WedgewiseCuneiform archive experiment

Limits are part of the result

Where it goes wrong

Finding signs is not reading them

The embedding model does not read cuneiform for this demonstration. Exact sign search works by matching Unicode strings; a nearby vector does not establish the meaning of those signs. The earlier four-archive probes found native-script meaning-search failures for hieroglyphic and Tibetan sequences. Those specific probes are not a cuneiform decipherment test. Here, English search matches words in machine drafts; it is not meaning search over signs.

Photographs can organise the neighbours

Image neighbours can follow background, aspect ratio and the arrangement of tablet views. In the earlier seeded image audit, all 15 within-archive neighbour positions stayed in their own archive. Forced cross-archive matches shared visual form, such as brown surfaces or thin line crops. Neither those neighbours nor the dots on the map establish historical relationships.

English can sound more certain than it is

Every English draft is unchecked machine output. Ur III administrative formulae are the easy case; readable formulae do not validate broken passages, names, quantities or other periods and genres. The model's own confidence estimate is not an expert judgement or a calibrated accuracy score.

600 of 600 tablets have drafts. Their self-reported confidence distribution is:

A fixed sample leaves things out

These 600 tablets come from a fixed 4,096-entry candidate frame, with deterministic seeded selection and a size limit on locally acquired photographs. They are not representative of CDLI or the history of cuneiform. 600 catalogue records are currently loaded; missing catalogue fields are shown as unknown and must not be inferred from a neighbour.

A map leaves most dimensions out

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.

These cautions draw on the repository's research/findings-first-demo-2026-11-19.md (19 November 2026) and the counts in this build. The original probes compared keyword, meaning and fused retrieval on ten fixed questions across four archives. They were exploratory results, not an accuracy benchmark.