A Workshop Cipher: A Generative Model Reproducing the Statics and Dynamics of the Voynich Manuscript Text
- 1Independent researcher
Abstract
The text of the Voynich Manuscript (Beinecke MS 408) combines language-like statistics (Zipf law, vowel-consonant alternation, topic structure) with properties no natural language shows, most notably a conditional character entropy h2 ~ 2.1-2.3, far below all tested languages. Greshko (Cryptologia, 2025) showed that a verbose homophonic table cipher ("Naibbe") reproduces much of the static profile; Parisel (2026) introduced four dynamic word-boundary signatures that his own generators and a Cardan grille fail jointly. We present a generative model that, for the first time, reproduces the static profile, the homophone "twin" layer, the serial clustering of homophones within lines, and all four Parisel boundary signatures simultaneously, with four interpretable parameters: a Naibbe-class cipher (six balanced tables, units of one or two plaintext letters) operated with serial reuse of recent units (lambda ~ 0.22), a weak graphotactic choice at word boundaries (K = 2), and moderate spelling noise (epsilon ~ 0.1). Homophone twin distributions are invariant across the five scribal hands, identifying the apparatus as shared workshop technology; any skew of table choice degrades the fit; the boundary preferences can be derived from intra-word transition statistics, removing the last circular element; and a plaintext-genre gradient moves serial clustering and twin rate together, implying the manuscript's high seriality at low twin rate is procedural, not textual. A self-citation ("meaningless text") generator fitted to the static profile remains separable on five of six preregistered windows, and long-range burstiness arises from section-level table dialects, not topicality. The model is a mechanism, not a decipherment: the key is not recovered, even when the model is inverted before attack; we argue on capacity grounds what the text can and cannot encode.
Replication package (87 scripts, corpora, labels dataset, CLAIMS, README EN/RU) is attached and also available at https://voynich.site/files/voynich_replication_v1.zip . Project site: https://voynich.site
v1.1 (2026-09-21). Replication package updated: added test_parisel4.py with its reports and pre-registration (test #61), and analysis/external/seeton and analysis/external/vgcmi - third-party modules imported by packaged scripts (Apache-2.0 and MIT, with notices); the package assembler now follows imports transitively. 333 files, 3.3 MB, sha256 c01688ebd93e9f6dc99a06dee26f13e27f71054c6672a0cbcc43af32f2865fc2. Reported by Igor Golub. The paper PDF is unchanged.
v1.2 (2026-10-22). Corpus correction. The parser reading the ZL transliteration dropped every word written directly next to an in-line transcription mark (paragraph start, a line interrupted by a drawing, comments): 3,530 of 38,190 words (9.2 %), mostly the first and last words of paragraphs. The parser was fixed (every mark is now a word break); all corpus-dependent numbers were recomputed with the same scripts and seeds, and the generator was re-run under pre-registered rules. Numbers shift mostly in the second or third decimal (e.g. twin co-occurrence R 1.87 -> 1.95, burstiness 1.75x -> 1.81x, Taylor exponent 0.669 -> 0.641). No conclusion changes; one previously listed limitation (word-length autocorrelation) is now reproduced by the model. Five passages are reworded and a short "Corpus correction" note is added to Section 2; a hard-coded figure in test_boundary_probes.py (commas ~16 %) is now computed (18 %). Replication package: 431 files, 43.3 MB (adds the multi-language battery corpora, Heaps-curve tests and the K=3 canon test), sha256 088fe2f41f94de1cd7635a4e769847d419d155328bc535e6137d219c89522229.
Related
- The Working Files of a Quintessence Stillroom: The Voynich Manuscript (Beinecke MS 408) as a Plant-Extraction Workshop Record
- A new perspective in linear Cauchy Elasticity: variational minimum principles for statics, dynamics, and heterogeneous materials
- 静的显化理论:自发动力学的公理化构造 / Statics as Manifestation: An Axiomatic Theory of Spontaneous Dynamics
- Training the public health workforce for generative AI: results from a practical workshop in Italy
- Statics–dynamics exchangeability in a quenched Gaussian lattice field: an exchange-rate theorem, an aging-rate selection rule, and a residue tower
BibTeX
@misc{averyanov2026workshop,
title = {{A Workshop Cipher: A Generative Model Reproducing the Statics and Dynamics of the Voynich Manuscript Text}},
author = {Vitaly Averyanov},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23552487},
url = {https://zenodo.org/records/23552487},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - UNPB
TI - A Workshop Cipher: A Generative Model Reproducing the Statics and Dynamics of the Voynich Manuscript Text
AU - Vitaly Averyanov
PY - 2026
DA - 2026/10/22
PB - Zenodo
DO - 10.5281/zenodo.23552487
UR - https://zenodo.org/records/23552487
AB - The text of the Voynich Manuscript (Beinecke MS 408) combines language-like statistics (Zipf law, vowel-consonant alternation, topic structure) with properties no natural language shows, most notably a conditional character entropy h2 ~ 2.1-2.3, far below all tested languages. Greshko (Cryptologia, 2025) showed that a verbose homophonic table cipher ("Naibbe") reproduces much of the static profile; Parisel (2026) introduced four dynamic word-boundary signatures that his own generators and a Cardan grille fail jointly. We present a generative model that, for the first time, reproduces the static profile, the homophone "twin" layer, the serial clustering of homophones within lines, and all four Parisel boundary signatures simultaneously, with four interpretable parameters: a Naibbe-class cipher (six balanced tables, units of one or two plaintext letters) operated with serial reuse of recent units (lambda ~ 0.22), a weak graphotactic choice at word boundaries (K = 2), and moderate spelling noise (epsilon ~ 0.1). Homophone twin distributions are invariant across the five scribal hands, identifying the apparatus as shared workshop technology; any skew of table choice degrades the fit; the boundary preferences can be derived from intra-word transition statistics, removing the last circular element; and a plaintext-genre gradient moves serial clustering and twin rate together, implying the manuscript's high seriality at low twin rate is procedural, not textual. A self-citation ("meaningless text") generator fitted to the static profile remains separable on five of six preregistered windows, and long-range burstiness arises from section-level table dialects, not topicality. The model is a mechanism, not a decipherment: the key is not recovered, even when the model is inverted before attack; we argue on capacity grounds what the text can and cannot encode. Replication package (87 scripts, corpora, labels dataset, CLAIMS, README EN/RU) is attached and also available at https://voynich.site/files/voynich_replication_v1.zip . Project site: https://voynich.site v1.1 (2026-09-21). Replication package updated: added test_parisel4.py with its reports and pre-registration (test #61), and analysis/external/seeton and analysis/external/vgcmi - third-party modules imported by packaged scripts (Apache-2.0 and MIT, with notices); the package assembler now follows imports transitively. 333 files, 3.3 MB, sha256 c01688ebd93e9f6dc99a06dee26f13e27f71054c6672a0cbcc43af32f2865fc2. Reported by Igor Golub. The paper PDF is unchanged. v1.2 (2026-10-22). Corpus correction. The parser reading the ZL transliteration dropped every word written directly next to an in-line transcription mark (paragraph start, a line interrupted by a drawing, comments): 3,530 of 38,190 words (9.2 %), mostly the first and last words of paragraphs. The parser was fixed (every mark is now a word break); all corpus-dependent numbers were recomputed with the same scripts and seeds, and the generator was re-run under pre-registered rules. Numbers shift mostly in the second or third decimal (e.g. twin co-occurrence R 1.87 -> 1.95, burstiness 1.75x -> 1.81x, Taylor exponent 0.669 -> 0.641). No conclusion changes; one previously listed limitation (word-length autocorrelation) is now reproduced by the model. Five passages are reworded and a short "Corpus correction" note is added to Section 2; a hard-coded figure in test_boundary_probes.py (commas ~16 %) is now computed (18 %). Replication package: 431 files, 43.3 MB (adds the multi-language battery corpora, Heaps-curve tests and the K=3 canon test), sha256 088fe2f41f94de1cd7635a4e769847d419d155328bc535e6137d219c89522229.
ER -