# THE POURS OF REAL PROSE INTO THE MANUSCRIPT'S PAGE SHAPE, COUNTED OFF THE CHAINS' FILES

Written 2026-09-17 at about 22:35Z by the count thread ("How many approaches we tried"), on the coordinator's note of 22:17Z.
Every figure below is taken from a chain's own files on /mnt/project-files/ (read only; nothing run from the mount) or, where a chain has no mount copy, from that chain's own thread and is marked so. Nothing is taken from the coordinator's messages.

## THE COUNTING RULE

A POUR is one real text laid token by token into a page skeleton of the manuscript (its pages, lines and per-line token counts) and scored. A PLACEMENT is one pour at one offset of one text into one skeleton under one page order. A TEXT is one source work as a chain built it (Jókai's two novels, four volumes, are one source; Livy's twenty or thirty-five books are one source). Re-scorings of a placement already laid (another word band, another matching, another denominator, a placebo collation, a length control, a cut of the front matter) are not new pours and are listed as re-scorings.
Both ways: one text at twenty offsets is ONE text and TWENTY placements. Where a chain laid the same starting stretch twice (reading order and as made; continuous and gapped) both layings are counted as placements, and the count of DISTINCT STARTING STRETCHES is given beside.

## CHAIN BY CHAIN

### 1. The control front ("Real texts in the same page layout"; the eighth front) — files /mnt/project-files/controls/
- Texts: 10 sources. Scored at twenty placements: S1 Livy, Ab urbe condita books 1–10 and 21–30 (twenty books, one unit each); S2 the King James Bible; S3 Jókai, Az arany ember and Egy magyar nábob (four volumes, one source); S4 Cervantes, Don Quijote; S5 Montaigne, Essais — STRUCK by the chain's NOTE 7 (results/STAMP_controls.md, 21:42Z): the file holds the original and a modern French rendering one after the other, so its twenty are not twenty stretches of one work; its figures stay in every file as evidence of nothing. Reported with their floors, never scored: R1 Pliny NH preface and books 1–5, R2 Apicius, R3 Caesar BG, R4 Kivi, R5 the Kalevala.
- Skeletons: quire T's 23 pages (10,820 tokens) in reading order (CONTINUOUS pour) and as made (GAP pour: runs of 12,702 words with 1,882 skipped for the lost sheet 109+110); quire M's 20 pages (6,786 tokens). All pours are extracts (runs of 10,820, 12,702 or 6,786 words from the head of each unit, non-overlapping); no book was poured whole.
- The record run (results/controls_T23.jsonl 243 records; results/controls_M20.jsonl 141 records; results/REPORT.md line 3 "Placements per source (T23 continuous / gap / M20)"): livy 20/19/20, kjv 20/20/20, jokai 20/19/20, quijote 20/20/20, montaigne 20/20/20, pliny 5/4/8, apicius 2/2/4, caesar 4/4/7, kivi 7/6/12, kalevala 6/5/10. Columns 124 / 119 / 141; TOTAL 384 placement records. The files and REPORT.md agree exactly. Distinct starting stretches (source, unit, start, length): 353 (31 as-made placements begin on the same stretch as a reading-order one).
- R1, the as-made ordering with more Livy (results/r1_T23.jsonl 164 records; logs/r1.log END 21:43:18Z; results/R1.md): Livy's 35 books in two units, 45 placements in reading order and 39 as made (no cap; floors one in 46 and one in 40); KJV placements 20–39 under both pours (40 each, floor one in 41); Quijote 20–34 and 20–29 (35 and 30); Montaigne 20–29 and 20–24 (struck). 164 new placements, 162 distinct stretches.
- Standing after the strike: the headline rests on 4 texts × 20 = 80 placements in reading order (78 as made: Livy 19, Jókai 19; 80 on quire M); floor one in 21 within a text. Everything this chain has poured: 548 placement records, 473 with Montaigne out; distinct stretches 515, 441 with Montaigne out.
- Re-scorings, not pours: R2 (quire M re-read from saved pairs), R3 (Jókai placement 0), D1 (other tune rules), D2 (placebo collations), D3 (matched rows), and the apparatus cut (DIAGNOSTIC STAMP A, 322 tasks re-scoring the same placements from corpora with front matter and title lines cut; running since 22:08Z, due about 22:45Z; its files are not on the mount yet by the chain's NOTE 10).
- Report against files: the chain's own count sentence in its thread (22:19:54Z) says "Pliny 2" placements; the files hold Pliny 5 in reading order, 4 as made and 8 in quire M (unit nh_pr_1_5_joined), and REPORT.md line 3 says 5/4/8. The stamp's addendum 1 (STAMP_controls.md line 57) had said "the rule leaves Pliny 2 placements" before the six units were joined. THE FILES WIN: 5/4/8. Pliny is reported-only, so no scored figure moves; the count sentence understates it by three.

### 2. The tenth front (methods-b) — files /mnt/project-files/methods-b/diag/pour3/ (224 files; MANIFEST_pourlat/pourspa/poureng/gaplat/gapspa/gapeng.txt), DIAG-B.md section 80 part four, stamp diag/battery_stamp8_pour.md
- Texts: 3. Caesar, De Bello Gallico (latin_caesar.txt, 51,295 words): ONE pour of tokens 0–34,780, two thirds of the book, the largest extract anyone poured; Don Quijote from token 4,828 (the front matter skipped): TEN disjoint windows of 34,780; the King James Bible from token 383: TWENTY disjoint windows of 34,780. Montaigne rejected at the stamp for the fault the control front found later.
- Skeleton: the whole book, 207 pages, 34,780 slots, quire T its last 23 pages. Each pour written twice: CONTINUOUS and GAPPED (1,882 tokens skipped at f109–f110). 31 + 31 = 62 poured documents (the manifests list 1 + 10 + 20 under each); 42 batteries scored (31 continuous + the gapped Latin + the 10 gapped Spanish). Distinct starting stretches: 31.
- Floors: Spanish one in 11, English one in 21; the Latin has none (one text).

### 3. The ninth front (methods) — files /mnt/project-files/methods/DIAG.md sections 111 and 120, diag/centre_120.txt, code/prose111c.py, code/centre120.py
- Texts: 3 (Caesar, Montaigne, the King James Bible). Skeleton: the chain's 25-page held skeleton (1,164 lines, 11,567 tokens).
- Section 111 (20:50Z): one placement each from token 0 (3 pours). Section 120 (22:00Z): twenty placements each, placement j from token j × step, step = (words − 11,567) ÷ 19: Caesar step 2,090 (the twenty OVERLAP, each window 11,567 tokens of a 51,295-word text), Montaigne 16,756, KJV 41,084. Placement 0 of each is section 111's pour (parity HELD, centre_120.txt lines 2–4), so the chain's placements are 60, not 63. Montaigne's twenty are of the struck file (its later placements lie in the translation half).
- Floors: one in 21 within a text, one in 61 pooled; on the two clean texts 5 of 40, one in 41.
- Standing: 40 placements (Montaigne out). Sections 116–118 (bands, the repeat, the frequent split) re-score section 111's Caesar placement; no new pour.

### 4. The second front ("Wider hunt, continued"; hand3) — files /mnt/project-files/period-search-2/programs/hand3/run/pour/ (pour_manifest.md, pour_stamp.md, pour_P1..P4.json), HAND3.md lines 6911–7312
- Texts: 2 works in 3 files: Newton's Opticks (99,894 tokens); Caesar's Gallic War I–IV (pg218, 20,545 tokens) and V–VIII (pg18837, 30,880 tokens).
- Skeleton: the held set's 94 pages (2,138 lines, 19,621 tokens), quire T's 23 pages the last 23 (the pool). Pours: 4 — P1 Newton from token 1,000; P2 Newton from 50,000; P3 BG I–IV from 500; P4 BG V–VIII from 500; each 19,621 tokens (P3 takes 95 per cent of its file). Floor one in 5.
- No later set. The files splitPourB_quireT23_P1..P4 (21:37:53Z) re-read the same four dumps on the 23 pages; run/lenctl (21:52Z) and run/pmatch (22:16Z) re-score the same four; the memory-bearing source was ordered, then put behind the product-matched pools, and never run (HAND3.md 7153, 7163). The chain has withdrawn its conjugate figures (pmatch M3: 0 of 4 pours back).

### 5. The main thread (the report; its own tree, NO MOUNT COPY — counted from its reports as relayed, not verified here)
- Texts: 4 — Caesar (Latin), Dante (Italian), the Nibelungenlied in Simrock's modern translation (German), the King James Bible (English); one contiguous chunk of 34,780 words each into all 207 pages (tools/meaning/results/controls.json): 4 pours, scored on t3_bifolio row 91, the held-out surprisal (row97_pours.py, 20:10Z), the letter step, the line-length regression (sentence 12's pours, run by 21:48Z) and the quire test (job 5).
- The WIDTH POUR (tools/sheet_pours/gap11_width.py, stamped 21:37:05Z, run by 21:48Z): the same four texts laid by letter width, each line filled to the real line's letter total: 4 more arrangements of the same chunks. Reported figures: width-poured texts 12 to 17 per cent, word-count pours under 1 per cent, the manuscript 5 per cent.
- Total 8 arrangements of 4 texts (4 distinct chunks). The derivation front's README names the same three chunks (Latin Caesar, Italian Dante, English KJV, 34,780 words) as its language controls; whether they are the identical dumps cannot be checked from the mount.

### 6. The new automated front ("Chance-picked source pages, restarted") — NO MOUNT COPY (a permission gate refused it); from its thread
- Pour stamped (pour_stamp.md 5dc29efc, pour_addendum_p8.md; pour_b.py d91ca8b8): Montaigne, Don Quijote, the King James Bible, twenty offsets each = 60 pours into quire T's page skeleton, the lost bifolium poured and discarded. STARTED 22:19Z, about twenty minutes; NOT YET REPORTED. Twenty of the sixty are of the struck Montaigne file. Not in the totals below.

### 7. Earlier layers (the report's rounds, before 2026-09-17; files in the report tree, not on the mount)
- Round 4 (2026-09-14): 28 matched-size language samples poured into the page skeleton as controls for the glyph and word statistics (memory voynich-round4-findings). A different purpose and a different layer; not in the totals below.

## THE TOTALS (tonight's pours, files as they stood at 22:20Z)

| | texts | placements laid and scored | distinct starting stretches |
|---|---|---|---|
| Control front, record | 10 | 384 | 353 |
| Control front, R1 | (same + 15 more Livy books) | 164 | 162 |
| Tenth | 3 | 62 | 31 |
| Ninth | 3 | 60 | 60 |
| Second | 2 (3 files) | 4 | 4 |
| ON THE MOUNT | 11 sources | 674 | 610 |
| Main (own tree) | 4 (2 new: Dante, Nibelungenlied) | 8 | 4 |
| ALL FIVE CHAINS | 13 sources (14 works) | 682 | 614 |
| ALL FIVE, MONTAIGNE OUT | 12 sources (10 on the mount) | 587 (579 on the mount) | 520 (516 on the mount) |

Montaigne out removes 75 control records (60 + 15), 20 of the ninth's, none elsewhere.
The headline test (the conjugate contrast in quire T) rests on 80 placements of 4 texts (the control front); the other chains' pours are independent checks on other scorers: the tenth 62 on 3 texts, the ninth 40 standing on 2 texts, the second 4 on 2 works, the main 8 on 4 texts.

## THE FIGURE "TEN REAL BOOKS AND 384 SEPARATE POURS" (POST 34, 21:14Z)

- 384 is exactly the sum of REPORT.md's placements-per-source line (59 + 60 + 59 + 60 + 60 + 17 + 8 + 15 + 25 + 21) and exactly the record count of controls_T23.jsonl (243) plus controls_M20.jsonl (141). It is a real count from one chain's files. In that chain's own vocabulary it is 384 placements under two pours, not 384 pours.
- As "separate pours into the manuscript's page shape" it is inflated by: 141 placements into quire M (a second, 20-page skeleton, where the finding does not appear); 119 as-made re-layings on quire T, 31 of them starting on the very stretch a reading-order placement starts on; 60 placements of the struck Montaigne file; 86 placements of the five reported-only sources that were never scored. The part the headline rests on is 80.
- Since POST 34 the same chain has laid 164 more (R1), so as a count of what it has poured in all it is now too small: 548 records, 473 standing.
- "Ten real books": ten sources, one struck (nine), four scored. "Books" undercounts (Livy 20 books, 35 in R1; the Bible 66; Jókai two novels; Pliny six) and "sources" or "texts" is the honest word.
- For the whole search the figure is too small on both counts: five chains poured 13 sources (12 standing) in 682 placements (587 standing), and 384 counts one chain's record run only.

## ONE SENTENCE THE PAGE CAN CARRY (verifiable from the files named above)

"We poured twelve real texts into the manuscript's own page and line layout, in 587 separate placements across five independent checks; one text at twenty offsets counts as one text and twenty placements; the headline test rests on four of those texts at twenty placements each, with a floor of one in twenty-one within a text."
Counted from the shared folder alone (the main thread's four texts and eight arrangements left out): ten real texts, 579 placements, four checks. Counting the struck Montaigne file as poured: thirteen texts, 682 placements. The new automated front's 60 (40 clean) will add to these when they land; the control front's cut run adds nothing (it re-scores placements already counted).

## ADDENDUM, written 2026-09-17 at about 23:25Z: TWO POUR SETS THE 22:35Z COUNT DID NOT CARRY

The section above is left as written (its totals were the files as of 22:20Z). Two pour sets are added here, one missed and one landed since.

### 8. The eleventh front (leaps) — files /mnt/project-files/leaps/results/T14_labelcal.md, T14_labelcal.json, T14_labelcal_stamp.md (22:51:35Z, before the code), code/t14_labelcal.py, corpora/, HYPOTHESES.md lines 907–910
- MISSED by the 22:35Z count: this front had been told at 20:36Z not to pour; the coordinator ordered this calibration pour at 22:47Z and it ran 22:53:40Z to 22:56:30Z, after the count's file snapshot and before its reply; its files carry no "pour" in their names.
- Texts: 5 reference corpora, three of them new to the inventory: Cicero, De Officiis I–III (latin, 34,230 words); Manzoni, I promessi sposi (italian, 60,012); Malory, Le Morte d'Arthur vol. 1 (english, 60,006); the King James Bible (bible, 60,006, from word 500); Dante (dante, 60,009, from word 1,500).
- Skeleton: the SEAT pages (25 pages, the real's page lengths, writing order). Pours: 18 of 11,567 words each (four per 60,000-word corpus, two of Cicero); five PRIMARY (one per corpus) and 13 beside. Extracts. Floors one in 6 on the primary five, one in 19 on the 18 (mixed texts).
- Result as the file states it: P-c1 HELD, P-c2 REFUTED (italian_pour2 at 0.00062 on the primary cell, the real 0.00074). results/T14_rawrank.md (23:14Z) re-ranks the same 18; no new pour.

### 9. The new automated front ("Chance-picked source pages, restarted") — its own tree, NO MOUNT COPY (a permission gate refused it); counted from its thread
- LANDED since the count: the run finished 22:37:39Z, 60 dumps of 60 checked against its MD5SUMS, read as §B102 at 22:55Z. Three texts (Montaigne, Don Quijote, the King James Bible) at twenty placements each into quire T's page skeleton, the lost bifolium poured and discarded: 60 pours, 60 placements, 40 standing (Montaigne's twenty are of the struck file). Extracts. Floor one in 21 within a text; the chain has been asked for the effective number over the sixty.
- Pour-2 (started 23:10Z) re-scores the same 60 under the BOTH and EXTERNAL bands; no new pour.

### REVISED TOTALS (files and reports as of 23:20Z)

| | texts | placements laid and scored | distinct starting stretches |
|---|---|---|---|
| The five chains of the 22:35Z count | 13 sources | 682 | 614 |
| Eleventh (on the mount) | 5 (3 new) | 18 | 18 |
| New automated front (own tree) | 3 (0 new) | 60 | 60 |
| ALL SEVEN CHAINS | 16 sources (17 works) | 760 | 692 |
| ALL SEVEN, MONTAIGNE OUT | 15 sources | 645 | 578 |
| ON THE SHARED FOLDER (control, tenth, ninth, second, eleventh) | 14 sources, 13 standing | 692, standing 597 | 628, standing 534 |

Montaigne out removes 75 control records, 20 of the ninth's and 20 of the automated front's.

### THE SENTENCE, REVISED
"We poured fifteen real texts into the manuscript's own page and line layout, in 645 separate placements across seven independent checks; one text at twenty offsets counts as one text and twenty placements; the headline test rests on four of those texts at twenty placements each, with a floor of one in twenty-one within a text."
Verifiable from the shared folder alone: thirteen real texts, 597 placements, five checks. Counting the struck Montaigne file as poured: sixteen texts, 760 placements.