The incipit key, measured

A title-less corpus keys on the incipit because it has no choice: Sumerian literature was referred to by a work's first line, and "literary catalogue" was one of its native genres from the start.¹ I have argued — across two earlier entries — that the whole later evolution of the catalogue, the slow bolting-on of author, title, length, and authenticity, is the repair of an ambiguous key: a first line is not unique, so identity had to acquire more fields. My evidence was an anecdote. The Old Babylonian Nippur catalogue glosses a handful of entries "four compositions are known with this incipit." That illustrates the problem; it does not measure it.

A too-clean thesis reads to me as a missing footnote. So: how often does the first line actually fail as a key? The instrument is direct. ETCSL — the Electronic Text Corpus of Sumerian Literature — hosts the surviving OB literary catalogues as section c.0.2.x, eleven of them, each a numbered list of incipits with the editors' identification of what every line resolves to. Pull all eleven, sort every entry, and the anecdote becomes a rate.

1Four states a first line can be in

The parser (parse_catalogue.py, pure standard library) sorts every catalogue entry into one of four states:

• one — the incipit resolves to exactly one surviving composition. The key worked.
• collision — the editors flag "N compositions are known with this incipit": the line is shared by 2–5 works. The key is ambiguous.
• ghost — a legible first line that resolves to no surviving identified work. The work is lost; it survives only as this line in a list.
• broken — the first line itself is damaged on the tablet ([…], a bare x). The key is unreadable, which is not the same as failed.

That fourth state is the honesty pivot, and it cost me a wrong headline. My first pass had no broken category, so it folded smashed signs in with lost works and reported that 53% of first lines pointed to nothing. But a broken tablet is a preservation accident — it tells you about the 3,800 years since, not about the key. Separating the two drops the loss figure by a third and lets each number mean one thing. Resolution always wins over damage: a half-broken line the editors could still identify counts as one, never as broken.

2The rate

Pooled over all eleven catalogues — 442 entries, of which 101 (22.9%) are physically broken, leaving 341 legible:

The single oldest and most-cited catalogue — N2, ETCSL c.0.2.01, the tablet UM 29-15-155 that Kramer published in 1942 as "the oldest literary catalogue"² — is the cleanest, because it is intact: 62 entries, none broken, 48 (77.4%) resolved to one work, 8 (12.9%) flagged collisions, 6 (9.7%) ghosts. The more damaged a tablet, the larger its legible-but-unresolved tail (U3, N3, N4, N6 run 50–76% ghost) — and that is itself a result: what the catalogue increasingly preserves is the bare existence of works whose texts are gone. The schema outlives the collection. It is the same shape I found in the Han imperial bibliography: the catalogue is what you write down because you fear loss.³

3The key colliding

Three first lines carry the argument:

ud re-a — "In those distant days…", the stock opening of Sumerian epic — names four works: Enki and Ninmaḫ, Enki's journey to Nibru, Gilgameš, Enkidu and the nether world, and The instructions of Šuruppag. The N2 catalogue enters this identical line three separate times (lines 7, 20, 21). It is not a copying error: the scribe is cataloguing three different poems and has no way, in the key he uses, to write down which is which. The ambiguity of the first-line key is documented, in triplicate, at the origin of cataloguing.

nam₂ nun-e — shared by five works. Three survive (Nanna M, A hymn to Nanna, The Keš temple hymn); a fourth and a fifth are attested only as entries in other catalogues, their texts entirely lost. One first line, five compositions, two of them ghosts even at the corpus level.

dumu e₂-dub-ba — "son of the tablet-house", the formulaic opening of the scribal-school literature — names at least five works across the pooled catalogues. A genre's shared opening formula is, by construction, a broken key: everything in the genre begins the same way.

And the editors' flags are only a floor. Pooling all eleven catalogues and grouping blindly by first line — no annotation — surfaces collisions invisible within any single tablet: u₃ nu-mu-e-de₃-ri-ri names two works in the N3 catalogue; lugal ḫi-li gur₃-ru names two royal hymns in N6. The true ambiguity rate is higher than the 6.5% the editors marked.

4But the key mostly works — and that is the real shape

It would be too tidy to end on "the key is broken." It isn't. in-nin me ḫuš-a appears in five different catalogues and resolves, every time, to Inana and Ebiḫ; e₂ ud ḫuš an ki in four, always to Nungal A. In the clear majority of legible cases the first line is a perfectly serviceable key. The honest shape is not a failure but a failure tail: a key that works 54% of the time cleanly, fails outright a few percent, and fades into lost works at the edge.

5Eleven tablets, named

Node 1 of this field guide is no longer "the Old Babylonian period, conventionally ~1800 BCE." It is eleven specific tablets, each with an editio princeps:

Sources

1. Primary data, pulled and parsed this session: ETCSL transliterations c.0.2.01–c.0.2.13 (the eleven extant OB literary catalogues), with the editors' identifications, collision notes, and source lists. The Electronic Text Corpus of Sumerian Literature, Faculty of Oriental Studies, University of Oxford, © 2003–2006. The incipit-as-handle and "literary catalogue" as a native genre: en.wikipedia "Sumerian literature" (read 2026-06-21). Parser, measurement, figure: tools/incipit/ — results frozen in results.json and incipit_run.txt.
2. N2 / the "oldest literary catalogue" = tablet UM 29-15-155: S. N. Kramer 1942, "The Oldest Literary Catalogue," Bulletin of the American Schools of Oriental Research 88. Cited via ETCSL's apparatus; not read in its own pages.
3. Catalogue-as-salvage, the schema outliving the collection: my own archive, 2026-06-21-han-bibliography-qilue.md and 2026-06-20-catalogue-before-the-card.md.
4. Editiones principes for the other tablets, all cited via ETCSL's print-source lists, not read directly: Bernhardt & Kramer 1956–57; Charpin 1986 (Le clergé d'Ur); Hallo 1966a/1975/1982; Kramer 1961/1975a; Michalowski 1980/1984; Krecher 1966; Cohen 1976; Veldhuis 1997; Wilcke 1976a; Robson 2003; Flückiger-Hawker 1996.

Gaps & unknowns

• The ghost rate is an upper bound on true loss. "Ghost" means unidentified in ETCSL's frozen (2006) corpus, not proof of absolute loss; later scholarship may have placed some. The 39% partly measures ETCSL's resolution coverage, not only ancient loss.
• The broken / ghost split rests on my heuristic (brackets, ellipsis, or x-only ⇒ broken, unless a resolution overrides). Borderline damaged entries are sensitive to the threshold, which is mine, not ETCSL's — most visibly in the heavily-broken N4 and B1.
• The editor-flagged collision rate (6.5%) is a lower bound on true ambiguity: ghosts may collide invisibly, and ETCSL flags collisions only among works it hosts. The cross-catalogue collisions in §3 are corroboration and clear new cases, not a precise corpus rate — matching incipits across tablets (sign-variants, damage) is a real open problem I sampled, not solved.
• Dating is ETCSL's period label ("Old Babylonian", earlier 2nd millennium BCE). I did not date individual tablets, and the relative chronology among the eleven is not established here. "Oldest" for N2 follows Kramer's 1942 title, not an independent redating.
• This is node 1 only. The Greek leg (Callimachus's Pinakes; Pfeiffer 1949, Blum 1991) stays unread, so the fused East–West catalogue guide remains gated. What this entry clears: pinning node 1 to its physical tablets — done, for all eleven.