A continuously updated public record
The Voynich Manuscript Beinecke MS 408
A century of failure. A new method — a coordinated team of AI agents working in parallel under human direction.
A continuously updated public record
A century of failure. A new method — a coordinated team of AI agents working in parallel under human direction.
This is an open log of an analysis in progress. No decryption is claimed. Every entry is provisional. The Voynich Manuscript remains unread.
For over six centuries, the Voynich Manuscript has resisted every attempt to read it.
Carbon-dated to the early 1400s, its 240 vellum pages hold an unknown script, impossible plants, figures bathing in green pools, and astronomical diagrams that match nothing in our world. World-class cryptographers — William Friedman and his team, who broke the Japanese PURPLE cipher — spent decades on it and produced no readable plaintext. Every published "solution" since has been refuted.
The manuscript may not be a cipher at all. It may be a constructed language, a glossolalia, a sophisticated hoax, or something we do not yet have a category for. The honest position, after a century of failure, is that no one knows.
Seven specialist AI agents work in parallel on every word. An adversarial Skeptic agent attacks every hypothesis. A human author authorises every label.
This is not "AI cracks the Voynich" theatre. Most outputs are labelled SPECULATIVE or FAILED, and that is the point. We publish the failures alongside the breakthroughs. Every step is reproducible, every dataset is open, every diary entry is append-only.
Decomposes each EVA word into prefix, root, and suffix candidates; identifies gallows characters and rare glyphs.
Computes per-section concentration, line-position bias, Currier-A vs Currier-B purity, neighbour tightness.
Tests similarity against Medieval Latin, Hebrew, Arabic, Aramaic, Old Italian, Old German. Negative results are recorded.
Proposes a mapping to one of four GENOM semantic quadrants: technical, entity, action, abstract.
Generates 2-5 candidate semantic categories (plant_part, container, function_word, …) with explicit uncertainty.
Maps strong left/right neighbours, identifies repeated bigrams, hypothesises syntactic role.
Synthesises the six chains into one falsifiable hypothesis with a preliminary confidence label.
Attacks the hypothesis. May concur, downgrade, or reject. Its veto cannot be overridden.
An append-only public log of every word analysed, every hypothesis raised, every objection registered. Nothing is deleted, only superseded.
Query the live dictionary. Type any EVA token (e.g. daiin, chedy, qokeey) and the API returns its frequency, section distribution, Currier class, neighbours, and current swarm-assigned label.
live endpoint: voynich.radimkaufmann.com/api/word/{eva}
The complete book — methodology, decipherer profiles, statistical analysis, and the full folio commentary — will be available as a paperback on Amazon KDP. The research presented on this site is the foundation; the book is the synthesis.
Coming to Amazon KDP — 2026
This page is the live research log. The book is the closed, edited synthesis. Both will exist; the book belongs on a shelf.