This project tests what happens when TRIBE, a brain-predictive AI model trained on fMRI responses to contemporary video, is applied to premodern manuscript pages. It does not try to predict how medieval readers saw a page. Instead, it uses the model as a diagnostic tool. Where the model's outputs stay meaningful, and where they break down, shows which features of manuscript visual organisation (diagrams, tables, illuminations) fall outside what the model learned. The data are several hundred folios from 40 manuscripts at the Schoenberg Institute for Manuscript Studies (University of Pennsylvania, via OPenn/Colenda), supplemented with CSMC holdings, and they span Mesopotamian, Egyptian, Islamicate, Hebrew and European traditions. The project asks whether the model can tell manuscript types apart, at which stage of its processing it does so, and where its groupings agree or conflict with established codicological categories.

The project relies on computer vision. Each folio is rendered as a short silent video and passed through TRIBE. Outputs are taken at three points: the raw visual features of its V-JEPA backbone, the brain-predictive embedding, and the predicted cortical maps. Layer-by-layer feature analysis and contrast maps compare categories such as diagrammatic and text-dominant pages, or Arabic and Latin scientific material. Controls on resolution and cropping separate responses to the page from responses to the scan. This gives a scalable way to compare page design across traditions without first sorting by culture or period. When the model groups folios from different traditions, it points to shared structure. When it fails to group folios from the same tradition, it suggests that their visual identity lies in learned conventions, such as hand, reading direction and decoration, rather than in spatial layout.

Davide Pafumi and Hussein Mohammed
starting date: January 2027