13-05-2026, 10:20 PM
Working on the recipe section (folios 103r–116r), we found that ch and sh appear to carry two independent signals simultaneously.
The first is paragraph-level structure. A logistic regression using ch/sh-based features predicts tail vs no-tail paragraphs at balanced accuracy 86%, confirmed by within-folio permutation testing.
The second is folio-level language. The ratio of (ch sh)o-type words predicts which Currier language a folio belongs to with ~99% accuracy manuscriptwide, even though all recipe folios are Currier B. The cho/che balance appears to encode a stable folio-level property operating above the paragraph.
The key thing is that these two signals are orthogonal.
The first is paragraph-level structure. A logistic regression using ch/sh-based features predicts tail vs no-tail paragraphs at balanced accuracy 86%, confirmed by within-folio permutation testing.
The second is folio-level language. The ratio of (ch sh)o-type words predicts which Currier language a folio belongs to with ~99% accuracy manuscriptwide, even though all recipe folios are Currier B. The cho/che balance appears to encode a stable folio-level property operating above the paragraph.
The key thing is that these two signals are orthogonal.