A statistical characterisation of the Voynich Manuscript's text: sixteen hypothesis-driven steps, from a Zipf's-law test to a mechanical-generation account of Voynichese. In short: the text's real anomalies are very low conditional character entropy (h2 ≈ 2.4 bits vs 3.4–3.6 for languages) and too-uniform word length; a four-parameter copy-and-mutate generator reproduces essentially the whole 17-metric statistical fingerprint; the one structure it under-reproduces is page-level vocabulary organisation; and a pre-registered label-to-image census (the decisive test) finds the labels behave as a formulaic register, not as a catalogue that names its pictures. No decoding is attempted. Each plate states a hypothesis, the test, the result, and the question it opened next. Collapsible IN PLAIN TERMS blocks give worked examples and plain-language explanations; ALGORITHM blocks hold the core code; dotted-underlined terms link to definitions.
All code, the transliteration inputs and the reference outputs are in this repository (code MIT-licensed). The replication pipeline is in the README; every external source, dataset and download is itemised in REFERENCES.md, with the key items in the References section at the end of this page.
The complete list, with per-item provenance, licences and fetch scripts, is REFERENCES.md in the repository.