Skip to content

Cole Simmons

I'm an engineer and researcher.

I used to be a San Francisco-based technical startup founder.

Now I'm at Oxford, working on the corpus of 500,000+ cuneiform texts.

Project, 2024–

Cuneiforge

A workbench for learning to read cuneiform from the copies themselves. Every sign I identify is a box on the image, and every reading, word, line, and translation is built from those boxes.

I built it to prepare the set texts of the Oxford MPhil in Cuneiform Studies: two years of Akkadian, Sumerian, and Elamite, examined from the cuneiform. The published tools keep the pieces apart. An edition gives the transliteration without the signs, a sign list gives tidy, typeset forms, and a dictionary gives the word without the places it is written. The reading in between (these wedges are one sign, these four signs are one word) usually stays in a notebook. Cuneiforge keeps it, tied to the pixels it came from.

project

languages Akkadian, Sumerian, Elamite
texts 67 on 153 images
sign boxes 35,683
words 12,855
lines 2,803, and 119 Elamite sentences
lexicon 1,019 lemmas
sign list 279 signs
built with TypeScript, React, SQLite
status Private; runs locally

A letter from Hammurapi, read in layers

Hammurapi writes to two officials about a man the enemy has taken: they are to pay his merchant ten shekels of silver from the temple of Sîn, and so ransom him. Each of the 71 signs on the copy is a box. Boxes are grouped into 23 words, and words into 10 lines; readings, normalisations, lemmas, and translations hang from those groups.

  1. ana luštamar-zababa

    To Luštamar-Zababa

  2. u bēlānum

    and Bēlānum

  3. qibī-ma

    speak:

  4. umma hammurapi-ma

    Thus (says) Hammurapi:

  5. sîn-ana-eridu-lippalis

    (Regarding) Sîn-ana-Eridu-lippalis,

  6. mār maninum

    son of Maninum,

  7. ša nakrum ilqû

    whom the enemy took,

  8. 10 šiqil kaspam ina bīt sîn

    10 shekels of silver in the house of Sîn

  9. ana tamkārīšu idnā-ma

    give (the silver) to his merchant and

  10. puṭrāšu

    redeem him.

YOS 2, 32, obverse, in H. F. Lutz’s 1917 copy (CDLI P305758). Old Babylonian, from the reign of Hammurapi. For each line: the transliteration, its normalisation, and a translation. Of the 23 words, 11 are linked to the lexicon so far.

One click, one box

Dragging out 35,683 boxes by hand would be the whole job, so a click proposes one. Around the click, the server separates ink from paper against the local brightness, thickens the strokes by a couple of pixels so broken lines hold together, and floods outward through every stroke the click touches. That shape’s bounds, padded and snapped to the ink, are the box.

Where signs touch and the flood spills, it falls back to walking out to the nearest empty columns and rows. Copies differ, from a fine 1917 pen line to heavy modern drawings, so its settings are tuned and saved per text; 32 texts have their own. On a cramped copy like the letter, many proposals still need a drag to correct, but clicks run concurrently, so one can work down a line and fix as one goes.

  1. 1 The click

    A window 267 pixels square, centred on the click.

  2. 2 Bridge and flood

    Strokes thickened by 2 pixels; the shape under the click floods, the rest stays out.

  3. 3 The box

    Its bounds, padded and snapped to the ink.

The sign nu in line 6 of the letter, chosen because it stands clear of its neighbours, re-run through a port of the detector with the default settings. Its box matches the one stored for that sign to 98% overlap.

Forms, not fonts

Every box is linked to a sign in a shared list, so each sign’s page gathers every form of it I have drawn, grouped by period, with the readings it takes in each language. Font renderings sit beside them but are kept visibly apart: a font shows what a sign is, not how anyone wrote it. The same sign in one reign can look like two: Hammurapi’s stele and Hammurapi’s letter.

Readings are sorted per language, because the same string means different things. In Akkadian and Elamite, lower case is a syllable and upper case a logogram; in Sumerian, upper case names a sign whose reading in context is uncertain. Braces mark determinatives everywhere, and damage brackets are set aside, so a half-broken ⸢di⸣ still counts as di.

Words work the same way. A lemma’s page tallies its spellings and shows each occurrence as the word looks on its tablet. A word that wraps onto the next line is imaged from its own signs only, cut from their boxes and set at their original positions, rather than by a crop that drags in its neighbours.

  1. Gudea, Cylinder A

    Sumerian, read an

  2. Laws of Hammurapi, prologue

    Akkadian, read d (determinative)

  3. Hammurapi’s letter, YOS 2, 32

    Akkadian, read d (determinative)

  4. Annals of Sennacherib

    Akkadian, read d (determinative)

Akkadian · 537 d (determinative) 311, an 104, DINGIR 38, AN 37
Sumerian · 247 d (determinative) 112, an 94, AN 27, diŋir 14
Elamite · 65 an 65

AN, the sign for heaven and for god, from four of the copies annotated here. It is boxed 849 times in the corpus; the rows give its commonest readings, counted from boxes that hold it alone. d marks it as the determinative written before a god’s name.

From annotation to review

Reading has to become recognition. A word I flag becomes a flashcard whose front is the word as written on its tablet, with its line where the language is read in lines, and whose back is the reading, the lemma, and its sense. Every line becomes a transliteration card and a translation card. 685 words are flagged so far.

The sync to Anki is a plan, then an apply. It loads notes from the database and the notes Anki holds under the same tags, diffs them into creations, updates, deletions, and skips, and shows the plan before anything is written. Notes are keyed by record, and images are named by a hash of their contents, so a re-cut image is an update rather than a silent mismatch. A plan that would delete more than ten notes and more than a fifth of a deck stops until DELETE is typed: that is usually a loading bug, not an intention, and skipping the prompt does not skip it. Card templates and their CSS live in the repository and sync the same way.

A second reader, with limits

Parsing and translation are the slowest part of preparing a text, so for Akkadian and Sumerian I wrote agent workflows that draft them, under rules. One agent analyses each line, word by word; in Akkadian it must derive every verb from root, stem, and paradigm to the form on the tablet. A second, briefed as a specialist from a different grammatical tradition, checks each analysis independently and attacks the draft. A third, which wrote neither, rules on what they dispute.

Writes go through a script that enforces what is allowed. The agents may store parses and line translations and append to a word’s notes, and in Akkadian fill an empty lemma link. They cannot touch a transliteration, a sign reading, the lexicon, or my own notes; a reading they suspect comes back to me as a question. Every change lands as a reviewable diff of the versioned data.

The corpus as text

Everything lives in one SQLite file, and beside it in version control sits the database as text: the schema, numbered migrations, and one dump of inserts per table, regenerated by a commit hook. Each session of annotation is a commit whose changes read as rows added and altered, and a checkout can rebuild the database from the dumps alone. One set of schemas types both the server and the client.

Three languages, one design

The languages share an approach but not a schema. Elamite royal inscriptions are read in sentences, which run across lines, rather than in lines. An Akkadian box may hold more than one sign from the list, where Sumerian and Elamite boxes hold one. Akkadian lemmas carry a root, a theme vowel, their logographic spellings, and the lemma they derive from, so a word’s page can show its family.