Endogenous Grammar Recovery in the Voynich Manuscript: Word Classes, Syntax, and Sentence Structure from Distributional Affix Profiles

We present a reproducible, corpus-internal pipeline for grammatical structure recovery inan undeciphered script, applied to the Voynich manuscript (∼33k tokens). Across 19 tests, weevaluate competing non-linguistic explanations (random generation, template systems, grille-likemechanisms, and simple substitution assumptions) using permutation-based null models, affix-profile clustering, syntax under within-line shuffle controls, and full-line template concentrationanalysis. Four robustness probes destroy the morphological rule set, the section labels, theanchor-vocabulary domain, and the language itself respectively, and confirm that the syntaxsignal halves under random morphology (p ≤ 0.02), section labels are recoverable above chancefrom blind clustering, pharmaceutical anchors strictly out-produce astronomical, alchemical, andreligious anchors on Voynich, and an identical pipeline correctly identifies pharmaceutical as thedominant domain on a Latin herbal and religious as the dominant domain on the Vulgate. We recover four endogenous word classes with distinct affix paradigms, section-dependent syntacticpreferences, and strong local ordering constraints. At the same time, long lines exhibit near-complete template uniqueness, indicating generated prose rather than repeated formula frames.Prefix-suffix composition is predominantly additive on high-coverage roots, consistent withagglutinative morphology rather than lookup-table behavior. The combined evidence supportslocal grammatical constraints with global utterance diversity and provides explicit falsifiableconstraints for future decipherment attempts.