WedgewiseCuneiform archive experiment

About and credits

A preview for careful reading

Wedgewise is a one-day experiment by Dominik Lukeš, AI Competency Centre, University of Oxford. It brings source readings, visual comparison and unchecked machine drafts together so that specialists can inspect what the methods do and where they fail.

Data and photographs

Catalogue records and photographs are credited to the Cuneiform Digital Library Initiative (CDLI). Signs and transliterations are drawn from CuneiML v1.2, licensed CC BY 4.0, which renders CDLI readings. Source readings and break markers are retained, rather than silently reconciled with newer catalogue inscriptions.

Photographs load in your browser directly from cdli.earth. No photograph files are included in this site. Photographs belong to their holding institutions; each tablet page names the collection where the catalogue is available and links to CDLI. A missing catalogue is shown explicitly.

How the sample was drawn

The sample contains 600 unique tablets selected from the first 4,096 CuneiML v1.2 JSON entries, with deterministic seeded-hash ordering (seed 20261119). Nine candidates were backfilled: five photographs above the local acquisition limit of 32 MiB and four records without transcription. The fixed frame and size limit introduce bias. This is not a representative sample of CDLI, periods or genres, and the complete publisher-file checksum was not verified for this bounded frame.

Vectors and the map

Embeddings use google/embeddinggemma-2, pinned to revision 914f7f89142e33e77833254d9c9b90c3cef7303b. Existing 768-dimensional, L2-normalised image and text vectors are reused; there are no model calls when the site is built or viewed. Each neighbour row compares only its own vector kind, excludes the selected tablet and shows eight nearest tablets within this sample. Cosine similarity is not a confidence percentage.

Shared PCA fitted to all image and text vectors in the existing four-archive index; two axes retain 28.2% of total variance. Most dimensions are omitted, and proximity does not establish meaning or historical relationships.

Machine drafts

600 of 600 tablets currently have machine drafts. The models represented are: claude-sonnet-5-5. Every displayed model statement carries a visible label and its model name. Drafts have not been checked by a specialist. English search and people, goods and document-type browsing use these drafts, so missing drafts leave those features incomplete. A rebuild picks up newly completed local records; the deployed files do not update themselves.

Corrections welcome

Use the CDLI links to compare each source record. Please treat English drafts as material for correction, not as published translations.