# I asked AI to read Voynich. The harder test changed the story.

Research case study by Luana Micheau, 15 October 2026.

The Voynich manuscript has no accepted decipherment. We used a published EVA transcription to test whether glyph-label patterns could predict a folio's illustrated section. EVA describes shapes. It is not a translation.

Held-out folio comparisons reached 0.8641 balanced accuracy for herbal versus pharmaceutical pages and 0.9500 for herbal versus biological pages, with scribal hand and Currier category held constant. A smaller herbal versus pharmaceutical comparison that also held quire constant reached 0.5000, chance by this measure, with only two folios per class.

The result does not prove there is no subject signal. It shows why high scores alone do not establish meaning. We did not identify a language, cipher key, author or translation.

Read the full case study at https://www.luana.systems/writing/voynich-ai-claim-check and inspect the code, method and predictions at https://huggingface.co/spaces/tropicanaca/voynich-ai-claim-check.