Figures shown: page images and complexity ratings — no model results
Page generated: 2026-09-15
Data (model runs) generated: 2026-04-11
Scoring inputs: ground truth cc2ed8f5, evaluate.py 845b70e9
Thumbnails of all 39 pages with their complexity ratings and mean accuracy under the guided condition. (Per-page and per-model comparison of guided vs naive accuracy.) (Complexity vs accuracy correlations and per-page ranking.) Click a thumbnail to view the full-size image. Structural complexity = sum of named travelers, unnamed travelers, con instances, e instances (from ground truth). Visual density = connected ink components from Sauvola binarization (≥5 px). Overall complexity and density = mean of structural complexity and visual density after each is min-max normalized to 0–100% of the corpus range. (Pearson r vs mean accuracy: structural complexity −0.75, visual density −0.77, overall −0.81.)
Accuracy (raw %), structural complexity (min-max normalized to 0–100% of the corpus range, 7–118), visual density (min-max normalized, 423–2,695 components), and overall complexity and density (mean of the two normalized values). Toggle traces in the legend. Note the inverse relationship: higher complexity → lower accuracy. Overall complexity and density correlates more strongly with accuracy than either component alone.