Constraining How the Voynich Manuscript Was Written Without Claiming What It Says

Every few months a new paper or preprint proposes either a plaintext reading of the Voynich manuscript (Yale Beinecke MS 408) or a simple mechanical generator that reproduces one of its famous statistical anomalies. When you test those proposals on the whole manuscript, two problems recur:

1. In-sample tuning: A cipher key, anagram table, or slot grammar is fitted on the same pages used to score it, or evaluated on only one summary statistic (such as conditional character entropy h_2 or Zipf word-frequency slopes).
2. Protocol drift across papers: One study concatenates words across line breaks while another stays within lines; one drops non-ASCII lines and collapses EVA ligatures while another uses raw tokens; one measures voynichese.com grid boxes as if they were physical ink gaps.

I directed the team (the AI coding and research agents I direct) to run a controlled structural investigation across three independent transliterations (ZL3b-n, RF1b-e, and IT2a-n), 223 Beinecke TIFF folio scans, and a 15-quire leave-one-quire-out (LOQO) evaluation (s1 through s106, summarized in F-01 through F-47).

The full illustrated project page is at /work/voynich/ and the complete findings and protocol reconciliation index is at /evidence/voynich/. Here is what the evidence shows, what we corrected along the way, and where every tested model still breaks.

1. What prior literature established vs what we measured

Nothing in our work claims first discovery of the manuscript’s slot grammar, line anomalies, or low character entropy. Those ingredients come from decades of work by Currier, Jorge Stolfi (2000), Sravana Reddy and Kevin Knight (2011), Marco Ponzi (2019), Lisa Fagin Davis (2020), Claire Bowern and Luke Lindemann (2020/2021), Daniel Gaskell and Claire Bowern (2022), Maurizio Zattera (2022), contributors to the September 5, 2025 voynich.ninja thread, and Liudmila Rozanova and Alexander Temerev (2026 preprint).

Our contribution is joint quantitative ablation and out-of-sample protocol bridging:

• Physical lines are compositional units (F-12, F-18, F-21, s33, s35, s36, s37): Gaskell and Bowern’s (2022) stats.py concatenated ~200-word subsamples across line breaks with rook weight w = 1, yielding I = 0.0415. Measuring strictly within lines across all 30,342 adjacent word pairs gives I_within = 0.1572 (44.4% line-multiset baseline + 55.6% intra-line ordering excess, p = 0.001). Extending Ponzi’s (2019) test across 223 folios (s35), true line ends carry 17.70% -m/-g markers and line starts carry 22.32% gallows, whereas internal <-> plant-drawing interruptions (n = 744) drop to 2.82% and 0.27%.
• Descriptive slot localization (F-12, F-14, s33, s36, L-189, L-190): Under our operational [Prefix, Core, Suffix] segmenter (SlotSegmenter), stripping prefixes and suffixes to Core_Only(C) removes 96.0% of full-fit within-line word-length ordering excess and 93.5% of held-out length covariance (166 training folios to 41 held-out folios). Important qualification: because SlotSegmenter is a statistical decomposition fitted on the transliteration, this is a descriptive ablation, not a causal explanation or proof of linguistic morphemes. In the 90.0% (P, S) cell shuffle, 77% to 79% of tokens sit alone in their in-line (P, S) cell (leaving 12% to 13% mutable), and under a fixed morphology segmenter cores retain 48.8% to 54.1% of length covariance.
• Inter-word prefix coupling and within-word 12-slot coupling (F-06, F-15, F-35, s34, s63, s69, s76): Across adjacent words inside a line, positive pairwise mutual information excess concentrates on the next word’s prefix (+0.3171 bits across P1->P2, C1->P2, and S1->P2, compared with +0.3683 positive-cell sum and +0.3190 signed grid sum; because the nine pairwise MI cells overlap, neither ratio is a share of total sequential information). Core-to-core C1->C2 sits -0.0315 bits below within-line shuffle on the full 1,860-core inventory due to short-range core repetition and sparsity, not negative MI or meaninglessness; at matched inventory size (K = 200, s80), common cores carry weak positive sequential order. Within words, Zattera’s (2022) 12-slot grammar treated as independent slots has entropy sum 16.71 bits vs empirical word entropy 11.06 bits (5.65 bits of cross-slot coupling, EXT-ZAT-2022, s63, s76).
• Four-decimal reconciliation with Rozanova & Temerev 2026 (F-03, F-19, s38, s45, s47, s51): Converting our raw vparse boundary MI (0.1380 nats on 4,135 lines) to bits, dropping 80 non-[a-z] lines (28,797 pairs), collapsing 9 EVA ligatures, and subtracting 100 within-line shuffles reproduces R&T’s Table 4 (+0.1972 1-glyph, +0.2426 2-glyph) and Table 11 (+0.1374 Currier A, +0.2477 Currier B) to four decimal places. However, s47 shows voynichese.com SVG boxes are quantized to a 3-pixel grid (R^2 = 0.004 vs Beinecke TIFF ink gaps), and s51 shows that at matched sample size (1,813 pairs), Voynich (0.1599 bits) overlaps Latin Priapea (0.1456 bits, p = 0.38).

2. Consequential corrections we made during the audit

Following our editorial rule to keep consequential corrections in the public record:

• Scribe vs section deconfounding (s60, s77): Broad Davis Hand (1 to 5) x Currier Language (A/B) differences 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 divergence between Hand 2 and Hand 3 disappears; section and quire mixture explain that gap.
• Volunteer gibberish corpus size (n7): Gaskell and Bowern’s volunteer gibberish benchmark has 38 writers (10,023 tokens). Our local 64 sidecar files include pooled and variant files rather than 64 independent corpora, and family X-vs-L counts are sample-size dependent at 79 to 480 tokens per document.
• Recipe paragraph count (s52), pre-registration audit rejections (D-064, D-065), and s96 provenance: Fixing a locus-parser bug that read f102r1 as folio 1021 corrected the Q20 recipe corpus from 295 to 285 ZL3b paragraphs (its non-formulaic incipit result held). Under our pre-registration audit rules, s93 (P-25) was rejected as a unified generator pass (D-064) and upstream s96 (P-28) was rejected after a post-freeze threshold overwrite (D-065), while the separate Daybreak workspace’s db-s96 forward-tail audit and s112 to s121 runs are tracked separately as user-supplied reported evidence.

3. Why the manuscript remains undeciphered

Every familiar single-mechanism family we tested breaks on held-out quires:

• Per-word Naibbe and homophonic/verbose ciphers (H-02 to H-04, s24, s69, s84, s102, s104): Miss within-line prefix memory (F-24!), conditional character entropy h_2, or word-length rhythm (F-35, F-39). Period-documented ciphers (substitution+nulls, homophonic+nulls, 200-word nomenclators, line-held nomenclators) over Latin, Italian, and German recipes fail 0/12 cells (F-45, s102), and spec-built verbose table encodings of real recipe text fail 0/18 cells (F-47, s104).
• Line-held sticky tables (H-14, narrowed under D-067, s81, s98): Match 1-step prefix memory (F-38), but their core streams fail 15-quire LOQO 0/15 (F-34) and their conditional lag-2 prefix dependence is roughly 2.5 to 3 times higher than Voynich (10 to 12 / 15 quires vs 4 to 5 / 15 in s98, F-46).
• Table-and-grille (H-05, s71) and self-citation (H-07, s66, s83): Weakened grilles fail LOQO 0/15 (F-26!); self-citation misses edge coupling or inflates core-length autocorrelation I_core (F-23).
• Our closest multi-layer synthetic producer (s93, rejected as a unified pass under D-064): Still overproduces short-range core recurrence by roughly 2.6x.
• External semantic & crib tests (s29, s35, s92, s99, s103, s106, and user-supplied Daybreak s112 to s121): Zodiac star angles (z = 0.00, p = 0.488), herbal rare-core plant features (p = 0.895), Penn LJS 419 / BL Sloane 4016 herbal runs (s103, INCONCLUSIVE), recipe star-color alternation (s106, HOLD at 17/277), pharmaceutical container/visual similarity (20/31 vs null median 20/31), and Antidotarium Nicolai order-free recipe networks all return null or negative.

Read the full illustrated case study at /work/voynich/ or inspect the complete F-01 to F-47 and Matrix-v2 reconciliation index at /evidence/voynich/.