Can Meaningful Plaintext Be Transformed into Voynich-Like Text That Cannot Be Uniquely Recovered? A Generative Experiment Using Multi-Stage Information-Loss Transformations

Abstract

This study investigates a constructive question concerning Voynich-like text generation: can meaningful natural-language plaintext be used as source material, subjected to non-injective and deliberately information-losing transformations, and still produce localized word-form regularities resembling those observed in the Voynich Manuscript? A fixed Latin-input generative model, v1.0 FINAL, was constructed using deterministic pseudo-word generation, chapter-specific word-form families, short-range copying, microvariation, word dropping, adjacent transposition, and a small amount of copying error. The study does not attempt to identify the historical production method of the Voynich Manuscript, nor does it aim to minimize an overall statistical distance from Voynichese. The first pre-specified metric, FT-NNS30, did not yield consistent results across three previously unused Latin works, and this negative result was retained. An exploratory measure, NPL300-60, was then introduced to quantify whether similar high-frequency word forms are disproportionately localized in the same chapters. Pliny the Younger’s Letters, Book I, was subsequently used as an independent confirmatory holdout under analysis conditions fixed in advance. The primary NPL300-60 result was observed = 0.104990, null median = 0.099836, with a one-sided empirical upper-tail p-value of 0.000999. The results provide a constructive example showing that meaningful input and reversible preservation of meaning are separable requirements. This record includes the English preprint and a lightweight supplementary package containing the figures and publication metadata. Research data, source corpora, Voynich transcriptions, analysis code, raw null distributions, and event traces are not included in this release. A fuller reproducibility archive may be released separately.