Auto-triggered at the close of the Volume 2 DEEP INGEST, 2026-08-11. Four research lines that are live, ordered by how close each is to closing. Every open item names the range that would close it.
State: strongest line in the build. Two pages, one collision, [n=2 SUGGESTIVE] on the winning half, [n=5 PATTERN] on the losing half.
The claim: in this corpus, arguments do not win or lose on their content in either direction.
The losing half is settled inside the volume. Five serious arguments go unanswered — Tāraka's indictment (159), Arihat's ethics (175), the messenger's reductio (205), Śiva's own objection (176), Gaṇeśa's counsel (296). Four instruments produce closure without touching content, and none of them requires knowing whether the argument was good.
The winning half is thinner and better. Two arguments win — Gaṇeśa (169) and Ghuṣmā (304) — and both work by converting into a citation the opponent has already endorsed. Ch 304 is close to a control: the same speaker's moral appeal fails and her citation wins four lines apart.
What would close it. A second corpus with enough persuasion attempts to run the count. The collision names the Aśoka edicts as the sharpest available test, because they are a real attempt to persuade at scale and their effectiveness is genuinely contested. The vault holds Olivelle and Lahiri material.
What would break it. Any instance in this corpus of an argument that wins on its merits and is acknowledged to have done so. Not found in ch 151–304.
State: two independent routes to the same conclusion, arrived at from different evidence. This is the build's most robust structural finding.
The claim: the corpus operates a working liability system and it is not a moral one.
Route one — disclosure and authorisation. Liability attaches to what was undisclosed and what was unauthorised, never to what was documented in advance, however bad the documented thing is. Arihat manufactures a scripture that gets three cities burned and bears nothing, on one sentence issued beforehand. Six women told nothing "saw that they had transgressed." Arundhatī is the control — the one party briefed is the one party not deluded. And ch 224, the fully disclosed self-requested impairment, is the one operation after which nobody is billed.
Route two — location and act. A tīrtha discharges a premeditated, pre-costed brahmin-killing (277) as cleanly as it credits an involuntary vrata performed by a starving blind woman who throws away leaves she cannot eat (280). Neither turns on intent.
⚑ Two systems, different evidence, same conclusion: intent is not a variable anywhere in this corpus's liability logic. That agreement is the thread's value; either alone would carry much less.
The unreconciled part, preserved. Ch 260 runs an exact four-debt karmic audit delivered by Śiva in person, including a life ended early for an interrupted pūjā. Ch 277 lets a murderer pay with a bath. Neither passage acknowledges the other and nothing in the range read reconciles them.
What would close it. A rule anywhere in the corpus governing which of the two systems applies to a given case. Not found in ch 151–304.
State: four pages, and the cleanest method line in the build. [n=1 OBSERVED] on the seam, [n≥3 PATTERN] on the enumeration failures.
The claim: this volume's silences are informative because its conventions are demonstrable.
Three convergent findings.
Doublets against disclosures. Four unacknowledged repetitions, against three disclosure mechanisms the volume deploys five times. The machinery existed, was in active use, and did not reach the four cases that needed it.
Enumeration failures, now four. Ch 219 announces 108 forms and enumerates 100. Ch 231 announces eight and explains seven, omitting Rudra entirely. Ch 230 assigns taste twice and sight never. Ch 272's upaliṅga list closes at nine of twelve with a completion formula. Only the first was known before this build.
Repair-annotation density. Endnote 501 flags the Andhaka block at a chapter heading, and the annotation layer's character changes across that block — from naming referents to reconciling contradictions and filling omissions, seven such notes in twelve consecutive. Corroborated by two survivals in the English.
⚑ And the finding that runs underneath all three: the translator discloses when he changes a word and does not disclose when he completes a structure. Endnotes 716 and 717 supply the full five-face and eight-form mappings for chapters whose bodies deliver them incompletely, unmarked.
What would close it. Run the formula-density instrument inside the flagged Andhaka block and compare it to the volume's baseline. The instrument exists, the block is identified, and the comparison has never been made — it would be a genuine test of whether the translator's stylistic impression corresponds to anything countable.
What would break it. A fifth enumeration that comes out exact would not break anything; the ch 233–234 lineage roll already does, at 112 across twenty-eight entries. The interesting break would be finding that the doublets are disclosed somewhere in the corpus and the build missed it.
State: the build's own record, and the least comfortable thing in it.
Two numbers govern everything downstream.
Quotation-level claims held 7 for 7. Sequence-level claims failed 3 for 3. Every claim of the form this passage says P survived. Every claim of the form X caused Y or this is the only instance of Z failed, because each had been assembled from passages read separately and never re-read in order. That converted into the rule that governed the rest of the build, and the rule worked.
The pre-specified lens ran 304 chapters and returned one contested hit. Everything substantive came from instruments nobody planned — list arithmetic, speech-tag coverage, doublet detection, whole-body construction sweeps. The pattern: the working instruments all measured features the author was not choosing.
Three failure modes, each caught in flight. Inline endnote markers break phrase matching. Sampling on the dependent variable returns the answer it started with. And a summary and a count look identical once written down — six chapters were recorded as read and were not.
What would close it. Nothing inside this corpus. This thread closes by being applied to the next build and either holding or not.
Thread 1 and Thread 3 meet at a single point and it is worth naming. Thread 1 says content doesn't decide outcomes; Thread 3 says the text's silences are informative. Put together: the volume leaves arguments standing because leaving them standing costs it nothing — attribution disposes of them — and it leaves repetitions unmarked for a structurally similar reason. In both cases the thing that would have to be done is the thing the text has a cheaper alternative to.
That may be one finding rather than two. It is not stated as one anywhere, and it should not be until somebody can say what the shared mechanism is in a sentence that survives contact with both bodies of evidence.
Do not promote any thread without explicit instruction. Thread 2 is closest to stable — two independent routes, both sourced, contradiction preserved rather than resolved.