Release 016 found the Voynich splits into two dialects— Currier A and Currier B, straight from the transliteration’s own metadata. 017 built a meaning-free machine that fakes the whole manuscript from three letters of memory. This one asks the question those two left open: the A→B shift is real — but is it a change of language, or the same engine drifting? A raw distance between them means nothing on its own, so we don’t report one. We build a ruler.

Where we land:held open. The A-to-B seam is real but only about a third of the way to the nearest language boundary we measured — far smaller than any language gap on our ruler; a blind, control-gated 10-model panel (LLMs, not independent of shared training) agrees — the six that cleared the gate, six to nil. Measured seam, unread system.

1 · A ruler, not a number

One instrument, three kinds of mark, all measured in the same unit — D, the extra bits per letter a foreign character-model pays over the nativeone (own-code interpolated add-α n-gram, every model capped to the same word budget so nothing looks closer just from more data):

• The floor— split Currier A into two halves, Currier B into two halves, and the meaning-free babble engine into two batches. These are one engine, two samples. They should read zero. Reported first, free-to-fail.
• The rungs— every pair among English, Latin, German, French. These are genuine language changes, close (same family) to far (cross family). This is what a real language boundary costs.
• The measurement— Currier A | B. Where does it fall between the two?

2 · What the ruler reads

Two facts, earned and marked. One: A and B are notthe same sample — at 0.273bits/char they sit far above the same-engine floor, so the A→B shift is a genuine structural change, not noise. Two: that change is far smaller than any language gap we measured.The closest pair of real languages we could find — French | Latin, which already share an alphabet and a deep common root — still costs 3.2× more. A and B are nowhere near that far apart.

And it holds when we deepen the instrument — raising the model from 2 letters of context to 3 pushes every distance up, but A | B stays low in the gap:

The read is the same both ways: whatever separates Currier A from Currier B, it is dialect-distance, not language-distance— one generative system re-tuned between scribes or sections, not two different tongues sharing a script.

3 · A second light

The ruler is the machine light. For a different kind of light we ran a blind cross-model panel— ten non-Anthropic models (OpenAI, Google, Meta, Mistral, DeepSeek, xAI, Alibaba) via the Vercel AI Gateway, each shown the same transliterated word-samples, pair order shuffled, the real question hidden among controls. Own-code (grade_gateway.py), and hard-gated: a model’s call on Currier A | B counts only if it first separates the controls — the same-split pair as one system, the two known language pairs as two. The gate did real work, discarding four models that failed in both directions (two over-split the same-system control; two called genuine languages one system). Of the six that got every control right, all six independently called Currier A | B ONE system re-tuned — six to zero, the same direction the ruler points. Two lights, two kinds, one read.

Why our math sees more

Because it refuses to read a distance in a vacuum. “Currier A and B differ” is true and empty; the whole question is differ by how much, next to what.So we built the scale before we took the reading — a same-engine floor that must land on zero, real language changes for the far end — and let the measurement fall where it falls, in public, on an instrument that could have put A | B out with the languages and didn’t. That is the method: measure the dark against marks you laid down first, and read only what the ruler can hold.

Sources & method