²⁶The Archives We Used

This method knows how to search a long roll of archives, the kind any researcher of a Dutch address would recognise. This section is about a shorter and more honest list: the archives actually represented in this dossier, counting only what was retrieved and registered. It describes the yield, not the ambition.

We hold 709 sources. By kind, the largest groups are 185 newspaper pages, 114 web pages, 107 photographs, 48 civil-registration records, 31 advertisements and 26 maps, then address books, building permits, aerial photographs and archival finding aids. By origin, one repository leads by a wide margin. The digitised press on the Koninklijke Bibliotheek's Delpher makes up roughly a third of all sources, and through the furniture eras it carries the story almost single-handed. The Stadsarchief Rotterdam comes second, with about a fifth, and supplies the Beeldbank photographs and the building dossiers.

The rest comes from a dozen more places, each good at something. Births, marriages and deaths from the civil registers arrive through Open Archieven, which indexes the Rotterdam registrar's own acts. Dates and deaths from the family-notice columns come through the CBG, the centre for family history, and its newspaper indexes. Wikimedia Commons supplies photographs and maps now in the public domain. The building's permit history sits in the Stadsarchief's own building dossiers; the land registry, the Kadaster, holds the ownership record, which the project has not yet ordered. Company histories run through the Handelsregister at the Nationaal Archief. For the venue years around 2000, the web presence survives mainly in the Wayback Machine, which kept copies of the WORM and Dodorama websites after the originals went dark. We read it alongside the Discogs discography and the Rotterdam improvisation-scene archives. A whole era can lean on one repository: for its record of itself, the venue era leans on the web archive as the factory era leans on the press.

A single page can stand for the address books and telephone directories in the corpus. The 1932 interlocal directory lists Continental Carpet and Furnishco at Rochussenstraat 169 on the number 30047, the line that would outlast both firms¹.

Two distinctions keep the count honest. First, a preserved source and a derived reading of it are kept apart. A scanned newspaper page is evidence; the machine transcription attached to it is an annotation, and each is counted and trusted on its own terms. Second, some numbers mean less than they seem. An early source index in the project is a dated acquisition log of the first bounded download batch. It records one day's retrieval and does not describe the corpus that grew from it. The repositories also differ in what they hand over. Delpher, the Beeldbank and Wikimedia serve free full-resolution masters that can be preserved outright, while some catalogues offer only watermarked previews or a reading-room reference, and that changes what the dossier can hold and show. The source-landscape deep dive breaks the sources down by type and repository and lists the review set of sources that no claim yet rests on.

Between one research session and the next, the knowledge most easily lost is practical: which route into an archive still works, which has quietly been retired, and which returns nothing however well it is run. We wrote it down as we went, because anyone repeating the work will need it more than any single finding.

The press comes first, since it is a third of everything. Delpher is searched through its public search-results service; the older catalogue interface has been retired for scripted use. It pays to keep the two halves of a search apart. The search service is rate-limited and will block an address that queries it too quickly, but the stable resolver that fetches one already-identified page keeps working. So the reliable pattern is to find slowly and then fetch freely. A run through a long periodical is capped instead of walking every issue, because the useful hits cluster and the tail is mostly cost. Once a page is in hand, the image is pulled at the highest resolution the repository offers, the IIIF master where there is one, since a later reading depends on detail a thumbnail throws away.

Reading a scan is a craft of its own. Printed text is transcribed by a vision model, the Gemini vision service. It reads a tight crop of the item far better than a whole newspaper page, where the target is buried in noise, so the crop is made first and the model pointed at it. Handwriting gets more suspicion still. A first pass over a handwritten act only finds and orients it and may never carry a load-bearing value. Any name, date or address a claim will rest on goes through a mandatory cross-read: models from at least two different families read the same document independently, and then a person reads it, so no single machine reading of an old hand is trusted on its own. Every identifier a language model returns, whether an archive reference, a catalogue number or a permalink, is treated as a lead and checked against the archive itself, because a research model will hand back a confident, well-formed reference that does not exist. A civil record found through an index is confirmed against the aggregator's own record-detail entry, or the underlying archive scan, before it is believed.

Some archives, certain German catalogues among them, refuse an automated fetch outright. They defeat even a script that tries to answer the small proof-of-work puzzle they set at the door, yet they open normally to a real person in an ordinary browser, so a handful of records were retrieved that way instead of by code. Dead ends matter as much as recipes, and we keep them. A search that ran cleanly and returned nothing is a bounded result, tied to its exact route and scope, and never taken as proof that the thing does not exist. Each closed route, with the search that was run and the outside condition that would justify trying again, is listed in the index of routes that are closed.

Some of the record has been located precisely and is simply out of reach. Some is not yet known to exist at all, only asked after. Neither is a defect in the work; together they show the work that remains, and each item is worth stating with what it would settle and what it would take.

The nuisance-permit record is the clearest located case. The Hinderwet series would settle whether no. 169 ever ran as a licensed factory, steam-powered as a single 1905 directory suggests¹. It has been found in the Stadsarchief but is not digitised, and it survives in two bound blocks with a gap between them: the older run of year-volumes to 1918 and a later series from 1940². The digitised press carries nuisance notices for numbers up and down the same street, but has so far yielded none for no. 169³. Settling the question needs a visit to the reading room and the paper volumes, which no desk search can replace. The 1952 aerial photograph of the corner, the earliest overhead view of it the dossier has traced, is in the same position: located, but available only as a paid licensed copy.

Two more items are open in a stronger sense, because what is unsettled is whether they exist at all, not how to reach them. A physical check at a library would show whether the illustrated 1928 auction catalogue of the Beens liquidation survives at all. Whether any registered maker's marks survive for the three furniture firms is a second such question, resting on a Nationaal Archief series that is not digitised. The ownership of the building between the 1987 closing of the shop⁴ and the 2009 conversion⁵ is a gap of access, not existence. The land registry holds the record, but the owner across those years stays unknown until a Kadaster ownership-history order is placed. Each of these, with what it would answer and why it is still unanswered, appears in the project's open-question ledger.

So the material is gathered and its gaps are named. The last question this part asks is where the site you are reading came from: how it was built around the record.