1. What was measured on the Voynich side

From the saved ink masks of the neutral extraction:

• connected ink pieces became symbols, grouped into 16, 32 or 64 neutral shape classes;
• pieces sharing a baseline became provisional lines;
• gaps wider than 0.8, 1.1 or 1.5 x-heights split lines into candidate clusters;
• broken strokes were either kept separate or merged.

That gives 3 × 3 × 2 = 18 segmentation variants. Coverage: 204 pages, 45,495 provisional lines, and 14,478 lines with at least 8 pieces used. Clusters are not words and symbols are not letters.

2. How the languages were blinded and made comparable

There were 16 corpora of 16,000 words each from Universal Dependencies treebanks, historical where available (Ottoman Turkish, medieval Latin, Old Italian, Old French, Gothic, Old East Slavic, Ancient Greek, Ancient Hebrew, Classical Chinese). Where no historical treebank exists, modern ones were used (Hungarian, Finnish, German, Czech, Persian, Arabic, Basque). Every corpus got the same treatment:

• several documents or authors;
• accents and vowel points stripped;
• characters replaced by random numbers;
• a random code (C01–C16) instead of a name.

Metrics, scoring, controls and the Voynich measurements were hashed and frozen before the code key was opened. The blind conclusions were also sealed before the reveal.

Structure, not alphabet. Every word-internal measure is scored as its excess over the same corpus with its symbols shuffled. Every between-word measure is scored as its excess over the same corpus with its word order shuffled. This removes alphabet size and word length, which a first test showed would otherwise dominate.

3. Revealed comparison table

The “change excess” columns form the agglutination test. Among pairs of similar forms that differ by one symbol, they show how much more often the difference sits at the end, the start or inside than it would by chance. Suffixing agglutinative languages were expected to show end excess and root-pattern languages internal excess. Voynich sits near zero on all three in every variant.

4. Which similarities disappear under controls

• Classical Chinese (robust scoring): reproduced by noise-degraded versions of other languages (31 of 32 degraded corpora nearest Classical Chinese (isolating control); 1 of 32 degraded corpora nearest Ancient Hebrew (root-pattern control)) and by Voynich-trained pseudo-text. It disappears as language evidence.
• Ottoman Turkish (secondary all-metric scoring, #1 in 13 of 18 variants). Its advantage comes from:
  • the adjacent-repeat excess (Ottoman is the only language slightly above zero);
  • low adjacent-word mutual information;
  • high type/token ratio and hapax rate.
  Ottoman Turkish is not closer than the median language on the word-internal measures (neighbour rate, internal change, rank-frequency slope). Voynich-generated self-copying and Markov pseudo-text are also nearest to Ottoman Turkish on this distance. It disappears as language evidence.
• High hapax / type-token rate (Voynich 0.79–0.95 hapax vs languages 0.43–0.77) is reached by languages degraded at 30–50% noise, so it is explained by extraction noise.

4b. Post-reveal check: removing parchment texture (not blind)

The class gallery above shows that many frequent “symbols” were blank parchment texture. The extraction was repeated keeping only ink-dark pieces (20th-percentile brightness below 0.8 of the local background; ink and texture separate cleanly). The unchanged frozen scoring was rerun. This check is not blind, because the labels were already known.

• Frozen robust metrics: still Classical Chinese (isolating control) first in 18 of 18 variants.
• Robustness rule re-applied to the filtered data: Ancient Hebrew (root-pattern control) first in 18 of 18, holdout-stable. However, a Markov pseudo-text built from the same Voynich clusters is also nearest to it, and noise-degraded languages go to Classical Chinese or Ancient Hebrew. It fails controls.
• All-metric distance: Ottoman Turkish (Turkic) first in 17 of 18. Voynich-generated Markov and self-copying pseudo-text are also nearest to it. The drivers are again the repetition statistics. It fails controls.
• Word-internal predictability rose into the normal language range, but there is still no edit-neighbour, suffix or prefix alternation above a shuffle.

5. Robust Voynich properties, and what no language explains

• Robust across all 18 segmentations. The candidate clusters have no word-internal structure detectable above a symbol shuffle: no end, start or internal alternation pattern, no family effect and no concentration of endings. This is unresolved as a property of the manuscript, because degraded real languages look identical.
• Unexplained by every language. Neighbouring clusters are identical or near-identical more often than a corpus-wide shuffle predicts (K32/1.1/merged: near-identical neighbours 0.202 vs 0.175 shuffled). A post-hoc check shows this is line-level sameness, not word-to-word repetition: shuffling clusters only within their own line gives 0.207. Whether this reflects the writing, such as the known line-level patterns, or local ink and scale drift in the shape classes is unresolved.
• Excluded as not robust. Cluster length, type/token ratio, rank-frequency slope and adjacency statistics all changed by more than half the between-language spread across segmentations.

6. What would make this test informative

The comparison is limited by symbol quality, not by the languages. The next test is the same frozen protocol on a human-verified glyph sequence for a subset of text-dense pages, for example the reviewed f1r plus a recipe page, with the same controls. If structure then appears above the shuffle, the language-type comparison becomes meaningful.