The empty corner: the place on the map of languages where only the manuscript sits

There is a simple way to see the central oddity of the Voynich manuscript in a single picture. Take any text and count two things: how predictable the next sign is given the previous one (the conditional entropy), and what share of all the words are distinct words (the richness of the vocabulary). Plot the point. Do this for two dozen languages. Then plot the manuscript.

It lands where nobody else is.

How the test was built

We defined the corner in advance, before the calculation: entropy no higher than 2.60 bits and a vocabulary share no lower than 0.15, at an equal volume of text - 100 thousand characters, with a repeat at 50 thousand. Twenty-seven languages in their ordinary orthography: Latin, English, German, Finnish, Russian, Turkish, Mongolian, Khasi, Santali, Occitan, Italian, Catalan and others - syllabic and morphologically rich, with long words and with short.

The result: natural languages in the corner, zero out of twenty-seven. In the corner sit the manuscript (2.261 bits at a vocabulary share of 0.230), its "language B" on its own (2.028 at 0.208), its "language A" on its own (2.293 at 0.298), and two artificial texts: Greshko's Naibbe cipher (2.092 at 0.212) and our own generator (2.106 at 0.212).

The languages form two camps. The syllabic ones - Mongolian, Blang, Khasi - give low entropy (2.18-2.50) but a beggarly vocabulary (0.02-0.04): few distinct syllables, many repeats. The morphologically rich ones - Santali, Turkish, Russian - give a rich vocabulary (0.20-0.41) but high entropy (3.2-3.5). Low and rich at the same time - it does not happen.

What pushes a text into the corner

The control for the test bench: we took Cato's Latin - "On Agriculture", a dry formulaic text - and ran it through our cipher. It moved from the point 3.22 at 0.423 to the point 2.148 at 0.223. Into the corner. A meaningful text with real syntax and a real vocabulary became statistically indistinguishable from the manuscript on these two axes. That is the answer to those who like to say "low entropy means nonsense": no, low entropy means a way of writing.

Next we asked which part of the cipher, exactly, pushes a text into the corner. We built a grid of synthetic ciphers on Latin and Turkish: k signs per letter against h variants per letter. The answer came out clean. Verbosity - two signs per letter - drops the entropy of Latin from 3.15 to 2.11 and of Turkish from 3.19 to 2.19 with the vocabulary unchanged. That is the corner. Homophony - several variants per letter - does the opposite: it raises entropy (2.11 → 2.78 → 3.40 → 3.68 at one, two, four and six variants) and inflates the vocabulary to 0.76-0.90. The manuscript's point is verbosity plus restrained homophony: few variants, reuse, anti-repetition.

The nearest miss: Ge'ez - and how it fell apart

In the first version of the test the language closest to the corner was Ge'ez, the Ethiopian liturgical language: 2.671 bits at a vocabulary share of 0.503, 0.07 bits from the boundary. In October 2026 we took it up again and found the cause: in the old corpus the word divider of the Ethiopian script had been transliterated as a syllable, so 92 % of the words carried the same tail. That tail was what lowered the entropy.

On a clean corpus Ge'ez sits 0.16-0.30 bits from the corner, and in its own script - one fidel sign per syllable - 1.3-1.75 bits away, further than almost any language. Used as plaintext under our cipher it behaves like Latin and Coptic. The miss was an artefact of the corpus - and a lesson: a match with a language is first checked for an artefact of the writing.

What the neighbours confirmed

In August 2026 De Stefano reached the same conclusion independently (the "entropy - productivity frontier", Zenodo): with different axes, with the same requirement of two linked components. In September we ran our numbers through the Lux test bench - three transliterations against a grid of spacing policies. The numbers survive a change of transcriber: two independent transliterations in the EVA alphabet give entropies of 2.248 and 2.252, and the corner holds in all 36 cells. But they do not survive a change of alphabet: the v101 transliteration raises entropy by 0.40 bits and the manuscript leaves the corner (2.648 at 0.360). The corner is a statement in EVA glyphs, and we now write the unit of measurement beside every number.

What follows from it

The empty corner is not proof of a particular model but a constraint on a class. Any theory of the manuscript must explain how the text ended up where languages never go. Simple substitution does not explain it. Nonsense with tuned statistics explains the corner but fails other checks (see Autocopying). A verbose cipher explains the corner by itself, without tuning, and along with it the word length, the doublets and the clustering. That is why we work with it - see Six tables and five scribes.

For understanding: the test speaks about the way of writing, not about the language and not about the meaning. Success would be finding a natural language in a historical orthography that lands in the corner by itself, without a cipher - in two years none has turned up.

← All findings