What the harder test revealed.
A program found patterns in an undeciphered manuscript. Then a stricter comparison changed what we could honestly claim.
The finding
We could distinguish some illustrated sections from patterns in the transcribed glyphs. Within the same quire, the signal fell to chance. That control prevents us from claiming we found word meanings.
The same-quire sample contained only two folios per class. This is a cautionary result, not proof that subject matter has no relationship to the writing.
Three comparisons
Scores are balanced accuracy, the average recall of the two classes. 0.50 is chance by this metric. They measure section prediction, not decipherment.
What we actually did
We used line-level EVA transliteration, counted glyph-label tokens in paragraph text, and held out entire folios. We compared pages whose metadata listed the same scribal hand and Currier category. A third test also kept quire constant.
EVA is a labeling system for visible signs. A token that looks like a word has no established meaning.
The result is exploratory. Section, layout, manuscript position and copying history may overlap. The same-quire check is small, and we could not run one for the biological pages in this dataset.
Earlier topic-modeling research already found that text clusters track a combination of scribe and illustrated subject. We do not claim to be first to see a relationship.
What would move us forward?
Repeat the experiment with an independent transcription, stronger controls for quire and layout, and larger comparable samples. Any proposed translation must predict unseen passages and agree with evidence outside the manuscript.
The text remains undeciphered. The useful contribution here is a checkable method for challenging confident claims.