Harvard physicist's take on the VMS

[hensing1]

I'm gonna be honest with y'all: I can't read this. Too technical or too AI?

Allright, I'm going to bite the bullet on this one. I am in no means qualified to put this into context of existing voynich research, and I am no expert in encryption or decryption of cipher texts, let alone medieval ones. But I do have a degree in computer science, which means I am used to reading fairly technical papers. I'll try to turn the AI babble into something a human would have written, with the intent that another human can read it.

Since the author is a Harvard professor, I'm going to give them the benefit of the doubt and assume that there is at least some merit to what they published (even if they say it's preliminary), and that they didn't just copy-paste some AI nonsense without ever looking at it.

So, let's trudge through the conclusions chapter and figure out what they're actually saying.

For four hundred years readers have asked what the Voynich manuscript says. We have asked instead, as others since Friedman have, what process produced its text. The questions are connected: a keyless procedure with no source text leaves nothing to decipher.

So far, so good. "Keyless procedure with no source" presumably means "random word generator", which indeed would not be decipherable.

On the first question, what the text says, our results are negative within a stated space. Where the test rejects false readings, every mechanical reading we tried, into any of eighteen languages and under either character parsing, fails on the test pages.

A convoluted way of saying "we can't translate the text either". They tried 18 languages and failed to find a decryption.

Where it passes true readings of the same kind, a failure excludes the reading. This holds at low transcription noise for substitution and fixed-rule anagramming of scripture in most languages, for Latin abbreviation and moderate Hebrew and Arabic homophony, and for substitution of Latin herbals and Latin or Italian recipes.

They have some test to check wether something was encrypted (true reading) or just generated (false reading). They encrypt some example text (latin, hebrew, etc.), and their test correctly identifies it as an encryption.

Elsewhere the test rejects even correct readings or, for wide homophonic tables and Naibbe-type ciphers, has no power, and the anagram test is uninformative for Hebrew and Arabic. Only character-level tests and the evidence for keyless production count against such readings.

Their test is a little too strict, in that it sometimes thinks a genuine encryption is actually just generated nonsense.

The space is limited in three respects: free anagramming cannot be tested at this text size and lies outside it [Newbold, 1928]; only Latin was searched at full verbose-table width; and the word-level criterion was calibrated on biblical prose, and on herbal or recipe prose only for Latin and Italian.

Some more limitations of their method. "The space" means "the area where they looked".

An unconstrained search for the most language-like reading finds several bits per word of spurious signal, which may explain the succession of announced decipherments.

"If you look really hard, there are some parts in the voynich that look like language by pure chance, which is why people keep claiming they found a translation".

On the second question, how the text was produced, we sharpen the picture that has emerged since the work of Rugg and of Timm and Schinner. A keyless writing procedure reproduces nearly all the summary statistics of the two principal hands, within its sampling noise, but not the coupling between adjacent words’ edge glyphs, which earlier generators also lack.

Their explanation (a text generator) accounts for everything apart from the dependence between the end of one word and the beginning of the next.

Each scribe’s words combine a small set of affixes, found in every hand, with a larger, mostly unshared set of cores. Yet two halves of one hand’s pages share cores little more than two hands do. Hence the contrast cannot show how many writers there were or what passed between them. Each scribe composed pages by reusing his vocabulary, with some coining, a little local copying, and no detectable dependence between words beyond the edge-glyph coupling.

Not only is there little coherence between two separate scribes, even two sections written by the same scribe on the same page have little to do with each other. Words seem to be random apart from the dependence of the edge glyphs.

Local copying, central to Timm and Schinner’s generator [Timm, 2014, Timm and Schinner, 2020], is a minor component of ours; copying from further back is indistinguishable from reuse, so their broader self-citation [Timm and Schinner, 2024, Timm, 2026a] is not excluded. Rugg’s table and grille [Rugg, 2004], as parametrized here, fits the text worse than reuse alone.

Comparison with earlier methods that I don't know about

Two weak traces of content survive controls for scribe and page order: distinctive label words and a faint subject dependence of word choice in the smaller sections. We read these traces as signs that the writers tracked the pictures, but neither trace shows that the text describes them.

Words weakly depend on the pictures next to them, but nothing else.

Our reconstruction has clear limits: it assumes the five-hand attribution of Fagin Davis [2020], which Timm [2026b] disputes. It is fitted on two of the five hands and extended to the others largely by inference, and even on the two fitted hands the complete model misses a few of the summary statistics it is scored on.

They trained their generator only on two of the five scribes. It's not perfect, not even for replicating the two scribes it was trained on.

Reference confined to single, unrepeated objects cannot be tested from a transliteration, and meaning carried by something other than word identity and word order would be invisible to every test used here.

If a word only shows up once, there is no way of knowing what it could stand for. If meaning is conveyed through something other than the words themselves and their order, this is undetectable for their test.

Why the book was made requires a further step of inference. The low cost of the writing and the genres the pictures imitate suggest a book made to be looked at and believed, as a credential or for sale to a collector. The reconstruction is falsifiable, most directly by a confirmed reading of any passage. The same tests apply to other disputed texts, such as suspected forgeries and undeciphered scripts.

They think the manuscript was made to fool a wealthy collector or merchant. If someone actually translates it, their theory is wrong.

---

Sorry for the long-ass post. I hope this helps explaining what they're on about.

On a more general note: the whole "hurr durr, just let the AI summarize it for you" is insane to me. If we have papers written by AI, and the only one who can read it is a different AI (or even the same AI), what the hell is the point? Who are these papers even for? The scientific process is about advancing the extent of human knowledge. If we publish nonsense that nobody can read, literally nothing is gained.

For what it's worth, I think chapter 5.6 ("Purposes, parallels and possible refutations") is at least a worthwile read. I'm not saying they're correct, but I think you can reasonably follow what they are saying.

-Henning