The Voynich Manuscript and Character Repair in Lexical Systems (CRILS)

The Voynich Manuscript exhibits strong local regularities and low character entropy, yet resists classification as natural language or conventional cipher text. Many prior analyses rely on global statistical resemblance, which can fail to distinguish constrained generative systems from linguistic structure. This paper introduces Character Repair in Lexical Systems (CRILS), a deterministic, leakage-free diagnostic framework that evaluates local admissibility and bounded structural variation at the token level. CRILS measures bigram legality and single-edit repairability relative to a lexicon learned from a disjoint training partition. Applied to the Voynich Manuscript and to natural-language, cipher, and randomized control corpora, CRILS reveals a distinctive regime characterized by low illegality, high lexical novelty, and unusually high single-edit repairability with smooth length-conditioned decay. These properties are not reproduced by natural language, classical encryption, or stochastic baselines, indicating a tightly constrained lexical system operating under regulated local variation.