Line-Conditioned Orthographic Organization in the Voynich Manuscript

The Voynich manuscript exhibits well-known restrictions on word forms and their distributions, but the relationship between local sequential structure and the physical organization of written lines remains contested. We analyze a conservative reconstruction of the ZL3b EVA transcription containing 37,602 accepted word tokens across 227 folio identifiers. Eighteen predetermined orthographic classes, defined from word-initial characters and length rather than interpreted morphemes, were used to predict the class of each succeeding word. Models conditioning on manuscript section and relative line position were compared with regularized first-order predecessor-class models under folio-grouped training, development, and held-out partitions. Three alternative hold-out partitions yielded gains of 0.0415, 0.0644, and 0.0413 nats per token, respectively, with folio-bootstrap intervals above zero. Position-matched predecessor randomizations did not attain the corresponding real-sequence performances. Additional exploratory analyses suggest line-position preferences and substantial changes in predecessor prediction at transcription line boundaries. However, those boundary claims await independent confirmation against high-resolution manuscript images. A position-aware word-selection-and-copying simulator recovered several aggregate positional characteristics but did not reproduce the observed first-order transition gain under the tested, incompletely reoptimized parameters. A separate opener-to-closer prediction test found no additional long-distance constraint once penultimate-word characteristics were controlled. We conclude that orthographically defined classes exhibit reproducible immediate sequential dependence in the analyzed transcription. These findings do not identify grammatical categories, establish linguistic status, specify a source language, or yield a translation. Data and source limitationsThe study used a filtered working corpus derived from the Zandbergen–Landini ZL3b EVA transcription. The official source is available at https://www.voynich.nu/data/ZL3b-n.txt. Source-transcription and derived word-token data are not redistributed in this release pending permission. Included CSVs contain aggregate results or predictive-loss outputs without EVA word strings. This deposit does not establish byte-identical source-to-working-corpus verification. Computational cryptology, Orthographic structure, Sequence modeling, Statistical language analysis, Line-position effects, Manuscript transcription