Reconstructed letters for each scribal hand, averaged from glyphs cut out of FJMS page images using the aligned transcriptions. Each template is the robust mean of the hand’s instances of that letter, size-normalized to a 64×96 canvas. Hands are labelled by shelfmark; one shelfmark can contain several scribes.
Consistency: median IoU of a hand’s instances to its own template. Held-out fit: median IoU of instances on an unseen page to templates built from the hand’s other pages. Hatched cells: fewer than 5 instances.
Positions from multidimensional scaling of 1 − IoU between aligned templates. Nearby letters look alike.
Each card: pixel IoU → IoU with thickness removed → IoU with geometry (slant, proportion, rotation) removed, then the structural residual: mean distance between the two centrelines after the geometry fit. same form means the residual is within this letter’s within-hand range (90th percentile of the distance between two halves of one hand); different form means it exceeds it. blurry marks a template averaged from mixed forms, whose structure cannot be trusted. Click a card to make it the reference.
Hands grouped by how their letters are constructed: centrelines redrawn at one width and aligned by slant, proportion and rotation, so pen weight and photography matter little. Each type has a prototype alphabet (the aligned mean of its members’ letters). Each scribe is then described as prototype + its own variation. resolvable: the scribe differs from the prototype by more than it differs from itself (two halves of its own pages). distinctive: the scribe departs from its type much more than the type’s other members do. Types are unsupervised and unnamed; the grouping is tentative (see stability).