¹⁹Sources and Evidence

Every sentence on this site that carries weight is built so you can follow it back to something you can look at. The path has four hops: a fact; the support entry that says how the fact was read; the source behind that entry; and the evidence record that holds the preserved file.

Take the masterpiece cabinet the furniture works built in 1895. The fact says that three named Rotterdam makers did its design, its sculptural modelling and its imitation-gobelin painting, and it is marked confirmed¹. The support entry spells out the reading. It quotes the source, the Stadsnieuws notice of 18 September 1895 that names the makers, and records what the researcher infers from it: that all three are documented figures in the decorative arts. It gives the locator, Rotterdamsch Nieuwsblad, 18-09-1895. And it states the uncertainty: one maker's role rests on that notice plus prosopography rather than a second independent source. Behind the entry stands the newspaper record itself, held at Delpher and catalogued with everything else in the source landscape. The evidence record gives its kind, a full-page scan, its locator and a sha256 checksum, here beginning f2f82e00, which proves the file has not changed since it was captured. The last hop is the file.

That chain is the rule, and we measure it rather than simply claim it. At the 10 September 2026 snapshot, 174 of the 385 facts carry a structured support block like this one, and 211 do not. That does not make any of the 211 false. The figure marks a gap left by older support: in some, the support still sits in a plain source list, and each needs checking claim by claim to reach the same standard. On the source side, 641 of the 709 sources name at least one evidence record whose file is present on disk, and 68 do not. That test checks only that a file is there; it proves neither quality nor the right to publish. How the two figures are defined, and what they do and do not certify, belongs with the methodology, in the measure of proof. One rule covers all of it: a generated reverse view, a graph or a summary helps you inspect the chain, but never proves anything on its own.

The size of the corpus is a dated snapshot and keeps changing. On 10 September 2026 the registry held 709 sources and 2,817 evidence records. The two count different things. A source is the record or publication something comes from; an evidence record can be a preserved copy of it or a reading made from it.

Two holdings carry most of the weight: Delpher, the Koninklijke Bibliotheek's digitised newspaper archive, and the Stadsarchief Rotterdam, through its reading room and image bank. The rest are record-label catalogues, the city's family-history centre, the national heritage photograph collection, an encyclopaedia's media library and the venue's own web pages. Newspaper pages and cuttings make up the largest group by type. Web pages, photographs, civil and population-register sources and advertisements follow close behind, and then come address books, maps, building permits, deeds, official gazettes, genealogical indexes, paintings, postcards and posters. One trade-paper feature on the corner shows the kind of material the whole record is read from.

At this snapshot, 315 of the 709 sources are cited by no fact's source list or embedded support entry. That makes them a set to review, not a verdict. Some back an entity record or a row in a register rather than a dated fact. Some document a search that was run and returned nothing, a finding the project keeps. Others are simply waiting to be linked more closely. The count points to work to be done; it does not prove that every one of them is needed, or that every one is properly connected. The source landscape sets out the whole picture by type and repository.

On its own, a scan is only a picture of a page. It becomes usable through the reading set down beside it: a word-for-word transcription of a document, a translation where the page is not in English, a description of a photograph, map or drawing, and a note of how the reading was made. When a claim rests on a hard handwritten value, that note must record who read it and confirm it against a separate, independent reading. This companion opens in the viewer when you enlarge a scan and turn its tabs.

Counting coverage honestly takes care, because the parts differ: a photograph needs a description and no transcription, and an English page needs no translation. At the 10 September 2026 snapshot, something like nine hundred of the evidence items carried a reading companion whose file is present on disk, and like every figure on this page, that number moves as records are added. If revised readings are counted too, the raw count of companion records is higher. That is a count of readings made, not of unique scans, and we deliberately do not turn it into a single coverage percentage. An earlier percentage was withdrawn because its denominator could not be reproduced: it never said which kinds of evidence need which components. A number worth defending needs a census component by component: which evidence items are in scope, which need transcription, translation or description, and whether those parts exist. The project treats that census as a backlog to work through. The annotation-coverage note explains the method and shows one transcribed household card in full.

Run on people and firms instead of documents, the same chain leads from an entity to the records that name it. That index is on the next page.