The Voynich Ninja

Rod Kinnison published a You are not allowed to view links. Register or Login to view. at Cryptologia: "The Voynich Manuscript’s positional entropy collapse: comparative evidence from languages and ciphers"

You are not allowed to view links. Register or Login to view.

Quote:The Voynich Manuscript exhibits unusually low conditional character entropy and rigid positional constraints that have resisted explanation within linguistic frameworks for more than a century. In this study, I apply a leakage-free, folio-segmented, character-level analysis to the Voynich Manuscript and to a set of natural-language, constructed-language, and cipher control corpora. The analysis measures bigram and trigram bits per character, the marginal gain from higher-order context (ΔBPC), positional cross-entropy within words, and local lexical overlap under minimal edit distance. All experiments employ strict recto–verso train–test separation, transcription-layer isolation, and identical preprocessing across corpora. The results show that Voynichese exhibits a pronounced collapse of medial positional entropy and a consistently small ΔBPC. These properties persist across manuscript sections and transcription layers and are not reproduced by natural or constructed languages under identical conditions. Applying the same analysis to a verbose substitution cipher derived from a Latin source text demonstrates that this cipher reproduces the Voynich statistical profile with high fidelity. Shuffle controls confirm that this profile is destroyed when internal adjacency is randomized, indicating structured generation rather than sampling artifacts.

Thanks Torsten!  This is my first peer reviewed paper to get published.  Taylor & Francis gave me a few free eprints to hand out.  If anyone is interested in reading the full paper and would like one just PM me.

Nice work, Rod..

(09-09-2026, 06:50 PM)DG97EEB Wrote: You are not allowed to view links. Register or Login to view.Nice work, Rod..

Thanks Ed!

Have 2 more in their queue but it's so nice to have one more item checked off the bucket list.  And I didn't know this at the time but
Cryptologia has an 11% acceptance rate so I guess I really screwed up and did something right.

Rod

Where can I find which „natural languages“ were involved into comparisons etc.?
Is there a list or such?

(09-09-2026, 08:40 PM)Stefan Wirtz_2 Wrote: You are not allowed to view links. Register or Login to view.Where can I find which „natural languages“ were involved into comparisons etc.?
Is there a list or such?

Unfortunately, with my first attempt at peer review, I didn't preprint this paper.  However, it does a comparison to English, Finnish, Esperanto, Latin, Naibbe cypher and shuffled controls.  It basically uses a train/test method where it will train on, for example, odd pages and then test on even pages or train one Voynich section (i.e. recipes) and test on another (i.e. herbal).

Reference check: Kinnison (2026), "The Voynich Manuscript's positional entropy collapse"

The new Cryptologia paper by Rod Kinnison (published 7 September, doi:You are not allowed to view links. Register or Login to view.) argues that a verbose substitution cipher reproduces the manuscript's statistical profile where natural languages do not. My first step with any new paper about the Voynich manuscript is to check the references. I did the same for Layfield and Davis (2026) and put the results on Zenodo (doi:You are not allowed to view links. Register or Login to view.). The results here are of the same kind, so I am posting them in the same descriptive spirit: what the entry says, what the source says, no argument about the paper's conclusions.

The bibliography has nine entries.

1. Bennett 1971. Given as "Scientific and engineering problems of the Voynich Manuscript, Proceedings of the American Philosophical Society 115(5):395–412", and it carries the paper's opening empirical claim about low conditional entropy. You are not allowed to view links. Register or Login to view., 15 October 1971, runs pp. 335–421 and contains six items: Baumgartner 335–340, Newman 341–349, Schultz 350–365, Kaminsky 366–397, True and Núñez 398–421, plus front and back matter. There is no Bennett in the volume and nothing on the Voynich Manuscript. The page range 395–412 does not correspond to any article; it straddles Kaminsky and True. The intended source is almost certainly You are not allowed to view links. Register or Login to view., W. R., Scientific and Engineering Problem-Solving with the Computer, Prentice-Hall 1976.

2. You are not allowed to view links. Register or Login to view.. Given as "Bowern, C., and S. Lindemann. 2021. The linguistic structure of the Voynich Manuscript. Annual Review of Linguistics 7:315–37." Four things are wrong: the co-author's initial (Luke, not S.), the title ("The Linguistics of the Voynich Manuscript"), the pages (285–308), and the missing DOI.

The content claim also does not hold. The citing sentence says the positional rigidity "remains stable across manuscript sections and transcription conventions and is not readily explained by known linguistic typologies, including agglutinative or highly regular constructed languages." 

Of the four assertions there:

• Stable across sections: they report h2 = 2.17 for Currier A and 2.01 for Currier B, attribute the difference to the qo- and -dy affixes, and note the two values converge (2.23, 2.24) only after those affixes are removed.
  
• Across transcription conventions: they state that conditional entropy is to some extent affected by how characters are divided, and give the -iin case (three characters in EVA, one in Currier). Their claim is that character division cannot lower entropy to Voynich levels, which is a different and narrower statement.
  
• Agglutinative and constructed languages: the word "agglutinative" does not occur in the paper. Their conlang discussion covers Balaibalan, Lingua Ignota and Enochian as hypotheses about the manuscript, not as controls. The Figure 4 sample is natural languages in alphabets and abjads. The clause names precisely Kinnison's own two controls, Finnish and Esperanto.
  
• Not readily explained by known typologies: this one is supported. They do write that the entropy of Voynichese is unlike that of any other language or script.
  

3. You are not allowed to view links. Register or Login to view. 2013. Cited twice, and in my view this is the substantive problem. First, for the claim that quantitative studies converged on properties that "sharply diverge from those of known natural languages"; second, that "later information-theoretic studies placed Voynichese as an extreme outlier in comparative entropy analyses of written language."

They performed no such analysis. Their measure is an information quantity in bits per word describing how the distribution of a word across text partitions identifies those partitions. There is no conditional character entropy in the paper, no bigram or trigram analysis, no comparative entropy ranking of scripts. And their comparative result is the reverse of what is attributed: the Voynich maximum falls slightly above English and well below Chinese, peaking at a scale close to the natural-language examples, while the yeast DNA sequence and the Fortran source code fall outside the language cluster. Their abstract states that the manuscript's word organization is compatible with real language sequences and that this supports a genuine message in the book.

Their Discussion additionally states that any fabrication model must explain how the linguistic-like structures arose, and that the statistical structure requires an explanation going beyond local features such as word forms and local word sequences. Kinnison's evidence is entirely word-internal character adjacency and edit distance between neighbouring word forms.

4. You are not allowed to view links. Register or Login to view.. The entry is correct (NSA Technical Journal 12(3):41–85; the declassified scan is on media.defense.gov). The use is not. It is cited as work "subsequent" to Bennett 1971 that "confirmed that this effect persists across different transcription systems and analytical methods."

Tiltman's paper is an expanded version of a talk to the Baltimore Bibliophiles on 4 March 1967, which he describes as an introduction for readers coming to the manuscript for the first time. Contents: provenance and the Marci letter, the Newbold, Feely and Strong solutions, the text of his 1951 report to Friedman, digressions on Cave Beck and on the history of herbals, thirty-one plates, and a three-item bibliographical note. There is no entropy calculation and no frequency table. The 1951 report is qualitative — positional behaviour of the commonest symbols, roots and suffixes, the A-groups — and he opens it by saying there is no opinion in it to which he cannot himself find plenty of contradiction.

On transcriptions: the paper contains one, his own, in which he reduced the script to 17 symbols with arbitrary letters and figures, similar to a scheme Friedman had used. That is not a comparison across systems. This is hardly surprising since multiple parallel transcriptions of the whole text did not exist in 1967. Currier's dates from the mid-1970s; Takahashi's, Stolfi's and EVA from the 1990s. Three of the four transcriber layers in Kinnison's own robustness table (C, F, H, U in the Landini–Stolfi Interlinear) postdate the source credited with the finding. Conditional-entropy analysis of Voynichese begins with Bennett in 1976.

5. You are not allowed to view links. Register or Login to view.. Cited for the claim that "character distributions at word boundaries differ markedly from those in medial positions, with medial transitions exhibiting especially high predictability." I have not been able to consult the article and make no claim about everything it contains. What can be established is how the field cites it. Bowern and Lindemann invoke Landini (2001) for higher-level document structure and, jointly with Reddy and Knight, for the finding that Voynich word frequencies follow Zipf's law, illustrated by a rank-frequency chart. Montemurro and Zanette write that Landini showed the collection of tokens satisfies Zipf's law, with a smooth, approximately inverse relation between occurrences and rank, and group this with the token-level rather than the character-based studies. Both of these papers are in Kinnison's own bibliography. The standard citation context for Landini 2001 is word-frequency distribution; I have not found it cited anywhere for word-internal positional character statistics, and its method — spectral analysis — targets periodicity and long-range correlation rather than position within words. Anyone with the Cryptologia article to hand is welcome to correct this.

6. Zandbergen 2025. Cited as the source of the transcription: "All Voynich analyses in this study use the LSI-format transcription of the Voynich Manuscript (Zandbergen, 2025)." The transcription used is Takahashi's, carried as transcriber layer H in the Landini–Stolfi Interlinear file, which is distributed in IVTFF format. Takahashi is not cited anywhere in the paper. The entry for Zandbergen gives no URL, no file version and no access date, although the interlinear file is periodically revised and its contents determine which transcriber layers are available for the robustness check in Section 4.

7–9. You are not allowed to view links. Register or Login to view. 1978 is cited for the folio ranges of the sections, which needs no source but is not wrong. You are not allowed to view links. Register or Login to view. is correct. "Hart, n.d." is a citation to You are not allowed to view links. Register or Login to view. as an institution rather than to any text: the four control corpora appear only as bare identifiers in table headers (pg57184, pg18837, pg11940, 345-0), in two different formats, with no titles, authors, lengths or genres, and they are not among the files listed in the data availability statement.

On what is missing. Six references. Absent are You are not allowed to view links. Register or Login to view., You are not allowed to view links. Register or Login to view. and You are not allowed to view links. Register or Login to view. — the generative-mechanism literature, most of it in the same journal — in a paper whose stated aim is to narrow the space of plausible generative explanations. Also absent are You are not allowed to view links. Register or Login to view., the entropy paper whose results are attributed to the review paper of Bowern and Lindemann 2021, You are not allowed to view links. Register or Login to view., and You are not allowed to view links. Register or Login to view., although the discussion contains a paragraph on what Currier A and B represent.

I am aware I am an interested party regarding one of the omissions and I raise it as a matter of coverage, not priority.

Summary. One entry does not exist as printed. One is wrong in four particulars, with a page range that again matches no article. Three are used for claims their sources do not support, one of them in the opposite sense. One is not a citation to a text. Two are sound.

All of this is checkable in a few minutes by anyone with the volumes to hand, and all of it is correctable in an erratum. I make no claim here about the paper's methodology or its conclusions; that is a separate discussion, and I may come back to it.

(09-09-2026, 10:47 PM)Torsten Wrote: You are not allowed to view links. Register or Login to view.All of this is checkable in a few minutes by anyone with the volumes to hand, and all of it is correctable in an erratum. I make no claim here about the paper's methodology or its conclusions; that is a separate discussion, and I may come back to it.

Timm,

Jesus Christ. You spent all that time counting pages and inventing "errors" like some unpaid librarian with a grudge. The Bennett citation was a typesetting error for the 1976 book; the rest of the points are either minor or disputable.  The paper stands on the data and the method. If you have a substantive critique of the entropy results or the cipher model, put it in writing with numbers. Otherwise this is just bibliography nitpicking.

Rod

(09-09-2026, 11:49 PM)Dunsel Wrote: You are not allowed to view links. Register or Login to view.Timm,

Jesus Christ. You spent all that time counting pages and inventing "errors" like some unpaid librarian with a grudge. The Bennett citation was a typesetting error for the 1976 book; the rest of the points are either minor or disputable.  The paper stands on the data and the method. If you have a substantive critique of the entropy results or the cipher model, put it in writing with numbers. Otherwise this is just bibliography nitpicking.

Rod

When it comes to undeciphered writing systems, nitpicking is the key since details matter. One clarification: a typesetting error changes a digit. It doesn't transform a 1976 Prentice-Hall book into a 1971 journal article in the Proceedings of the American Philosophical Society with a fabricated volume, issue, and page range.

(10-09-2026, 12:11 AM)Torsten Wrote: You are not allowed to view links. Register or Login to view.When it comes to undeciphered writing systems, nitpicking is the key since details matter. One clarification: a typesetting error changes a digit. It doesn't transform a 1976 Prentice-Hall book into a 1971 journal article in the Proceedings of the American Philosophical Society with a fabricated volume, issue, and page range.

Timm,

You asked me for a free eprint of my first published paper. I gave you one.

Your response was not to test the analysis, challenge the numbers, rerun the code, examine the controls, or show that the cipher comparison fails. You explicitly said you were not addressing the methodology or conclusions. Instead, you produced a public bibliography prosecution and even managed to include your own uncited paper among the supposed omissions.

That tells me quite a lot.

You also wrote that all of this is “checkable in a few minutes by anyone with the volumes to hand.” Not everyone doing independent research has a university library in their backyard, institutional subscriptions, or immediate access to every historical volume you happen to have available. That is one of the reasons independent researchers rely on public archives, digitized sources, repositories, and whatever tools they can get their hands on.

Let me make this clear so there is no misunderstanding. I am a truck driver. I am not an academic. I am not even college educated. I do not have an institution behind me, a research library at my disposal, or a department full of colleagues to lean on.

And yet I still managed to do the work, make it reproducible, survive peer review, and get published.

I work with the tools I can scrounge.

So spare me the holier-than-thou implication that everyone can simply walk over to a shelf and verify every historical citation “in a few minutes.” Some of us actually have to build the ladder before we can climb it.

Bibliographic details can be corrected. That is what errata are for. That does not immediately demand a public crucifixion. What you have not shown is that a single reported entropy value is wrong, that the train/test procedure leaks, that the natural-language controls overturn the result, that the shuffle controls fail, or that the Naibbe comparison does not reproduce the measured Voynich profile.

The paper makes its contribution through reproducible experiments, not through ceremonial obeisance to every Voynich paper you think should have been cited. The code and data are available. If you believe the result is wrong, falsify it.

If your real complaint is that I did not cite you, say that plainly. But do not dress a bibliography audit up as though you have dismantled the work when you have deliberately avoided engaging the work itself. 

You, sir, are why I will no longer do research on the Voynich. You, sir, are why I will stop posting on this forum. You read my paper not to understand, but to respond. And you just provided the perfect example. You wrote an entire dissertation on the references rather than the work. And that is too often what passes for discussion on these pages.

You had the paper. You had the data. You had the code.

And you chose the fucking footnotes.