# Cuneiform: research note

Research snapshot: 2 October 2026.

This note reports sign-recognition experiments. The wider project also has basic
translation capability; the results below do not measure translation quality.

## Sign-recognition results

The sign-recognition pipeline detects possible signs inside a supplied tablet-face region,
classifies the resulting crops, and retains original-image coordinates and ranked
alternatives. The frontsite demonstrates recorded output on tablet P335575.
These are provisional visual sign identities, not phonetic readings or translations.

The baseline detector locates signs with 71.9% precision and 71.3% recall on a
reused research benchmark of 51 face views from 30 tablet groups. Combined detection
and recognition correctly locates and identifies 39.3% of labeled signs as its
first choice, or 52.1% within five candidates. These measures include missed signs.
They are not estimates of corpus-wide transcription accuracy.

Later short localization experiments explored sign scale, normalization, extra
manual annotations, and weak supervision. The best observed localization F1 was
77.12%, averaged across two seeds. This is an exploratory test result, not a new
production reader or an independent benchmark. The existing default remains in place.

## What comes next

Improve sign isolation across collections and imaging conditions, establish
reliable reading order, and improve transliteration and basic translation alongside
visual recognition. Scholarly review and traceable evidence remain central.

## Reproduction and evidence

This summary derives from the project's DETECTION.md and LOCALIZATION_TRIALS.md.
The recorded example retains model proposals, original-photo coordinates,
and the source photograph's SHA-256. The model scores are not calibrated
probabilities of correctness. Rankings are shown without probability claims.

The full training code and checkpoints are not distributed in this frontsite draft.

## Source work

- CDLI tablet record: https://cdli.earth/P335575
- Dencker et al. (2020): https://doi.org/10.1371/journal.pone.0243039
- Open Richly Annotated Cuneiform Corpus: https://oracc.museum.upenn.edu/

Image and dataset terms remain those of the originating institutions and projects.