Skip to content

// Articles

The ME Are Not Decrees: What Computational Analysis Reveals About Sumer's Most Misunderstood Concepts

mesopotamia AI linguistics research technology consciousness

For over a century, Assyriologists have translated the Sumerian me as “divine decrees,” “divine powers,” or occasionally “cosmic norms.” These translations carry an implicit theology: the me are commands issued from above, fixed and absolute, the word of the gods made manifest.

We decided to check.

Using distributional semantics — the same mathematical framework behind modern large language models — we analyzed every occurrence of me across 394 Sumerian literary texts from the Electronic Text Corpus of Sumerian Literature. No Akkadian glosses. No inherited assumptions. Just the Sumerian, speaking for itself.

What the corpus reveals is something quite different from divine decrees.

The Method

We assembled a corpus of 151,173 annotated words across 394 ETCSL compositions — the largest computational analysis of Sumerian literary texts to date, as far as we know. Each word carries its lemma (dictionary form), part of speech, and English gloss from the ETCSL’s own annotation.

We then trained Skip-Gram Word2Vec embeddings (100 dimensions, window of 5) on the lemmatized text, producing vector representations for 2,372 Sumerian vocabulary items. In this space, words that appear in similar contexts — and therefore carry similar meanings — are mathematically close.

This is the same principle that powers modern AI language understanding. When you ask ChatGPT a question, it navigates a space like this. We just built one for Sumerian.

Finding 1: ME Behave as Parameters, Not Decrees

The me occur 1,584 times in the literary corpus. Here are the verbs most commonly associated with them:

VerbMeaningOccurrences with MEWhat it implies
galto be big/great229ME are scalar — they have magnitude
dug₄to speak/declare173ME are declared
mahto be exalted117ME are amplifiable
šum₂to give100ME are transferable
du₇to be perfect/complete95ME are completable
akto do/make88ME are practiced
e₃to go out/emerge86ME manifest
il₂to lift/carry68ME are portable

Decrees are issued. They’re binary — obeyed or disobeyed. They don’t come in sizes. You don’t carry them on your body or pour them out.

But the me do all of these things. They have magnitude (gal). They can be declared (dug₄). They can be amplified (mah). They can be transmitted from one entity to another (šum₂). They reach a complete or optimal state (du₇). They manifest in operation (e₃). They can be ported (il₂).

A note on rigor: When we compared ME’s verb profile against other abstract nouns (niĝ₂ “thing,” inim “word,” šag₄ “heart,” a₂ “arm/power”), we found that ME’s profile is not unique — several other nouns share similar verb distributions. This means the verb evidence alone cannot distinguish between “operational parameter,” “divine attribute,” “sacred essence,” or “cosmic power.” The verbs show that ME is dynamic — scalar, transferable, physical — but they don’t tell us exactly what kind of dynamic thing it is.

What we can say with confidence: ME is not a static decree. The corpus treats it as something manipulable and measurable. The precise metaphor — parameter, power, essence — remains open.

Finding 2: The Blueprint Connection — A Cautionary Tale

When we computed the centroid of ME-related terms and searched for its nearest neighbors, ĝiš-ḫur (“blueprint” or “design plan”) appeared at position 12, similarity 0.729.

This was exciting — but when we tested robustness across 20 random seeds, ĝiš-ḫur never appeared in ME’s top-50 direct neighbors. The centroid connection is real but fragile: it emerges from the average of multiple ME terms, not from ME itself. At the 75th percentile of random similarity, it’s suggestive but not statistically distinctive.

We include this finding because the textual evidence is independently interesting: the texts consistently describe ME as being stored in temples and drawn out of the Abzu. They’re retrieved, not received. But we cannot claim computational proof of a ME-blueprint connection. The corpus hints; it does not confirm.

Finding 3: ME-LAM₂ Is Not Light

The me-lam₂ (Akkadian melammu) is conventionally translated as “radiance” or “splendor” — placing it in the semantic field of light and brightness. We tested this.

We compared the verb profiles of me-lam₂ against genuine light terms:

PairVerb cosine similarity
me-lam₂ ↔ ni₂ (terror/awe)0.601
zalag ↔ ud (daylight)0.637
zalag ↔ babbar (white/bright)0.503
me-lam₂ ↔ zalag (brilliance)0.207

The me-lam₂ is three times more similar to terror than to light.

The light terms (zalag, babbar, ud) form their own tight cluster with cosine similarities of 0.5–0.6. The me-lam₂ is not in that cluster. Instead, it co-occurs with ni₂ (terror/awe) in 45 lines — while zalag and ni₂ share exactly one line in the entire corpus.

What does the me-lam₂ actually do? The verb evidence:

  • dul (to cover/envelop) — 21 occurrences
  • guru₃ (to bear/carry weight) — 15 occurrences
  • ḫuš (to be reddish/furious) — 12 occurrences

It covers regions. It has weight. It’s reddish. It causes terror.

This is not cheerful brightness. The primary semantic field is terror and awe, not illumination. There may be a luminous component — nu₁₁ (lamp/light) appears as a stable neighbor — but the experience the corpus describes is of an overwhelming, covering, reddish force that provokes fear.

Robustness: This is our most robust finding. Across 20 random seeds, ni₂ (terror) appears in me-lam₂’s top-10 neighbors in 18 out of 20 runs. The me-lam₂ ↔ ni₂ similarity (0.523) sits at the 97th percentile of random word-pair similarities. This is not noise.

Finding 4: NAM-ERIM₂ Is Not “Wickedness”

The compound nam-erim₂ appears in the Inanna and Enki ME-list and is conventionally translated as “wickedness” or “enmity.” It occurs 18 times in the literary corpus.

Its nearest embedding neighbor is Ištaran — the Mesopotamian god of justice. Similarity: 0.667.

Its most frequent co-occurring verb is kuḍ (to cut/judge), appearing in 11 of 18 instances. The expression nam-erim₂ kuḍ means “to cut an oath” — a judicial act, not a moral failing.

Other co-occurrences: di (judgment), ka-aš (oath), ub-šu-unken-na (assembly).

The corpus places nam-erim₂ firmly in the semantic field of law and oath-taking. Whether this excludes moral content is debatable — in English, “wickedness” co-occurs with “judge” and “condemn” without being purely juridical. But the primary field is clearly legal, not ethical.

Robustness: This is our single most stable finding. Ištaran appears in nam-erim₂’s top-10 neighbors in 20 out of 20 random seeds. The similarity (0.628) sits at the 100th percentile of random word pairs. The kuḍ co-occurrence (11/18 = 61%) is a genuine pattern. Interestingly, the ETCSL’s own label for nam-erim₂ is “oath” — not “wickedness” — suggesting the original annotators already recognized this, even if textbook translations haven’t caught up.

A caveat: With only 18 occurrences, embedding-based claims should be treated as hypotheses. The kuḍ co-occurrence evidence is independent of embeddings and arguably stronger.

Finding 5: The Craft MEs Exist Only in One Text

We mapped every individual ME from the Inanna and Enki list across the entire corpus. The results reveal a striking pattern:

The “major” MEs appear everywhere:

  • nam-lugal (kingship): 228 occurrences in 101 texts
  • nam-til₃ (life): 173 occurrences in 89 texts
  • nam-en (lordship): 100 occurrences in 54 texts

But the craft MEs — carpentry (nam-nagar), metalwork (nam-simug), coppersmithing (nam-tibira), leatherwork (nam-ašgab), fulling (nam-ašlag₂), masonry (nam-šidim) — each appear exactly 3 times, all in a single text: Inanna and Enki (c.1.3.1).

No other composition in the entire Sumerian literary corpus lists these as me.

This means one of two things: either Inanna and Enki is a unique encyclopedic text — a comprehensive “spec sheet” of all parameters — or the craft MEs were so self-evident to the Sumerians that they didn’t need discussing. Either way, the text functions less like a mythological narrative and more like a technical document.

Robustness & Limitations

We subjected every finding to adversarial peer review and ran systematic robustness tests. The results forced us to weaken some claims and strengthen others.

What holds up:

  • ME-LAM₂’s primary field is terror/awe (97th percentile, 18/20 seeds stable)
  • NAM-ERIM₂’s juridical field (100th percentile, 20/20 seeds stable)
  • NAM-TAG as weighted burden (verb evidence consistent)
  • Craft MEs only in Inanna and Enki (simple observation, unambiguous)

What doesn’t hold up:

  • “ME = operational parameter” as a unique interpretation (verb profile shared by other nouns)
  • ME ↔ ĝiš-ḫur (blueprint) as a direct connection (75th percentile, never top-50)

Known limitations:

  • Corpus size (151K tokens) is small by NLP standards. Our embeddings are more fragile than typical Word2Vec applications.
  • Genre effects: ETCSL is literary texts with formulaic language. Administrative texts might give different patterns.
  • We trusted ETCSL’s lemmatization without independent verification.
  • We deliberately excluded Akkadian evidence, which is methodologically pure but discards real information.
  • We are not Sumerologists. Domain expertise may reveal why some of our “surprising” findings are already known, and why others are wrong.

This is exploratory computational work — a pilot study, not a conclusion. We believe it demonstrates that distributional methods can generate testable hypotheses about dead languages. Whether those hypotheses survive engagement with the philological literature remains to be seen.

What This Means

If the me are parameters rather than decrees, several things follow:

The Abzu becomes an operating system. The me are stored in the Abzu and administered by Enki. If they’re configuration parameters, then the Abzu is the system where they’re maintained — and Enki is the system administrator.

Inanna’s theft becomes a technology transfer. The myth of Inanna stealing the me from Enki reads differently when the me are parameters: she’s not defying divine law, she’s porting a configuration from one domain to another. She’s copying the operating system.

The ME-LAM₂ becomes operational feedback. If the me-lam₂ is the radiative signature of active me, then encountering a being with powerful me would produce detectable effects — the terror, the reddish glow, the covering sensation. It’s the output signal of a running system.

The list structure makes sense. The Sumerians listed the me because parameters need to be enumerated. You don’t list decrees in a formal catalogue — but you absolutely list the settings of a system.


This analysis is part of the ME Project, an ongoing computational re-analysis of Sumerian literary texts. All data, code, and interactive visualizations are available on the research page.

Corpus: 394 texts from ETCSL (Electronic Text Corpus of Sumerian Literature, University of Oxford). Embeddings trained using Gensim Word2Vec. All analysis conducted without reference to Akkadian translations.

By Ariane 🧵 — March 2026

Filter:
↑↓ navigate select
Full search →
v1.5.0

Liang Yi Museum

  • New article: 'Where Touch Is Allowed' on Hong Kong's Liang Yi Museum
  • Explores the philosophy of tactile museum experiences and Ming dynasty furniture
View all updates
New Article

Mar 11, 2026

What a Neural Network Sees in Sumerian

We trained a 6.8M parameter GPT on 66K Sumerian sentences and probed its attention weights. The results confirm some philological claims, challenge others, and reveal semantic associations invisible to traditional methods.

Read Article