes every sentence in this vault instantly attributable.
That is not an accusation of anything. Citation is right and the vault's insistence on it has repeatedly saved this build from itself. But the two facts sit together badly enough to be worth writing down:
⚑ A perfectly-sourced corpus is also a perfectly-dismissable one. Every claim in here is one click from who said that, and the volume just spent 154 chapters showing what people do with that click.
The stronger the attribution layer, the cheaper speaker-based dismissal becomes — and the two grow together, because the same infrastructure serves both.
Three things follow, and the third is the one I'd want tested.
First, a practical asymmetry. Attribution helps you check a claim and helps someone else skip it, and skipping is cheaper than checking. So a well-cited corpus does not distribute its benefits evenly: it rewards the careful reader a little and the dismissive reader a lot.
Second, this vault has a specific exposure. A large share of what's in ARCHIVES is sourced to practitioner transcripts, popular-source authors, influence manuals and devotional texts — categories the vault itself tags with warnings ([POPULAR SOURCE], [PARAPHRASED], the red-flag list). Those tags exist to protect against overclaiming. They also constitute a ready-made discrediting index. A reader who wanted to discard a finding never has to engage it; they can find the tag.
Third, and this is the live one: what would a claim look like that had been separated from its speaker on purpose?
The volume shows two arguments actually winning, and both win the same way — by converting into a citation the opponent has already endorsed. Gaṇeśa hostages the Vedas' truth-value; Ghuṣmā hostages Śiva's own prior teaching. Neither wins on content. Both win by relocating the claim into a mouth the opponent cannot discredit without discrediting himself.
⚑ Which suggests the counter-move is not better sourcing. It's finding the version of your claim that your reader has already agreed to — and that is a completely different research operation from the one this vault is built for. The vault indexes what a source said. It has no index of what the reader already holds.
A case where a well-sourced claim was dismissed on the source rather than engaged. This should be trivially findable in public argument and I have not gone looking. If the base rate is low, the collision deflates.
A test of the counter-move. Take one finding from this build — say, content is not the operative variable when arguments win or lose here — and try to restate it entirely in premises a skeptical reader has already granted. If that's possible, the move generalises. If it collapses into hand-waving, it was a feature of two Purāṇic scenes.
Evidence from META/what-worked.md. The vault already tracks which ideas reached output. If the ideas that landed were ones the reader already half-believed rather than ones that were best sourced, that is confirming evidence from inside the practice — and it is checkable today without any new material.
And the deflationary reading, which is strong. Purāṇic narrative discredits speakers because that's how narrative works: characters have standing, and standing is a storytelling tool. Extending it to how arguments get evaluated by actual readers may be importing a plot device into epistemology. ⚑ That reading is live and this collision does not defeat it — though it's worth noting that the volume's own author appears to have believed it, since he gave a figure he wanted destroyed the two best sentences in the book and never bothered writing a rebuttal.
Filed from the Shiva Purana Volume 2 DEEP INGEST tension log, 2026-08-11. Pairs with Argument Never Wins Here, which this sharpens rather than repeats: that one asks whether argument wins, this one asks what the vault's own citation architecture does to the answer.