Shaping History at Scale
Computational Diplomatics and the Diachronic Evolution of Cuneiform Tablet Morphology
Shaping History at Scale
Computational Diplomatics and the Diachronic Evolution of Cuneiform Tablet Morphology over Three Millennia
Shai Gordin · Danielle Kapon Epshtain · Michael Fire Ariel University · Ben-Gurion University of the Negev · Open University of Israel
94,936 tablet images
21 historical periods
3,300 years (3500–64 BCE)
86% shape info retention
About this site
This companion website provides interactive access to all analyses published in:
Gordin, S., Kapon Epshtain, D., & Fire, M. (2025). Shaping History at Scale: Computational Diplomatics and the Diachronic Evolution of Cuneiform Tablet Morphology over Three Millennia. PNAS. DOI: [pending]
Each section corresponds to a section of the paper and a Jupyter notebook in the code repository. The two Interactive Tools pages provide live access to the VAE feature explorer and interpolation widget via Binder — no installation required.
Analyses
Height-to-width ratios across 21 periods. The three-millennium portrait→landscape shift. Log-ratio methodology.
Coefficient of variation as a proxy for institutional control. From Ur III (CV = 0.251) to Neo-Babylonian (CV = 0.807).
Decision tree → CNN → ResNet50. Shape information retention fraction: 86%. Confusion matrices and per-period performance.
12-dimensional VAE bottleneck. Discrimination (X7) vs. trend (X2, X8). Diplomatic genealogies in latent space.
What does each VAE dimension physically control? Filmstrip traversals with genre control and geometric descriptor analysis.
Is the trend real or a sampling artifact? Variance decomposition, zone analysis, site-level panel, maps.
Interactive Tools
🎛️ Try it yourself — no installation required
The two interactive widgets require only a browser. Click Launch Binder to start a cloud Jupyter session with the trained VAE model pre-loaded (~2 min to start).
Adjust individual VAE dimensions with sliders and watch the decoded tablet silhouette change in real time. Filter by period and genre.
Select two tablets or two latent vectors. Visualize the morphological path between them across 11 decoded waypoints.
Data & Code
All code, data, and the trained VAE model are publicly available:
| Resource | Link |
|---|---|
| Code repository | github.com/DigitalPasts/ShapingHistory |
| Dataset (VAE encodings, ratios) | Zenodo DOI: 10.5281/zenodo.12787745 |
| Trained VAE model | Zenodo DOI: 10.5281/zenodo.11668219 |
| CDLI Catalogue | cdli.mpiwg-berlin.mpg.de |
| DigitalPasts Lab | github.com/DigitalPasts |