Here is a question no institution on earth can answer: has this book been scanned?
Not “does my library have a scan”; any catalogue can answer that. The real question a scanning program needs answered: has anyone, at any of the thousands of institutions digitizing books, already put this particular printing online? Because scanning budgets are finite, and the marginal scan should be a book the world lacks, not a fourth copy of one it already has.
There is no registry to ask. This post is about what we had to build instead, and why the world should build the real thing together.
What it took to answer the question once
This week we tried to answer it for one library: the Bibliotheca Philosophica Hermetica in Amsterdam, whose 29,879-record catalogue we help steward. The library wants to digitize; it wanted to know which of its books deserve the scanner first.
To find out, we had to compare the catalogue against every digitization we could see. “Every digitization we could see” is doing enormous work in that sentence. There is no list. We have spent months harvesting one: 3.5 million records of scanned books, gathered institution by institution (the Internet Archive, the Bavarian State Library, e-rara, Gallica, Biblissima, the Vatican, and a dozen more), because the IIIF standard that makes scans interoperable deliberately has no central registry of what exists. In the IIIF community this is politely called the “discovery problem.” It has been open for a decade.
With that private census in hand, the answer took an afternoon: 3,655 books printed before 1830 in this one collection have no digitization anywhere we can find, plus about five hundred manuscripts, unique by definition. Meanwhile 324 of its books are printings that are already online in full, where a rescan would duplicate the world's coverage almost exactly.
That is the scanning queue any collection would want. And here is the absurdity: every library that wants to scan responsibly has to rebuild this same apparatus privately (the harvest, the identity matching, the census) or scan blind. Most, reasonably, scan blind.
The registries that almost exist
The pieces are scattered across a landscape of heroic, partial efforts. The Universal Short Title Catalogue knows 1.65 million European editions printed before 1700 and links about 600,000 digital copies: the best attempt anywhere, and still tagged to under a third of its records, with nothing after 1700. The incunabula catalogues (ISTC, GW) do this well for the fifteenth century. The German VD16/17/18 do it for German print; the Dutch STCN for Dutch. The English ESTC links to scans that mostly sit behind subscription walls. Europeana and HathiTrust aggregate tens of millions of digitized objects but at the level of files, not editions; they cannot tell you whether two records are the same printing. Google scanned perhaps forty million books and publishes no list at all.
Each covers one period, one nation, one language, or one silo. None joins to the others. A librarian in Amsterdam holding a 1688 Latin alchemical treatise would need to consult half a dozen of these and would still miss the digitization sitting in Munich under a variant spelling.
That last detail is not hypothetical. In our spot checks, a 1688 Tollius volume showed up as “never digitized” until we found its later German edition at the Bavarian State Library, catalogued under Manvdvctio Ad Coelvm Chemicvm, with the author's name fused into the title and coelum spelled against our caelum. Early modern books do not agree on their own names. Matching them takes bibliographic identity work (author, title, year, volume, orthography), not string comparison.
What a real registry would take
We have learned some of this the hard way, building the private version. A global registry of digitizations would need four properties:
- Edition-grain identity. The unit is the printing, not the title string and not the file. “Some copy of some edition of this work exists somewhere” and “this 1688 Amsterdam printing is online” are different facts, and a scanning decision needs the second.
- Evidence, not flags. A record should carry the manifest URL itself: a checkable claim. We inherited a dataset whose bare “has been scanned” flag turned out to undercount by more than a quarter, and there was no way to tell which rows were wrong because the flag carried no provenance.
- Orthography-aware matching. u/v, i/j, ae/oe, names fused into titles, titles translated between languages. Exact matching quietly fails on exactly the books that matter most.
- Openness. Harvestable in bulk, IIIF-native, contributed to by the institutions doing the scanning, the way OCLC built a shared catalogue of holdings, but for digitizations, and open.
None of this is speculative. The USTC has proven the bibliographic spine works at the scale of a million editions. IIIF has proven the interoperability. The missing piece is the join, maintained as shared infrastructure rather than re-derived in private by every project with a harvester and a deadline.
An offer
Source Library is a small project, and this is not an announcement that we are building the world's registry. But we hold a working prototype of its parts: a 3.5-million-record harvest of digitized books across nineteen sources, matching tooling tuned for early modern print, and a live case study in what the census makes possible: a real library's scanning priorities, computed rather than guessed.
If you work on USTC, on IIIF discovery, at CERL, at an aggregator, or at a library trying to decide what to scan next: the data and the tooling are open, and we would rather contribute them to a shared registry than keep maintaining a private one. Write to us.
Every year, digitization money is spent rescanning books that are already online, while books that exist nowhere else wait. Not because anyone chose that, but because nobody can see the whole board. The fix is a list. Libraries have known how to make lists for four thousand years. This one is overdue.
