²³The Proof Chain

Every claim on this site is meant to rest on four linked layers, and the links between them matter more than any one layer. The fact is the statement itself. The source is the archival item behind it, a gazette page or a birth register as the archive holds it. The evidence is our preserved copy of that item, the scan or photograph kept on disk, and the bytes are the actual file underneath. So the small mark that opens a citation here is not a footnote number: it stands for all four layers, and a reader can follow each one.

The chain only works if these layers stay apart. A crop of a page, a machine transcription of it, or a copy mirrored from another site still represents the same archival item; putting it in a separate file does not make it a new source. A reading we write ourselves is a companion to the source, never a source of its own. For the same reason, each preserved file is kept with a sha256, a checksum that changes if a single byte of the file changes, rather than with its dimensions or a thumbnail. The checksum cannot prove that the file is the right document, or that it is cleared to publish. It fixes the exact bytes a reading was made from, so a later substitution cannot slip through unnoticed.

Two separate checks are built into this. The checksum guards our own copy: a file swapped afterwards is caught the next time it is checked. And because the source is kept as a record of its own, apart from our scan, the archive reference travels with the claim, so our copy can be held up against the archive's version wherever that version can be reached.

The sheet above is one page of the 1899 building-permit dossier, the cadastral extract that maps the corner parcel. It is exactly this kind of preserved file: an archival source, scanned, kept on disk and fixed by its checksum.

A single confirmed claim, followed the whole way, shows what a citation is made of. In 1920 a Rotterdam physician in this story was made an Officer in the Order of Orange-Nassau by royal decree; the line is printed word for word in the state's own gazette, and we keep the full page on disk¹. That walk-through sits in one fact followed end to end, beside an older finding that is just as certain but carries its support more loosely.

The chain can also be measured, and we publish the measurement instead of assuming it. At the 10 September 2026 snapshot, 174 of the 385 facts carried a structured support block, with the locator, the source wording and the inference written into the record, and 211 did not. That gap is a legacy of older records, not a fault, and some of those facts cite their sources through a plain list instead, left visible with the limitation shown. Of the 709 sources, 641 name at least one preserved file that exists on disk and 68 do not. That is a simple filesystem test: it cannot tell a documented reason for missing bytes from a gap nobody recorded. The authored companions, the readings written beside the scans, number on the order of nine hundred, though the count shifts with the method used to take it. The proof-closure panel computes each measure at build time and prints its exact definition beside the number. These figures show coverage. They do not certify that any one claim is right.

Every claim carries one of six status words, and the site is strict about which. Confirmed means a primary source directly attests it in a supporting role. Probable means a well-supported inference that no source states outright. Disputed means the evidence pulls two ways. Open means a question the record has not yet answered. Lead means a thread worth pulling that is not yet a conclusion. False-positive marks an association that looked real, was tested and was rejected; it stays on the record so it cannot be adopted again by mistake.

All six are in use, and the labels travel with the sentences. The 1920 honour above is confirmed. Number 169 is probably the oldest building on its stretch of the Rochussenstraat, with later infill built up around it, but that is a probable reading of the house numbering and not a documented build order¹. Who owned the building between 1987 and 2009 is an open question. Two Rotterdam firms called Mico turn up in a business index and are only a lead toward an answer: they have been checked but not yet tied to the address. Then there is a 1930 newspaper notice that prints "Rochussenstraat 169" but turns out to mean an address in Den Haag. It stays on the record as a false-positive so that no later search adopts it again, which is exactly why a rejected association is never written up as a plain claim.

One rule governs all of this. No model, no validator, no count of agreeing sources and no completed form can promote a claim on its own. Status is a reviewed human judgement, and only a human review changes it.

A project earns trust by correcting its mistakes in the open, not by never making them. Here every correction gets its own record. The new value sits beside the old one, with a required reason and references to the sources and any preserved scans behind the new reading. The earlier wording stays on the record as a superseded reading, and the history of the change is kept. Keeping the dead value in view means no later draft can repeat it by accident. A correction can also run the other way: a confirmed fact is never frozen, and downgrading it reopens the question instead of sealing it.

The project's corrections are gathered in a ledger. Some were machine misreadings. The approval of the 1899 permit was first recorded as 23 June, read wrongly off the stamp, and later corrected to 23 May¹. A painter's family street, read from a difficult hand as Eendrachtsstraat, proved on a careful re-reading to be Eenhoornstraat². One error credited the wrong man, until four separate documents agreed on the architect Nicolaas Tinbergen³. Two figures had simply been invented: a table of ceiling heights the sheets never carried and a floor area more precise than they allow; each was struck out or replaced by an honest estimate⁴. The dead-claims ledger lists more of these corrections, with what replaced each value and why.

An empty search can be a finding in its own right, though it is the part of the method a general reader is slowest to trust. It earns that standing only when the search is defined, run and recorded: this index, these years, these spellings, this repository. The absence then has a scope, and a later researcher knows what not to repeat.

Several such searches hold up parts of this site. After the last furniture firm held its closing-down sale in 1987, no successor, takeover, merger or continuation of it appears anywhere in the digitised press that was searched¹. No nuisance-permit notice naming number 169 appears in that press either, though the same street's notices are plentiful². And a search of the occupation-era Möbelaktion for any tie to the firms at number 169 found none; that absence is recorded at its exact scope and never stretched into an acquittal³. Because each of these states what was searched, the list of closed routes can keep every closed route together with the outside developments that would reopen it.

Good intentions alone would not hold any of this together. Software enforces the chain, the status words and the corrections, refusing any change whose evidence does not resolve, and that software is the subject of the next chapter.