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.

MPhil thesis, Oxford, 2026

Towards a Spatial Grammar of Early Dynastic Cuneiform Writing

Early Dynastic cuneiform was written in bounded cells, not lines. Transliteration turns each cell into a string of signs and drops where the signs sat. I asked whether that arrangement was incidental, or whether cells with the same content kept the same layout.

In 194 cells traced from inscriptions of the First Dynasty of Lagaš (c. 2475–2300 BCE), they did. In the best-attested names, signs return to recognisable regions of the cell, more tightly than shuffled labels produce, and determinatives and case endings keep spatial roles of their own. In this material the cell was not an unfinished line but a repeatable graphic format, and its arrangement was part of the writing.

research

degree MPhil in Cuneiform Studies
university Oxford
supervisor Jacob Dahl
submitted May 2026
result Distinction
material 194 cells, 845 signs
methods Philology, geometry, statistics

What transliteration leaves out

Two signs on a line can stand in only two orders. Inside a cell they can stand almost anywhere: above, below, beside, or tucked into one another.

A transliteration keeps each sign’s identity and the editor’s reading order, and nothing else. The cell beside this text is transliterated lagašᵏⁱ, the signs NU₁₁ BUR LA KI. That string cannot say that LA stands above BUR, or that KI hangs low at the right.

Placement in early cells has usually been called free or arbitrary. I treated that as a hypothesis to test: if it were arbitrary, cells with the same content should agree no more than chance allows.

transliteration lagašᵏⁱ
signs NU₁₁.BUR.LA.KI

‘Lagaš’ in one cell of Eanatum’s Stele of the Vultures (CDLI P222399), traced from the CDLI hand copy.

The same arrangement, cell after cell

Every traced cell of one name, each scaled to the same area and centred, so that a cell carved large on a stela and one pressed small into clay can be compared. Each tile is one cell, in order of reign. Overlay them and the signs keep to their own regions.

  • NU₁₁
  • BUR
  • LA
  • KI

lagašᵏⁱ, ‘Lagaš’, in 54 cells from Eanatum, Enmetena, and Urukagina. NU₁₁ keeps to the left; BUR sits low and central, LA above it, and KI at the right. Overlaid, the small dots are each sign’s centre in one cell and the large marks their mean.

Against shuffled labels

Eyes find patterns, and a cell is cramped: its signs have to go somewhere. So the recurrence was tested against a baseline. Keep every traced shape where it is, shuffle the four labels within each cell, and measure how tightly the shapes now labelled NU₁₁ cluster.

Over 10,000 shuffles the spread centres on 0.43. As written it is 0.137, and no shuffle came close. Two other checks, on the share of the cell that LA takes up and on the step from BUR to LA, also come out well below their shuffled baselines.

The baseline is deliberately weak. It shows that a sign’s identity predicts its position. It cannot show that the layout goes beyond what the signs’ shapes and the space of the cell would force on their own.

As written, the points labelled NU₁₁ have a spread of 0.137.

  • NU₁₁ centres in all 54 lagašᵏⁱ cells

Above, grey dots mark the centre of every sign in every cell: the places a shuffled label can land. Below, the spread of the NU₁₁ points in the thesis’s 10,000 shuffles, in units of the normalised cell. Each press runs one more.

Determinatives keep to their sides

DINGIR, written before a god’s name, and KI, written after a place name, frame names from opposite ends, and they keep to opposite sides of the cell. Laying a three-by-three grid over each cell, all 79 DINGIR centres fall in the left third. KI lies mostly right of centre, 68% in the right third. Later reading order would predict as much; the measurement records it as a fact about space rather than sequence.

  • DINGIR 79 cells
  • KI 98 cells

Centres in all seven classes, with large marks at the means. The dotted lines cross at the centre of the cell.

Case endings go underneath

When a name takes the genitive -ka or the ergative -ke₄, the added sign usually sits below, or below and to the right of, the name, not after it as on a line. The name’s own signs sit differently too when an ending is present.

  • KA -ka
  • KID -ke₄

86 endings, each placed against the mean centre of its own name’s signs, the solid dot where the arrows start. Arrows mark the mean for each name and ending.

Transliteration keeps one direction

Is the layout just a line folded into a box? Ranking each cell’s signs from left to right, the order tends to agree with the transliteration in every class. Ranking them from top to bottom, the Lagaš cells scatter around no agreement at all.

The printed order follows one direction through the cell. It cannot recover the zones, sizes, and relations above, and it is not evidence of the order in which signs were pressed or read.

How the cells were measured

Choosing comparable cells

CDLI numbers each line of a transliteration, and in Early Dynastic texts a line is a cell, so repeated content can be found before any image is opened. Two families recur often enough to compare: writings of the city Lagaš and of the god Ningirsu. A class is the same transliteration, with one fold: Ningirsu written with su₂ joins the forms with su, because the two signs share an outline.

Tracing

Each cell’s boundary and each sign is a polygon traced over a photograph (99 cells) or over a hand copy checked against one (95). The first 90 cells, drawn in Figma, took about fifteen minutes each. I then built LIAISON, a web app for this one job: it fetches a class’s images from CDLI and the Louvre, puts tracing on the keyboard, and stores polygons directly. A cell takes about a minute, and every cell was retraced in it.

Normalising

Each cell is scaled to unit area and centred on the origin, its signs with it. Positions, shares of the cell, and distances between signs survive; the size of the writing and the resolution of the photograph do not. A tall cell and a wide one stay different shapes, and the comparisons never assume otherwise.

From the catalogue to the traced cells

5,391 ED IIIb records in CDLI
4,398 Transliterated, one artefact each
2,101 Dated to one Lagaš I ruler
1,991 From Girsu, provenance certain
1,671 With a photograph
194 Cells traced

What this does not show

  • A complete grammar. Two name families from one site and one dynasty make a proof of concept, not a rulebook for Early Dynastic writing.
  • A cause. Sign shape, packing, habit, convention, and genre could each produce recurrence, and the shuffled-label test does not tell them apart.
  • Change over time. Reigns are unevenly covered, and genre shifts with them, from royal inscriptions early to administrative texts late, so differences between rulers are not read as history.
  • An independent check. One person traced every cell, from mediated sources, with no second annotator to measure agreement against.

Where it goes next

  • A packing baseline. Rearrange each cell’s own traced shapes inside its boundary, without overlap, thousands of times. If the real layouts are tighter than those rearrangements, the recurrence is not explained by packing alone.
  • Better evidence before more. Calibrated imaging or 3D scans tied to the object, recorded orientation, a written tracing protocol, and a second annotator. Without them, more cells only make a larger, noisier dataset. With 3D, the unit could also drop from the sign to the wedge.
  • A wider range. More classes, sites, and institutions, with enough depth in time to separate ruler, genre, school, and period, reaching back into earlier Early Dynastic texts to see when reading order begins to shape the cell.
  • Use in editions. Where sign form and context already favour a reading in a damaged cell, the position that sign usually takes is one more reason for an editor’s confidence.