Structural Investigation, Beinecke MS 408, Not Deciphered
Constraining how the Voynich manuscript was written without claiming what it says.
I directed the team (the AI coding and research agents I direct) to test published statistical claims, paleographic transcriptions, scan-aligned word geometry, and candidate production mechanisms on the Voynich manuscript (Yale University, Beinecke Rare Book and Manuscript Library, MS 408). The evidence localizes a disciplined, line-bounded two-layer writing system. It does not decipher the text, establish a validated historical production mechanism, or claim first discovery of the underlying paleographic and statistical ingredients.
Status & boundary (October 18, 2026): The manuscript is not deciphered. Under our pre-registered decipherment standard, decipherment requires one fixed key, a named language or code, a continuous reading across both Currier languages, all five scribe hands, and the major sections, illustration labels that match without per-page retuning, and failure on a shuffled control. Nothing in our runs meets that standard. Local runs through s106 are directly inspected (with s93 and s96 rejected under pre-registration audit rules D-064 and D-065, H-14 narrowed under D-067, and s94 in flight), while results from the separate Daybreak workspace (db-s96 forward-tail audit and s112 to s121 label/recipe runs) are labeled throughout as user-supplied reported evidence rather than locally re-executed checks.
Epistemic separation
Four categories kept separate on this page
1. Established facts & prior literature
Vellum radiocarbon dating (1404 to 1438 CE, 95.4% CI), Currier's A/B split, Lisa Fagin Davis's five scribe hands, Jorge Stolfi's (2000) core-and-crust slot grammar, Sravana Reddy and Kevin Knight's (2011) prefix/stem/suffix and within-line analyses, Marco Ponzi's (2019) drawing-interruption vs line-margin distinction, Claire Bowern and Luke Lindemann's (2020/2021) conditional character entropy (h_2 = 2.11 to 2.19 bits), Daniel Gaskell and Claire Bowern's (2022) volunteer gibberish corpus and word-length Moran's I, Maurizio Zattera's (2022) 12-slot regular expression, the September 5, 2025 voynich.ninja thread on within-line vs cross-line word-length correlation, and Liudmila Rozanova and Alexander Temerev's (2026 preprint) cross-boundary mutual information tables.
Controlled decompositions and out-of-sample tests across three EVA/v101 transliterations (ZL3b-n, RF1b-e, IT2a-n), 41 held-out folios, and 15 leave-one-quire-out (LOQO) folds across runs s1 to s106 (F-01 to F-47). Our contribution is joint quantitative ablation, exact protocol bridging between external papers and our pipeline, and testing candidate generators and ciphers against the combined multi-scale constraints, not first discovery of the individual ingredients.
3. User-supplied reported evidence (Daybreak lab)
Reported results from the separate Daybreak workspace covering the db-s96 forward-quire tail audit, pharmaceutical label visual/container alignment (s112 to s114), and order-free recipe network comparisons against the Latin Antidotarium Nicolai and medical/culinary controls (s119 to s121). These are reported as negative or constraining findings from that workspace and are kept distinct from directly inspected local run files.
4. Hypotheses & open questions
Why our six-stage line-and-quire structural synthesis fits internal statistics better than plain prose, free gibberish, simple/period/verbose ciphers (F-45, F-47), word-level lexical codes (H-06, weight 2.0 / 3.7), or fixed line-held affix tables (H-14, narrowed to 2.0 / 4.7 under D-067), where our closest synthetic producer (s93, rejected as a unified pass under D-064) still overproduces short-range core recurrence by ~2.6x, and what falsifiable external prediction is needed next.
Public-domain manuscript evidence
Six Beinecke MS 408 folios and what our measurements test on them
Folio 1v (Herbal, Currier A, Hand 1, IIIF 1006077):
Shows why the physical line is a production unit rather than continuous text wrapped around drawings. Extending Ponzi (2019) across 223 box-aligned folios (s35), true line ends carry 17.70%-m/-g markers and true line starts carry 22.32% initial gallows, whereas internal <-> drawing splits (n = 744) drop to 2.82%-m/-g before the drawing and 0.27% gallows after it. Bare p/f gallows are enriched 57.46x at paragraph starts (46.77% vs 0.81% mid-line, s37).
Folio 75r (Biological, Currier B, Hand 2, IIIF 1006208):
Dense Currier B prose shows the strongest within-line word-length autocorrelation (I_within = 0.1673 vs shuffle 0.0665, excess +0.1008, p = 0.001 in s33) and boundary character MI (+0.2380 bits raw / +0.2477 collapsed in s38). Under operational [Prefix, Core, Suffix] slots (s34), Core_t -> Prefix_(t+1) (+0.1598 bits) and Suffix_t -> Prefix_(t+1) (+0.1554 bits) in Currier B are 70% to 72% as strong as internal Prefix <-> Suffix binding (+0.2223 bits).
Folio 57v (Cosmological four-cycle ring refrain, IIIF 1006187):
Tested in s38 Part 3 (O-021) across eight transliterations and the native 3028x3769 scan. In the four-cycle outer ring refrain, Slot 8 (f, f, p, p) is confirmed by pixel lobe-separation measurement as a genuine calligraphic switch from single-loop f in Cycles 1 and 2 (lobe_sep = -0.025, -0.007) to split-head p in Cycles 3 and 4 (+0.008, +0.039), whereas Slot 15 (I, c, c, c in ZL3b) is the same curved c glyph with a Cycle-1 top plume.
Folio 99r (Pharmaceutical jar and plant-part labels, IIIF 1006246):
Same-type label pairs share their IVTFF illustration class 46.3% of the time vs 36.2% under within-illustration shuffle (z = +5.24, p = 0.005 in q1). However, s35 Part 3 shows a collision-free 837-type synthetic generator conditioned on illustration class with a 46% shared pool reproduces this score (z = +5.11, p = 0.005) without decoded meaning. In user-supplied Daybreak reported runs (s112 to s114), shared label material did not predict visual plant similarity, container placement carried no transferable marker, and repeated labels agreed on broad class at 20/31 (the null median).
Folio 68r (Astronomical star-ring foldout f68r1-r3, IIIF 1006196):
Illustrates the label prefix register (o-/y-/d-). In s34 Part 3, extracting operational core stems raises label-in-body overlap on the same folio from 12.1% (words) to 48.7% (cores), but a within-(section, hand) permutation null rises identically from 10.6% to 46.8% (excess +1.95%, z = +1.07, p = 0.161). On the zodiac angular test (s29 Test C), planted noisy homophonic ciphers are detected in 40/40 runs (z = +26.6 to +42.8), whereas the manuscript's 344 zodiac labels show zero angular recurrence (0.0081 vs permuted 0.0082, z = 0.00, p = 0.488).
Folio 116v (Final-page marginalia, IIIF 1006277):
Historical crib proposals on f116v, star names, and plant identifications have varied widely across readers without surviving held-out checks. Our pre-registered decipherment standard requires any proposed reading or key to generalize out of sample across withheld pages and quires without free per-page parameter tuning.
Interactive protocol & ablation workbench
Test the protocol transformations and slot ablations yourself
Use the two controls below to inspect how external preprocessing choices change cross-boundary mutual information (s38, s51) and what happens to within-line word-length rhythm when operational prefix and suffix frames are preserved or stripped out of sample (s33, s36, s37).
1. Protocol bridge (s38, s51, EXT-RT-2026)
From raw EVA vparse to Rozanova & Temerev (2026)
Select a step to see how each preprocessing choice shifts corpus size and boundary character MI.
Currier B vs Currier A split (s38)B: 0.2380, A: 0.1319 bits
Matched 1,813-pair control (s51): Voynich vs Latin Priapea0.1599 vs 0.1456 bits (p = 0.38)
Step 0 (Raw vparse): Reads all 4,135 paragraph lines (34,475 tokens, 30,342 adjacent within-line word pairs) in natural log units (0.1380 nats). R&T (2026) report in bits on a filtered, ligature-collapsed subset.
Toggle the word representation to see how stripping or preserving operational [Prefix, Suffix] frames affects Moran's I and held-out length covariance.
Out-of-sample 41-folio length covariance retained (s36, 166 train to 41 test)100.0% (+0.2801 char^2 full ref.)
Illustrative line segmentation & token lengths under current view:
qokededy8 chars
qotaiin7 chars
chedy5 chars
daiin5 chars
okam4 chars (-m end)
Full Word (P + C + S): Words average 5.20 chars at line start and taper to 4.71 chars at line end. 44.4% of I_within = 0.1572 comes from page/line length drift and 55.6% (+0.0875) from within-line word ordering. Qualification:SlotSegmenter is a descriptive statistical representation fitted on the transliteration, not proof of linguistic morphemes.
Quantitative decompositions & protocol bridges
Four diagrams summarizing our measured findings and their qualifications
Figure 1: Within-line word-length rhythm and operational slot ablation (s33, s36, s37; EXT-GB-2022):
Gaskell & Bowern's (2022) stats.py concatenated word lengths across line breaks (w=1). Measuring strictly within lines (also noted in the September 5, 2025 voynich.ninja thread) yields I_within = 0.1572 on ZL3b (44.4% line-multiset drift + 55.6% intra-line ordering excess). Under SlotSegmenter, stripping operational prefixes and suffixes removes 96.0% of full-fit excess and 93.5% of held-out length covariance (166 train to 41 test folios, s36). Qualification:SlotSegmenter is a descriptive representation fitted on the transliteration; in the 90.0% (P, S) cell shuffle, 77% to 79% of tokens sit alone in their in-line (P, S) cell (12% to 13% mutable), and under a fixed morphology segmenter cores retain 48.8% to 54.1% of covariance (L-189, L-190).
Figure 2: Inter-word slot coupling (s34) and within-word 12-slot over-generation (EXT-ZAT-2022, s63, s76):
Across adjacent words inside a line, positive pairwise MI excess concentrates on the next word's prefix P2 (+0.3171 bits across P1->P2, C1->P2, S1->P2, vs +0.3683 positive-cell sum and +0.3190 signed grid sum; overlapping pairwise MI terms are non-additive). Core-to-core C1->C2 sits -0.0315 bits below within-line shuffle on the full inventory (raw 0.7289 vs shuffled 0.7604) due to short-range core repetition and sparse-inventory bias, not negative MI; at matched size (K=200, s80), common cores carry weak positive sequential order. Within words, Zattera's (2022) 12-slot product has entropy sum 16.71 bits vs word entropy 11.06 bits (5.65-bit coupling, pairwise MI 1.4041 bits, z = +373.4).
Figure 3: Exact protocol bridge to Rozanova & Temerev (2026 preprint) and scan-aligned box checks (s38, s45, s47, s51; EXT-RT-2026):
Applying R&T's four steps (bits conversion, dropping 80 non-[a-z] lines to leave 28,797 pairs, collapsing 9 EVA ligatures, and subtracting 100 within-line shuffles) reproduces Table 4 and Table 11 of arXiv:2608.17096 to four decimal places (+0.1972 bits 1-glyph, +0.2426 bits 2-glyph). On Beinecke TIFF projections (s45, s47), voynichese.com 3-px SVG grid boxes have R^2 = 0.004 with physical ink gaps, whereas transcription-aligned boxes match certain (.) spaces at kappa = 0.866 and track transition surprisal (rho = +0.168, p = 0.0002), closing O-048 only under s45 box alignment rather than transcription-free segmentation. At 1,813 pairs/sample (s51), Voynich (0.1599 bits) overlaps Latin Priapea (0.1456 bits, p = 0.38).
Figure 4: Constrained six-stage structural synthesis versus tested generator and cipher families (s1 to s106, F-01 to F-47):
Each baseline matches one or two projections of the manuscript and fails the joint 15-quire LOQO battery. Line-held affix tables (H-14, narrowed to 2.0 / 4.7 under D-067) reproduce one-step prefix memory out of sample, but their core streams fail LOQO 0/15 (s81) and their conditional lag-2 prefix dependence is ~2.5 to 3x higher than Voynich (10-12/15 quires vs 4-5/15 in s98, F-46). Period ciphers fail 0/12 cells (s102, F-45) and spec-built verbose recipe tables fail 0/18 cells (s104, F-47). Weakened grilles fail LOQO 0/15 (s71). Our closest synthetic producer (s93, rejected as a unified pass under D-064) still overproduces short-range core recurrence by ~2.6x.
External literature reconciliation
How our protocol compares with published Voynich studies
Rather than treating differences between papers as contradictions, we ran each external method alongside our pipeline on ZL3b-n, RF1b-e, and IT2a-n to isolate what changes when line boundaries, sample sizes, and held-out splits are controlled. Full register mappings appear in the complete findings & reconciliation index.
Summary of external protocol reconciliations and scope boundaries (Matrix v2, October 18, 2026).
Within-line I = 0.1572 splits into 44.4% multiset drift + 55.6% intra-line ordering; Core_Only(C) removes 96.0% full-fit / 93.5% OOS covariance. Gibberish pool has 38 writers (64 sidecars include variants/pools) and is size-dependent (n7).
Drops 80 non-[a-z] P-lines (28,797 pairs), collapses 9 EVA ligatures, subtracts 100 shuffles in bits; uses voynichese.com 3-px SVG boxes.
4-step bridge (s38), Beinecke TIFF ink-gap projections on box-aligned tokens (s45, s47), matched 1,813-pair Latin subsampling (s51; c-368, c-384..c-387, L-139).
Reproduces +0.1972 (1-glyph) and +0.2426 bits (2-glyph) to 4 decimals. SVG boxes have R^2 = 0.004 vs TIFF ink gaps (kappa = 0.866 with certain spaces). At 1,813 pairs, Voynich (0.1599 bits) overlaps Latin Priapea (0.1456 bits, p = 0.38).
Whole-stream conditional character entropy h_2 = H_2 - H_1 = 2.11..2.19 bits without line-frame stripping or Hand x Language stratification.
Core_Only(C) wrapper ablation + Davis Hand (1..5) x Currier (A/B) + section stratification across 165 folios (s60c-506..c-507, s77c-510, L-303, L-305).
Whole-text h_2 combines line-edge wrapper predictability, cross-slot coupling, and scribe/section shifts (JSD = 0.0253, p = 0.0002). Within Herbal-B (s77), the Hand 2 vs Hand 3 grammar split disappears (explained by section/quire mixture).
Independent 12-slot product has entropy sum 16.71 vs word entropy 11.06 bits (5.65-bit over-generation; pairwise MI 1.4041 bits, z = +373.4). Held-out generalization requires cross-slot coupling + 1-step prefix memory (0.0434 bits).
Consequential corrections & scope limits
Where earlier internal claims were narrowed or withdrawn
Descriptive slot ablation vs linguistic morphemes (s33, s36, L-189, L-190):
Stripping operational prefixes and suffixes removes 96.0% of full-fit intra-line word-length Moran's I excess (93.5% held-out covariance), and preserving (Prefix, Suffix) retains 90.0% (90.2% held-out). However, in the 90.0% (P, S) cell shuffle, 77% to 79% of tokens sit alone in their in-line (P, S) cell (12% to 13% mutable), and under a fixed morphology segmenter cores retain 48.8% to 54.1% of covariance. Because SlotSegmenter is fitted on the transliteration, this is a descriptive statistical ablation, not proof of linguistic morphemes or historical scribe rules.
Non-additive inter-word MI and core-to-core sign (s34, s80):
In the 3x3 inter-word slot grid, next-prefix column excess (+0.3171 bits) is 86.1% of the positive-cell sum (+0.3683) and 99.4% of the signed grid sum (+0.3190); overlapping pairwise MI terms are non-additive. Negative C1->C2 shuffle excess (-0.0315 bits; raw 0.7289 vs shuffled 0.7604) reflects short-range core repetition and sparse-inventory bias, not negative MI or meaninglessness; at matched size (K = 200, s80), common cores carry weak positive sequential order.
Transferable word order (F1) and generation discriminators (n2, n5, n6):
Pre-registered rebuttals showed a 4% word-pair habit reproduces quire-split transferable word order F1 at small vocabulary size (n2), and glyph position lock (1.661) sits inside both encoded (1.571..1.931) and meaningless (1.381..1.793) producer ranges (n5). Both were retired as standalone family discriminators.
Hand 2 vs Hand 3 grammar deconfounding (s60, s77):
Broad Currier A/B and Davis Hand 1 to 5 shifts are real across the manuscript (JSD = 0.0253, p = 0.0002), but within the same section (Herbal-B, tested in s77 under nested LOLO/LOBO/LOQO, c-510, L-305) the apparent grammatical separation between Hand 2 and Hand 3 disappears (D_sym = -0.116..-0.269, p >= 0.08); section and quire mixture explain that divergence.
Recipe paragraph count (s52), pre-registration audit rejections (D-064, D-065, D-067), and s96 namespace:
Fixing the f102r1->1021 parser bug corrected the Q20 f103r..f116r recipe set from 295 to 285 ZL3b paragraphs (its non-formulaic incipit result survived). Under pre-registration audit rules, s93 (P-25) was rejected as a unified generator pass (D-064), upstream s96 (P-28) was rejected after a post-freeze threshold overwrite (D-065), and H-14 was narrowed under D-067. Separately, the Daybreak lab's db-s96 forward-quire tail audit has distinct provenance in the findings index.
What this lets us do & what it costs
Why the structural boundary matters and what would count as progress next
Having a falsifiable multi-scale benchmark lets us reject proposed decipherments and synthetic generators quickly when they match one global average (such as character entropy h_2 or Zipf's law) while failing line-reset framing, cross-slot coupling, or 15-quire LOQO transfer. The cost is that internal statistical triangulation cannot by itself bridge from production mechanics to semantic content.
Another internal correlation will only refine the six-stage architecture. A genuine decipherment advance requires an out-of-sample external prediction: a model fitted without the target answer must predict an unseen object class, astronomical or calendar structure, historical codebook feature, or continuous reading that fails on shuffled controls. Explore the full Voynich findings & reconciliation index or read the introductory research note.
File details on this page
Format, size and a short SHA-256 identifier, measured when this page was built.
