Eight Arguments and a Pretty Face — eight authorities argue correct theology at one woman and all eight fail; a cosmetic change succeeds in one sentence.
Tāraka Abandons Heaven When Asked — the volume's three successful persuasions, none of which is an argument: renegotiated terms, a devotion threshold, a change of appearance.
The Satī Khaṇḍa's spine — five accurate forecasts across five ranks, and not one altered an outcome.
Against all of that sits the vault's own founding assumption, visible in how it is built: that a well-sourced, well-argued case is what moves people, and that the job is assembling one.
A text of 150 chapters, containing dozens of persuasion attempts, in which argument has a success rate of zero and three non-arguments succeed.
That is not a moral about rhetoric. It is a structural claim the volume makes without ever stating it, and it is unusually clean because the sample is large and the failures are explicit. Chapter 139 is close to a controlled experiment: same target, same objective, eight sequential attempts by escalating authorities, all recorded, all failing — and Viṣṇu's is marked as forming understanding and softening her mind while changing nothing.
Understanding was achieved and it did not matter. That sentence is the collision.
The vault holds a large behavioural corpus about influence, and almost all of it agrees with the Purāṇa rather than with the vault's own practice. The Hughes material, the Brunson material, the Greene material, the premium-branding material — all of them locate the lever somewhere other than the merits. And the vault, having ingested all of it, continues to operate as though the merits were the lever.
The three successful instruments in this volume are not three; they may be one, and the vault's own influence corpus may be describing the same single thing under a dozen names.
Look at what the three have in common:
In all three, nothing is added to anyone's beliefs. Something in the world is rearranged so that the target's existing position produces a different output.
That is a different operation from persuasion, and it may be the only one that works. It also explains why Menā's capitulation contains no retraction: she didn't change her mind, she changed her circumstances. Her position — I'll give her if his form is beautiful — was satisfied, not defeated.
If that generalises, the practical form is stark: you do not change what someone believes; you change what their belief implies.
A second long text with enough persuasion attempts to run the same count. One volume of one Purāṇa is a single case, and Purāṇic narrative may simply not be interested in dramatising successful argument. The vault holds candidates — the Greene corpora, the Dimsdale historical cases, the Olivelle and Lahiri Ashoka material. The Ashoka edicts are the sharpest test available, because they are an actual attempt to persuade at scale and their effectiveness is genuinely contested.
Evidence from the vault's own record. META/what-worked.md tracks which ideas reached output. If the ideas that landed were re-descriptions rather than arguments, that's confirming evidence from inside the practice.
A counter-case: any documented instance of someone changing behaviour because they were shown they were wrong. These presumably exist. The interesting question is their base rate against the alternative.
And the deflationary reading: narrative requires obstacles, and a story where the first good argument works is a story with no second act. The volume's zero success rate for argument may be a property of storytelling rather than a claim about people. That reading is live and this collision does not defeat it — though it is worth noting that the volume happily lets other things work immediately, including a demon vacating heaven in four lines.
The collision asked for a second body of material with enough persuasion attempts to run the count. Volume 2 of the same Purāṇa supplied one — 154 chapters, ch 151–304 — and the count was run, then re-run in sequence after the build's diagnostic showed that sequence-level claims from this corpus fail without a fresh read.
The bare claim does not survive as stated. The mechanism underneath it does, and it is sharper.
Two arguments win in Volume 2, not zero. Gaṇeśa at ch 169 and Ghuṣmā at ch 304 — and the second closes the book. [n=2 SUGGESTIVE].
But neither wins on its content. Both win by converting into a citation the opponent has already endorsed — hostaging the Vedas' truth-value in one case and Śiva's own prior teaching in the other. Ch 304 supplies an internal control that makes this close to decisive: the same speaker's moral appeal fails and her citation wins four lines apart.
And the losing side is never beaten either, which is the finding the collision did not anticipate. Five serious arguments in Volume 2 go entirely unanswered — Tāraka's largely accurate indictment of Viṣṇu and Indra at ch 159, Arihat's ethics at ch 175, the messenger's reductio at ch 205, Śiva's own objection at ch 176, and Gaṇeśa's counsel at ch 296. Four instruments produce closure without touching content: charge the speech to the speaker, move the category, relocate the agency, invoke a prior obligation. None requires knowing whether the argument was good.
⚑ So the corrected form of the collision is not "argument never wins." It is: content is not the operative variable in either direction. Arguments win when they become citations and lose when their speaker is settled, and in both cases the quality of the reasoning is doing no work. That is a stronger claim than the original and it is falsifiable the same way.
One correction to the record. The count was briefly three winners; ch 301 was removed after a sequential re-read showed Śiva concedes before Pārvatī argues, "bound by the boon he had conferred." The earlier figure should not be cited forward.
Both halves are now carried at page level — What Makes an Argument Win and Closure Without Refutation — which means this collision has done its job and the outstanding test is now the external one it named: the Ashoka edicts, and whether a real attempt to persuade at scale behaves the same way.
Filed from the Shiva Purana Volume 1 DEEP INGEST tension log, 2026-08-10. This one indicts the vault's own method and should be read before the next COLLISION session, per the feedback loop in META/what-worked.md.