²⁷How This Site Was Made

The design was fixed before a word of the site was written, so every page behaves the same way. The book's four narrative parts sit in the header, and each chapter is a page. As you scroll, the section headings feed a rail down the left, so you can find your way round a long chapter without losing your place. Deep dives are pinned in a right-hand column beside the passage they belong to, and do not interrupt the text. Where a passage has no deep dive pinned to it, that column shows the evidence behind it, in order. Every plate opens in a viewer with the reading its page calls for: a description of what it shows and, where the plate bears text, a word-for-word transcription of the lettering, translated into English wherever the page is not already in English.

The drawing shown here is a crop of the 1900 municipal-works (Gemeentewerken) plan of Rotterdam, covering Middelland and the Land van Hoboken. It is drawn at district scale, so the house itself is not marked. What the reading includes depends on the plate. A photograph or a drawing is described. A printed page, such as a newspaper advertisement, is transcribed and, since it is in Dutch, translated. A handwritten civil act is transcribed and translated too, and any value in it that a claim rests on is read independently by models from more than one family before it is trusted: the blind cross-read described below. Some plates still lack part of the full set, and they are listed openly as a backlog. The site is multilingual by design. The reading layer is written in English first, with Dutch and Italian editions following it, and the original Dutch of every quoted document stays beside its gloss instead of being replaced by it.

The site is the last step of a longer pipeline that runs in one direction. A source is taken in and its bytes are preserved. Its text is read and transcribed. A claim is drafted with its support and its uncertainty attached, a person accepts or refuses the change, and only then is it applied to the registry and the site rebuilt from it. Machines do the reading in the middle and people check it. Printed scans are transcribed by a vision model reached through the Gemini API. Any handwritten value a claim will rest on is read independently by models from at least two different families and then confirmed by a person, so no single machine reading of an old hand is ever trusted alone. The acceptance step is a real gate. Automated helpers may fetch and stage material but may never write accepted knowledge on their own; only a reviewed proposal, checked against the existing record, changes what the registry holds.

The outline decides which stories are told and where each deep dive appears. It comes at the end of that pipeline, and it is the clearest single example of how the project used artificial intelligence, because the work was never left to one model's judgement. It ran in stages, and every stage was kept. First a machine digest of the whole registry was put together: a compressed catalogue of every fact, source, entity and open question, so that a writer could see everything held at once. Two large language models were then each given that digest and asked to propose a full outline, independently and without seeing each other's work. Codex drafted one through its command-line tool; Claude Opus drafted the other. A third model, Claude Fable, merged them. It reconciled the two proposals under written rulings that recorded why each choice was made, and set the chapter titles. The owner then approved the merged outline. An independent reviewing model scored the result and filed its objections in ranked order, and the rulings on those objections drove one revision after another. Counting the two proposals as the first round, the outline went through four rounds of drafting and review. Every intermediate version and every review survives, so the outline can be traced back to the reasoning behind it. The digest, the rival drafts, the rulings and the review rounds are all set out in the deep dive on the outline pipeline.

The rules that tie the design to the outline are written as automated checks, so they stay enforced from one working session to the next. A checker compares the outline with the registry and reports every mismatch.

It measures against an exact contract. Every identifier the outline carries, for a fact, a source, a piece of evidence or a deep dive, must lead to a real record. An image slot must hold a visual record, not a text document dropped in where a picture should be. A probable claim must carry its sign of uncertainty, and a living person may not be named without a publication gate. The checker enforces these rules outright: a page that breaks one does not publish. Other rules are a standard the outline is working towards, not a line it has already crossed. Every registered fact should be placed somewhere in the reading layer, so that nothing the research established goes unsaid. Every chapter should carry its evidence column, and every deep dive a picture and at least one reference a reader can follow. No single section should be swamped by more dossiers than it can hold. The checker lists what still falls short, and that running list is kept in the open as part of the work. The pipeline that has to answer to it is the outline pipeline.

Two guards stand at the two ends of the work. The review gate described in the section above decides what may enter the registry as accepted knowledge, and this checker measures what that knowledge may become on the page: one watches the way in, the other the way out.

The checker's last rule asks whether a living person has been given a gate. The harder question behind it, what should never be put on the page at all, is one no check can decide, and it is the subject of the next chapter.