Categorical Naive Bayes Model of Morphological Slot Dynamics in the Voynich Manuscript: A Rigorous Attribute-Tagging Architecture and Corpus Audit

Abstract

The Voynich Manuscript (MS 408) has long resisted decipherment under standard natural language translation paradigms and classical substitution ciphers. In this paper, we present a mathematically rigorous computational framework reframing MS 408 not as an unmapped spoken natural language, but as a synthetic, low-entropy attribute-tagging system operating under strict phonotactic constraints. We establish a 16-slot morphological decomposition architecture (S0–S15), functionally categorized into a 4 + 8 + 4 macro-region framework, coupled with a Laplace-smoothed (α = 1.0) Categorical Naive Bayes classifier. To address potential methodological vulnerabilities, we introduce extensive empirical validations: Leave-One-Out Cross-Validation (LOOCV) yielding a 76.44% out-of-sample accuracy, a Null Baseline control proving structural divergence from stochastic noise, and 95% confidence intervals via non-parametric bootstrapping for log-likelihood scores. Furthermore, sliding-window audit protocols reveal structural hybrid zones and codicological misplacements (e.g., f26v and f96v), demonstrating that morphological signatures can assist in identifying historical binding inconsistencies. Rather than asserting a definitive decipherment, our quantitative findings provide strong statistical evidence incompatible with unguided stochastic generation models, establishing a robust foundation for computational codicology.

Keywords