The Babble Engine.

Last time we measured the Voynich’s structure and refused to read it, because the same structure fits a language, a cipher, or a machine that means nothing. That last option deserved more than a word. So we built the machine. Not a cipher, not a theory — a dumb generator with three glyphs of memory and a fixed coin, no dictionary, no grammar of meaning. Then we asked it to forge the most famous unreadable book in the world, and we scored the forgery against the original with our own code. If a machine this stupid can’t fake the manuscript, the manuscript has real depth. If it can, that depth was never evidence of meaning.

Where we land: still open — on purpose. A meaning-free text generator reproduces every Voynich signature, which retires the old “too structured to be a hoax” argument — without proving the book meaningless. Both crowns stay off.

1 · The machine

It is an order-3 Markov model over glyphs — a fancy name for a very simple habit machine. Train it on the real Voynich word-shapes and all it learns is: “after these three letters, here is how often each letter comes next.” To write a word it starts blank and rolls its loaded dice, glyph by glyph, until it rolls a stop. That is the whole mind of it — local letter habits, nothing more. No meaning can hide in three letters of memory. We seed the coin (408, the manuscript’s shelf-mark) so every number here is reproducible.

2 · The test, written down before the run

Before running anything, we wrote down in advance what result would count against us. The signatures that make the Voynich look too structured to be a hoax are its word length, its low entropy, its positional grammar, and its Zipfian vocabulary. The prediction: if the manuscript’s mystery really is just low-order letter habit, this toy reproduces all four. If instead the book carries structure a three-glyph memory can’t hold, the forgery breaks on at least one — and that break would be the first hard evidence of depth. One honesty flag up front: entropy h₂ is a two-letter statistic, so any model with at least one letter of memory matches it by construction. The real tests are the ones the machine was never shown: the length curve, the Zipf slope, and how many real words it accidentally rebuilds.

3 · The forgery holds

It reproduces all four, to within a few percent — on signatures it was never fit to.

4 · Read them and guess

Twenty words from the manuscript, twenty from the machine. One of these was written by a human being six hundred years ago and has defeated every cryptographer since. The other is dice.

The machine (order-3, no meaning): tchody al olkeedar shos okeey arar qoteol or taiin cthalody qkain arodly daiin daiin dain oteopcheodain cheopchedy ckholdy oly aiin

Beinecke MS 408 (real): fachys ykal ar ataiin shol shory cthres y kor sholdy sory ckhar or y kair chtaiin shar ase cthar cthar

The machine rebuilt 37% of its word-types into actual Voynich words — including daiin, the manuscript’s commonest word — having only ever seen three-letter fragments. The rest are new, and you cannot tell which.

The turn — and why we still don’t call it a hoax

Here is where every “the Voynich is a hoax” headline overreaches, and we won’t. A machine matching the manuscript proves the structure is reproducible without meaning. It does not prove there is no meaning — because a real language that happens to be highly regular (heavy abbreviation, a syllabic script, rigid morphology) would land on the same low-order statistics too. Low-order structure is silent both ways. What the engine actually retires is one specific claim — the decipherers’ favourite: “the text is too structured to be meaningless.” That argument is done, by construction: this much structure is exactly what a meaningless three-glyph habit produces. The manuscript may yet mean something. Its statistics are not the evidence, in either direction.

The structure is cheap — a coin with a three-letter memory buys all of it. And cheap structure can’t testify. It cannot say “I mean something,” and it cannot say “I don’t.”

One place the toy does fall short, kept on the page: it was trained on all words at once, so it does not reproduce the two-dialect split (Release 016) or the section-specific vocabularies — the manuscript carries a little structure above the third order. But a mode-switching generator handles that with one more coin, and it is still not meaning. The order sweep shows the honest trade: raise the memory and the machine matches better only by memorising more real words — the structure it invents is always cheap.

The sweep, order by order (h₂ · Zipf slope · share of real Voynich words rebuilt): order-1: 2.16 · -0.59 · 16%. Order-2: 2.16 · -0.80 · 29%. Order-3: 2.16 · -0.85 · 37%. Order-4: 2.16 · -0.92 · 48%. The real manuscript: h₂ 2.15, Zipf slope -0.89.

Why our math sees more

Because it will not take either bribe. The decipherer collapses the manuscript to a meaning; the debunker collapses it to a hoax; both claim an answer the evidence doesn’t earn. We built the cheapest possible machine that fits the data and stopped exactly there — a test that could have exposed real depth, run in public, coming back “cheap.” That is the whole method: measure how much structure the mystery actually holds, and refuse to read a verdict it cannot support. The Voynich’s order is genuine, and it is three letters deep. What it means, if anything, is still nobody’s to claim.