# MAP1 - consolidated research map of the Voynich project (17/11/2026)

This is a map of what we know, not a reading. No word value or meaning is claimed anywhere in it.

**Sources.** Every number is taken from the cited round's RESULT.md.
- Exceptions: phys1 has no RESULT.md, so it is cited as phys1/REPORT.md. The line-reset numbers come from bot_notes/chainprobe/FOR_MAC.md, as audit2 already checked. Some audit-level numbers are quoted via audit2/RESULT.md, and corr1/month names via the cited FINDINGS lines.
- [unverified] means the number appears only in SYNTHESIS/FINDINGS and was not found in a round file.
- "Power" means detection on planted positives. "no real positive" means the instrument was certified only on our own constructions (rule 1 unmet).
- G0/G4 are the no-message word machines. They are **fitted on the MS** (audit1 grades them MED), so "G4 writes it" is a fit, not independent support.

---------------------------------------------------------------------------------------------------------------------
## 1. The physical book (sourced; lit2, lit3, prov1, phys1/REPORT.md)
| fact | value | source / strength |
|---|---|---|
| Parchment date | 4 leaves (f8, f26, f47, f68). Combined 516+-18 BP gives 1404-1438 (lab); our IntCal20 rerun gives 1405-1435. Every single leaf keeps a 14th-c. lobe (f68: 42.5% in IntCal04) | lit2 table 1, lit3 s0. It dates skins, not ink. Quires 10-20 were never sampled |
| Material | calf leaves (eZooMS 2014); goat cover by a parchment maker's eye (protein inconclusive); iron-gall ink + gum; text and drawing ink "in all probability" the same; azurite, ochre, Cu green | McCrone 2009, Yale 2014 (data unpublished) - lit2 |
| Other inks | folio numbers: richer in Fe. Quire marks: very low Fe. n=1 fleck each | lit2 |
| Hands | 5 hands. Hand 1 = Currier A; hands 2-5 = B. Exception: hand 3 writes A on f58. f115r lines 1-12 are hand 2 inside a hand-3 quire (s1: undecided) | Davis (self-described as "subjective"), phys1/REPORT.md |
| Collation | 18 of 20 quires and 102 of ~116 folios survive. No ruling or pricking is evident (one ruled foldout paragraph) | voynich.nu via phys1/REPORT.md, lit2 |
| Sequence | spill -> disbinding -> bifolia misordered -> THEN 15th-c. quire marks -> later q9 resewn -> 17th-c. foliation | Davis 2025 via lit2/prov1 |
| Marginalia | f116v and f17r: one German hand near production. Month names: a later Romance hand (region weak; Occitan-Gascon fits per e3) | prov1, FINDINGS:1370 |
| Custody | Tepenec ex libris (1608-1622) = earliest secure owner -> Baresch -> Marci -> Kircher 1665/66 -> Collegio Romano -> Voynich 1912 -> Kraus -> Yale 1969. The Rudolf II purchase rests on one third-hand source | prov1 |
| MSI (Lazarus 2014, released 2024) | 10 folios; "no palimpsest or underwriting" | phys1/REPORT.md, msi1 |
| Corrections in the text | about 1 word-level change in 11,878 words (10th-14th pct of 73 German MSS) | corr1 via FINDINGS:873 |

---------------------------------------------------------------------------------------------------------------------
## 2. What the text IS (positive description)
**2a. The line is a closed unit.**
- Mid-line, a word's first sign depends on the previous word's last sign. Normalised MI is .082 mid-line vs .0048 across a line break (chainprobe 32,844 / 4,415 junctions). A .055 / B .110 mid-line (chainprobe).
- Real sandhi/mutation languages carry the dependence across the line: ratio .61-.92 vs Voynich -.01 (xmi2).
- One-entry-per-line lists reproduce the stop and the absence of split words, but not the edge preferences or the varied line ends: PARTIAL (xmi3).
- Line-end endings are MORE varied, not rhymed (xmi2). Plain Timm-Schinner text does this too.
- The last gap of a line is squeezed: -.506 SD, replicated on a fresh half; real hands -.00 to -.13 (lend1). The mechanism (planned vs out of room) is INCONCLUSIVE.
- Line-final -m/-g words carry the B line-final preference. They are ~18% of B lines (margin1) and are 2.5x rarer on paragraph-last lines (margin2, "reported, not voted").
- The space is a real unit boundary; under 15% of spaces could sit inside cipher units (seam1). That dose was certified at ~316/~120 units only.

**2b. The edge chain is a writing habit on word shapes.**
- The previous word sets only the next word's first sign(s): REACH .03, against .44-.53 for whole-word controls; 99% of the gain is on sign 1 (deep2).
- deep1 agrees by another method (RR .05 vs >=.55).
- The previous word is read deeper than its last sign: +.055 bits/word, z 8.2 (deep2).
- No lag-2 effect: .009 (z 1.0). The bar would see a running key at >=.032 (deep1).
- Edge-local shape: shuffle-corrected rc .745 vs languages .05-.21. A no-message G4 gives .863, so this describes the writing rule (spread1).

**2c. Word structure (morph1).**
- Nearest a SLOT GENERATOR: H4 6/6 bars on SCORE and FIT. Verbose 4/6, abbreviation 2/4, proclitic 0/6.
- No small prefix set, unlike Hebrew/Arabic w b h l m d. 20/22 real texts differ on >=3 bars.
- Above its own sign 3-gram model, the MS carries +.04 of prefix-stem restriction; real texts carry +.06 to +.47 (post hoc).
- The '57% sign + word' decomposability is produced by the MS's own sign chain.

**2d. Repetition and word order.**
- No repeat avoidance within a line: LINE 1.010, PAGE 1.297 (sher1). Every language tested avoids repeats.
- Class contrast: -aiin .66 vs qo- 1.31 (sher1). Identical qo neighbours .114 vs generators .006-.039 (mech2/r6).
- Repeated runs are 2-word pairs, mostly one edit off: L>=3 runs are 2.9/2.8%, vs 8.5-68% in real texts. No set phrases (gk3).
- The word-order excess over G4 is first-order and habit-like (gk4):
  - a deeper shape seam (37-68% of it);
  - pair re-use within ~400 tokens;
  - a line-slot habit.
  - Static pairs transfer between halves at .0003/.0000, vs real texts .10-.34. No class grammar.
  - One borderline axis: pair direction beyond the seam, z 2.5/3.3.
- No body-to-body dependence. Topic signal: .06-.11 vs .24-1.14 in real herbals; INCONCLUSIVE by halves (body1).

**2e. Page and paragraph habits.**
- Pages differ 3-5x more than any single-table generator [3-5x unverified]. ~75-90% of that spread lies BETWEEN Currier A/B x section groups (spread1).
- Page-first words are rare on most sections = a generic page-start habit. fw2's verdict is CONFOUNDED-section; herbal-A z +3.3 (audit2).
- The paragraph first line is not a heading and not a formula: power .85-1.00 vs Macer/F292/F120 (head1).
- Its one-off forms are FRESH one-edit variants of body words: R1 .945 (A) / .950 (B) vs .48-.76 in real texts with chapter names (head1/r2).
- Labels: 87% are one word; seam present inside multi-word labels, z +4.91 = same routine as the text (lab1).

**2f. Currier A / B.**
- Same rules, different settings. The seam is ~2x stronger in B: A .055, B .110 (chainprobe); hand 1 .123 vs hands 2/3 .250/.249 (deep1).
- Depth is the same size in hands 1/2/3 (.035-.042, deep1).
- Dialect, hand and section are entangled; hand vs dialect is separable only at f58 (struct1 A6).

---------------------------------------------------------------------------------------------------------------------
## 3. EXCLUDED / WEAKENED hypotheses
Status words are used as follows:
- **excluded** = failed with power, within the stated scope.
- **weakened** = failed, but power is partial or the scope is narrow.
- **untestable** = no instrument has power.
- **open** = not tested with power since audit2.

**3a. audit2 (b) families (11) and what happened since**
| # | family | rounds | power on planted positives | MS result | status |
|---|---|---|---|---|---|
| 1 | key-placed letter in word ENDINGS (<=2.2 bits/word) | fwd1, gk1 | gk1: unseen at <=1 bit/word; ending/prefix key seen only from 2 bits/word vs its own twin | gk1 INCONCLUSIVE | **untestable** below 2 bits/word; at >=2 bits it fails page burstiness, conditional on G0 being the writer |
| 2 | verbose ~800 units WITH borrowed habits | rck7 family #2 | habits add median +1 (bar +2) | plants pass 1-4/13 facts vs G0+habits 9-11/13 | **excluded** at scorecard level |
| 3 | syllable unit, size-matched | piece2, piece3, rck5 add-on | piece3 family-A ceiling .22-.47 (< .60) | nothing decoded | **untestable** |
| 4 | verbose Arabic | rck7 (gate 3/8), rck8 (gate 8/8) | Arabic U800 .89-.94 letters | agreement B .180 / HA .267 (bar .8) | **excluded** ~400-800 units, line-stretch keys |
| 5 | heavy per-token free choice (homophones + vowel deletion) | j1 only | - | - | **open** (no powered test since audit2) |
| 6 | low-reuse code book | - | - | - | **open** (no test since audit2) |
| 7 | fitted chain table / word-triggered relabel | chain1, chain2 (19/19 families fail) | constructions only | - | **open**; chain2 leans to the generator reading |
| 8 | divider+letter space; gap-width code | seam1, width2, glyph2 | width2: planted letter caught at >=.5 SD [unverified] | width beyond flanks .0047 < real hands .009-.011 | width code **excluded** at >=.5 SD; divider+letter **untestable** |
| 9 | heavy-strip encodings | body1 (strip ~7% vs 16-30%) | - | - | **open** |
| 10 | message in a small share of words | gk1 (low-rate keys) | unseen at <=1 bit/word | - | **untestable** by statistics |
| 11 | no message (G0/G4-type routine) | fwd1, mech1 r2-r4, mech2, gk1-4, morph1 | fitted on the MS | mech1 r2: simple procedures hold <=2/9 targets; r4 best 3/11; mech2 5/11 | **not closable** by text statistics; leading fit, but no short procedure writes it |

**3b. Rounds since audit2 (and the pre-audit lines they rest on)**
| hypothesis | round(s) | power on planted positives | MS result | status |
|---|---|---|---|---|
| Naibbe-type verbose cipher, ~400-800 units, Latin | rck6 ADD4, rck5 | Latin gate 3/3 (rck5 only via P5 polish) | agreement .232 (rck6) / .167 (rck5), bar .8; TS fits B as well or better | **excluded** (2 decoders) |
| same, Italian | rck6 ADD3 (rck5 gate failed 2/3) | gate 4/4 cells, letters .926-.976 | agreement .252 | **excluded** (1 decoder) |
| same, German (MHG) / Catalan (modern proxy) | rck7 | gate 8/8 each | agreement .251-.377, fail x3 | **excluded** |
| same, Hebrew / Arabic | rck8 | gate 8/8 (Heb .948-.986, Ar .892-.963) | agreement .125-.267, below the wrong-language null .348 | **excluded** |
| same family, per-WORD key choice or tables >1000 units | rck8 limits | clustering assumes key stretches | - | **open** |
| plain language written letter-for-letter (proclitic, suffixing, abbreviated) | morph1 | 22 real texts as controls | H4 slot generator 6/6; H1a 0/6 | **weakened** (no certified H2/H1c controls) |
| generator + key vs generator alone | gk1 | 1,800 worlds; nothing at .5 bit/word; whole-word key unseen to 3 bits in G0 (G4 r3, TS r2) | INCONCLUSIVE | **untestable** at low rate |
| page/line reuse writes the word-order excess | gk2 | recovery 3/3 | line copy reaches CZ_W (79%) but loses L3, N, F, P | **weakened** (3 variants only; reinterpreted by gk3) |
| repeated set phrases / formulas / refrains | gk3 | 11 real corpora; 42/43 windows self-call | L>=3 runs 2.9% vs 8.5-68% | **excluded** (phrase-level formula); a variant cipher over a formulaic text NOT separable (.42-.46) |
| syntax / class grammar in word order | gk4 | real texts .10-.34 static pairs | .0003/.0000 | registered HABIT-LIKE: transferable pair grammar **excluded**; longer-range order at language strength UNRESOLVED; code-hidden syntax not excluded |
| zodiac labels = day counter / ordinals / dominical letters | zod5 | real calendars z 7-64; power 1.00 at >=50% (counter) / >=75% (letters) | max z +1.42, 0/12 cells | **excluded** |
| zodiac labels = lunar-mansion names | zod5/r2 | mansions XSIM +63.8 | MS +13.4 | **excluded** |
| zodiac labels = per-day saints names | zod5/r3, r4, r5 | r4: .73-1.00 if ring starts known within one figure; <=.27 if unknown | r4 z +.52 | **excluded** only for known ring starts; r5 found no ring-start clue -> line CLOSED pending imaging |
| labels as spelled-out names under the word-per-letter code | lab1, lab1/r2 | arithmetic + plants | 1.17 words per label vs 6-10 needed; drift fits habit (D1 .87, D2z -1.30) | **excluded** (names); 1-letter labels UNDECIDED |
| paragraph first line = heading / subject / formula | head1, head1/r2 | .85-1.00 vs Macer/F292/F120/Cury | R1 ~.95 = generator level | **excluded**; weak-subject headings, power 0-.34 |
| seam = running key / whole-word state / plain grammar | deep1, deep2 | running key lag-2 >=.032; whole-word REACH >=.44 | lag2 .009, REACH .03 | **excluded**; shape-keyed key untestable |
| one-letter-per-word context-keyed code (code2) | code1, code2 r1-r4 | constructions only (no real positive) | passes the bars, but random letters and G0 pass too | **untestable** by statistics ('G0 + key') |
| the same code recovered by decoding | code3, code3/r2 | truth-init ceiling Latin .79-.80, Italian .81-.84; search .06-.14 | no MS decode (gate failed) | **untestable** at MS size |
| bath labels = organ names (Lia); zodiac = numerals; herbal openers = plant names | crib1 | M1 power ~1 at 50% wrong cribs; B/Z under M2/M3 .00-.05 (at 50% wrong) | 0/28 cells pass Holm | **excluded** for letter-per-glyph; verbose/syllable **untestable** |
| herbal text = cipher of the crib source books (alch. herbal, Megenberg, Crescenzi, Mishnah) | crib2 | in-book M1 1.0; other-book ~0 | no cell above any null | **excluded** for those books only; wider **untestable** |
| public crib on MSI pages (f101v2 vessel/fold; f17r/f116v tails) | msi1 | planted strokes 10/12 hits, FA 2/12 (missed bar) | only known ZL labels; one fill `daraiindy` (PCA, medium) | no crib; "nothing else under paint" INCONCLUSIVE |
| public physical route to ring starts / key | phys1 | - | MSI covers 2 signs (6+ needed) | **closed publicly**; needs Beinecke |
| North-Italian / Chigi copy source | capkey1, capkey3 | a real family member also fails the bars | Voynich 55% (p .074) ~ Sloane 52% | Chigi copy **excluded**; family lean untested |
| bath quire drawings = organs (Lia) | anat2, anat3 | anat3 valid instrument | bath 25/36 vs organ 2/36, but coders not blind; without drawn liquid organs lead 19/36 | **open** (neither supported nor excluded) |
| parallel/twin pages | twin2 | 10 planted AxB copies: power 1.00 at retention .3 | 0 twin pairs | **excluded** down to ~.3 retention (second-cipher parallel invisible) |
| alchemical-herbal template (closing slot) | alch1, chap1 | 1.00/.90 for word-level codes; verbose 5% | FAILS | **excluded** for plain/word codes; verbose **untestable** |
| one-letter-per-word homophonic over Italian | sher2 | 759 plants | 0 match; best 8/13 | **excluded** as built |
| cipher chain over list plaintext | chain1, chain2 | constructions | 19/19 families fail | **weakened** (not 'closed', audit2) |

---------------------------------------------------------------------------------------------------------------------
## 4. What remains OPEN, and why text statistics cannot close it
**The core limit (gk1).**
- A message steering a slot generator at <=1 bit/word is invisible to every statistic tested. A whole-word key in G0 stays unseen to 3 bits/word.
- At MS size a code2-type key is not recoverable even from the true key: the ceiling is Latin .79-.80 and Italian .81-.84, below the .90 needed (code3, code3/r2).
- So "no message" and "message hidden in word choice" give the same text. That would hold even if the word machine fitted perfectly.

| open question | why statistics stop | exact evidence that would close it |
|---|---|---|
| message vs no message (generator + key) | gk1 power limit; code3 ceiling; G0/G4 fitted on the MS | a crib (known plaintext tied to a word); a guessed key rule; or physical evidence of process (ink campaigns, corrections, ruling) |
| per-day saints names in the zodiac | power .27 at most with ring starts unknown (zod5/r4) | an imaged ring-start mark on >=6 signs (B2), then a rerun of zod5/r4 as registered |
| Lia's organ reading of the bath quire | anat3 coders recognised the MS; crib1 verbose/syllable power .00-.05 | blind human coders with hand-masked text (anat3 kill); >=30 label-specific cribs for M2 |
| verbose cipher with per-word key or >1000 units | rck8 clustering needs key stretches | a decoder gated on per-word-key plants, then one MS decode |
| syllable cipher (family 3) | objective ceiling .22-.47 | a new objective reaching .60 on size-matched plants |
| line-end squeeze: rule or out of room | room and "paragraph continues" are confounded (lend1) | right-margin drawings or ruling under raking light; a room dose on lines with no paragraph end |
| what the word-order excess is (gk4) | 'habit' is relative to G4, which is fitted on the MS | the gk4 next-step G4 variant reaching PAIR/ASYM/SLOT and keeping the card; ASYM on a second transliteration |
| the f96v plant kind | a kind (tuberous climber with berries), not an identification (parts10) | botanists blind to the MS naming the same kind from the line crop |
| `daraiindy` under the red paint (msi1) | PCA only, single class | an IR/raking view of the lower-vessel red label (msi1's own request) |
| scribes 3 and 5, quires 10-20 | no date, no ink sample | B4 XRF per scribe; radiocarbon of more leaves (lit3 D1) |

Full untestable register: struct1/UNTESTABLE.md, 71 cells (CORPUS 22, INSTR 33, DESIGN 8, EXT 8).

---------------------------------------------------------------------------------------------------------------------
## 5. Corrections to our own earlier claims
1. Round 49 "only the abbreviation family matches" is retracted: G4 and T&S match too (body1: MS 34.3, G4 35-38, T&S 41-43). Round 49 had no generator controls.
2. "MS cross-word xMI .180 is the highest of 9 languages" is a raw artefact. Alphabet-matched, the MS gives .153 (below Breton/Sanskrit) [unverified: xmi1 corr / bot, not in a RESULT file]. spread1 calls it TIED-TOP with Tongan (z 1.31).
3. fw1-A "opener shape = alchemical-herbal family" is absorbed by the page-start term. But fw2's own verdict is CONFOUNDED-section (audit2 L7).
4. "Cipher-chain family closed" was overstated. chain1 does not exclude a fitted table or word-triggered relabels (audit2 L21).
5. "Three results point to an alchemical herbal" were three non-detections from ONE control text, not positives (audit2 3a).
6. The figural-root "12-20% vs 3%" contrast was overstated: 20% = Pal. 586 (not a family member), 3% = one MS, and the Voynich was never censused (audit2 L31, zooroot5/6).
7. The "North-Italian/Chigi pull" (82/129) was mostly an artefact; hub-corrected 55%, p .074 (capkey3).
8. f96v "solid anchor" is a plant KIND, not an identification. It is Currier A, not B (parts10, twin1).
9. "Verbose ~800 is the one family still open" was wrong twice:
   - ~800 was a found count that inflates;
   - the scorecard had already failed those plants;
   - ~400 was only weakened (audit2 L38, rck2).
   - The family was later excluded in 6 languages (rck5-rck8).
10. Null-dilution "<=5% of words" had no source round; it is UNSUPPORTED (audit2 L40).
11. piece3 was stopped without a RESULT.md: its gate.sh never ran (zsh word-split bug) and nothing was decoded (audit2 L46).
12. "Paragraph openers = body word + gallows" is UNSUPPORTED by a certified instrument (e5 inconclusive).
13. The chap3 "page-bottom formula" hint is closed as layout. Its kill test cannot run because no held-out herbal-B text exists.
14. 1c line-position profile "REPLICATED": the matched value is size-dependent (2.6 vs 6.6, chain2 E10 VOID).
15. The lend1 paragraph-last weakening is unstable across halves (mech1 r4).
16. mech2/r5 "sign + word = proclitic-like" is corrected by morph1: the rate comes from the sign chain.
17. mech2/r7 "-aiin = function-word-like" is corrected by r8: a class-blind anti-repeat writes it, and the MS repeat profile is unlike any language.
18. mech2 "memorised depth" was a wrong null (corrected A4, mech2/r2).
19. sher2's X1/A1 trade-off came from its own construction (code1).
20. gk1 "Timm-Schinner reaches CZ_W .04" was our re-implementation; the published TS gives .021-.028 (gk2).
21. gk2 "ordered word runs" is corrected by gk3: runs carry <=~12% of the excess; it is diffuse type order.
22. rck5 PREREG defect: the P5 rule was not fully blind. Without P5 the Latin gate fails 2/3 (rck5 Disclosures).
23. code3 PREREG ceilings were superseded after a v/u fold mismatch (ADD2); the verdict is unchanged.
24. Month names: a Picard reading was misquoted as corroboration; it is weak (FINDINGS:1370).
25. audit1 (15/11) downgraded many FINDINGS verdicts. Examples: c1 UNDECIDED, d3 "not a STRONG record only", det2 "instrument valid" withdrawn (FINDINGS:82-83, 700).

---------------------------------------------------------------------------------------------------------------------
## 6. Next steps
**Text (no outside contact):**
- gk5: the G4 variant with a recency pair memory + 2-3-sign seam + line-slot habit (gk4 Next). Folder gk5/ exists with dev output only; no RESULT.md and no live process (checked 17/11).
- Matched-null Ilex as a registered round (morph1 Next).
- Per-word-key verbose plants and >1000-unit tables: gate first (rck8 Next).
- crib1 instrument fix: lift the M2 readings cap, then re-measure power (crib1 Next).
- Voynich figural-root census with zooroot5's method (audit2 #3).
- Families 5, 6, 7, 9, 10 have no powered test since audit2. Their closing tests are listed in audit2/RESULT.md (b).

**Beinecke requests (B1-B6, phys1/REPORT.md) - AWAITING LIA'S DECISION. Nobody has been contacted.**
| # | request | what it would decide |
|---|---|---|
| B1 | Zyats 2016 data + the remaining Lazarus pages | ink/material per page; more MSI targets |
| B2 | MSI of the zodiac f70v2-f73v | ring-start marks -> the only route to reopen zod5/r4 (saints names) |
| B3 | MSI/XRF of marginalia vs text ink | are f116v/f17r/month names the same campaign as the text |
| B4 | XRF ink per scribe | ink batches per hand; scribes 3 and 5 |
| B5 | >=740 ppi masters | glyph/gap measurements |
| B6 | spine CT / raking light | sewing history, ruling/pricking |

Unsent drafts: lit2/REQUEST_DRAFT.md, ir1/REQUEST_NOTE.md, lit3/DRAFTS_UNSENT.md.

**Bottom line.**
- The text is a rule-driven, line-closed writing system.
- Its words are built like a slot generator; it carries no language-like syntax, phrases, headings or counters at testable strength.
- Every classical cipher family we could gate fails.
- Neither a simple meaningless procedure nor a tested cipher explains it (mech1 r2: mystery open).
- A low-rate message hidden in word choice cannot be excluded by text statistics; only a crib, a key rule or physical evidence can decide it.
