Questions & answers
Everything people ask.
Where the tablets come from, how the engine reads them, what you can trust, and how one person keeps this growing.
The project
What is Sumerian Book?
A curated, readable history of Mesopotamia told through its primary sources: over 102,000 catalogued cuneiform tablets, presented both as a browsable catalogue and as a narrative book with a prologue, period chapters, and an epilogue. Every claim on the site is anchored to a tablet you can inspect — transliteration, translation, photograph, and provenance included.
Start reading the bookWho is behind it?
One person — a French citizen with no academic affiliation, no institutional funding, and no team. Not an Assyriologist, not a museum: an independent builder assembling what already exists in the open scholarly record and presenting it rigorously and beautifully.
About the projectHow was the site actually built?
With modern web tooling (Next.js) and heavy AI assistance: the site, its data pipelines, and its translation engine were developed with Claude, Anthropic's AI model, working alongside its creator. The data layer aggregates open scholarly datasets — CDLI, ORACC (SAAo, RINAP, RIAo, ETCSRI), ETCSL, Wikimedia Commons — into a single corpus that a nightly automated pipeline keeps improving.
Is this affiliated with CDLI, ORACC, or any museum?
No. The site reuses their open data with attribution and deep gratitude, but it is an independent project with no affiliation to any university, museum, or digital corpus project. Errors on this site are its own, never theirs.
Provenance, sources & ethicsWhy is it called Sumerian Book when many tablets are Akkadian?
The corpus spans the whole cuneiform world — Sumerian and Akkadian, from archaic Uruk to the Achaemenid period. The name honours where writing began: Sumer, whose script and scribal tradition everything later grew from. And 'book' because the ambition is a narrative you can actually read, not a database.
Does it cost anything? Is there advertising?
No ads, no paywall, no tracking-based monetisation. The project runs on its creator's own time and a modest infrastructure budget. If it ever needs support, the ask will be transparent — see the contact page for how to help.
Contact & supportThe tablets and the data
Where do the tablets come from?
From the open scholarly record: the CDLI catalogue (the backbone — identifiers, transliterations, photographs), ORACC projects including the State Archives of Assyria online, the royal inscription corpora RINAP, RIAo and ETCSRI, Oxford's ETCSL for Sumerian literature, and Wikimedia Commons plus open museum collections for photographs. About 96,000 tablets currently have a photograph on the site.
Full source list & licencesHow many tablets are translated?
Roughly 8,000 of 102,000 tablets currently carry a real translation. Most are established scholarly translations; a growing share are produced by the site's AI engine, always labelled as such with a confidence tier. The remaining tablets are catalogued with their transliteration or photograph while they wait their turn — the nightly pipeline translates a new batch every day.
Why do so many tablets share the same date, like ~2050 BCE?
Because most tablets can only be dated to a period, not a year. The site assigns each period's midpoint as an approximate date, which is why dates cluster. When a tablet's own text contains dating evidence — a year-name formula, a king's regnal year — the engine refines the date and records its reasoning, but that has covered only a small fraction of the corpus so far.
What does 'unclassified' mean in a tablet's address?
That the tablet has not yet received a trustworthy thematic classification. Bulk-ingested tablets arrive without one; the engine classifies a batch every night, and a tablet moves from /unclassified/ to its real theme (economy, law, mythology…) once it has actually been read. Old links keep working — the site redirects them to the tablet's current address.
Can I use the photographs and translations elsewhere?
It depends on the item — every tablet page states its sources and licences. Scholarly translations from ORACC projects are CC BY-SA; ETCSL and CDLI materials carry their own terms; most photographs remain © their holding museums and are reproduced here for non-commercial educational use. Check the tablet's attribution block and the terms page before reusing anything.
Terms of useThe AI engine
How does the AI translation work?
The engine is Claude wrapped in a retrieval pipeline. For each tablet it receives the transliteration, the photograph, glossary entries for the words involved, and parallel passages already translated by scholars. It examines the photo, cross-checks the transliteration against what it can actually see, and returns a literal translation, a plain-English interpretation, a modern rendering, and a confidence assessment with its uncertain terms flagged.
The engine, explained in fullDoes the AI learn from the tablets over time?
The model itself doesn't — its weights never change. What grows is the knowledge base around it: a glossary of Sumerian and Akkadian terms harvested from every translation, a bank of parallel passages, refined dates and themes. Each night's translations retrieve from a richer base than the night before. The system gets smarter; the model doesn't.
How do you stop it from making things up?
Several layers. The prompt enforces strict honesty rules: damaged text becomes brackets, uncertainty is flagged, never invented. Every night the engine takes a confabulation test — trap tablets whose photo and transliteration deliberately mismatch — and a single failure blocks that night's publication. And a fixed panel of 35 scholar-translated tablets is continuously re-translated blind and graded against the published references by a second, stronger model.
The nightly examWhat do the confidence badges mean?
Five tiers. 'Scholar-verified' is reserved for translations vetted by trained Assyriologists or drawn from established editions — the engine can never award it to itself. 'High', 'medium' and 'low' reflect how well-attested the text and reading are. 'Experimental' marks AI-assisted hypotheses: useful starting points, not substitutes for scholarly work.
What does 'read from photo' mean on a tablet?
That the engine visually examined the tablet's photograph while translating and cross-checked the transliteration against the actual clay. Tablets translated from transliteration alone — because no usable photo exists — are explicitly labelled as indicative readings and are held to a humbler standard.
Which AI models are used?
A frontier Claude model does the translating; a stronger Claude model serves as the independent judge for the nightly quality benchmark and as the editor that nominates the best new translations for the book. Model versions are recorded on every translation as part of its audit trail, so you can always see which model produced what you are reading.
Reading and contributing
What is the difference between the book and the catalogue?
The book (/read) is the curated narrative: a prologue, twelve period chapters and an epilogue built around a small set of editorially chosen tablets, written to be read in order like a history book. The catalogue (/tablets) is the full corpus — every catalogued tablet, sortable and filterable. The book is the story; the catalogue is the evidence room.
Open the bookI found a mistake. How do I report it?
Please do — corrections are the most valuable contribution there is. Use the contact form and mention the tablet's identifier (visible on its page). Confirmed corrections enter the knowledge base and improve future translations of related tablets.
Report an errorHow can I help?
Three ways matter most: scholarly expertise (reviewing engine translations, even a handful), tablet photographs (museums and collections willing to share images of uncatalogued or unphotographed tablets), and support for the translation effort. The contact page has a dedicated topic for each.
Get in touchCan AI crawlers and other tools index the site?
Yes, deliberately. The site publishes a llms.txt manifest, structured metadata on every tablet, and chunked sitemaps covering the full corpus, so that search engines and AI assistants can find and cite the primary sources. If you are building on top of the data, start from the original scholarly corpora — and check each item's licence.
The llms.txt manifest