Methods
How the atlas is made
Every record in the atlas is a public record, and every record can be traced back to the place it came from. This page describes where the records come from, how one monument is told apart from another, how a site is kept out when it is not old enough, how the text is written and checked, and what the three-dimensional views are.
The records we read
The public sources, each with its own licence
Nothing is ingested from a site that does not publish for reuse. Each source below is used for one thing, and the licence column is the licence the atlas publishes under.
- Wikidata
The structured record of a monument: its name and other names, its type, its country, the statement of when it was made, and the references those statements carry. The list of monuments comes from here too: the atlas asks Wikidata which items are of a kind of monument at all.
What it is used forThe site list, the facts, the links
LicenceCC0 1.0
- OpenStreetMap
The map. The atlas asks the map what monuments it holds and merges what it finds with the record; it is also where a reader is sent when they want to see the place for themselves. The map never decides what a monument is called, how old it is, or where it stands: a merged site keeps the record and gains the map identifier.
What it is used forThe map records merged into the catalogue, and the map a reader is sent to
LicenceODbL 1.0
- Wikipedia
The reading. The encyclopaedic text of a monument, in the language it was written, linked from every page that draws on it.
What it is used forThe text a summary is written from
LicenceCC BY-SA 4.0
- Wikimedia Commons
The photographs. Every picture is hot-linked from Commons, with its author, its licence and the page it lives on.
What it is used forThe gallery on a site page
LicencePD, CC0, CC BY, CC BY-SA
- Radiocarbon dates (p3k14c)
The measured dates. The p3k14c database holds 173,946 archaeological radiocarbon dates with the places they were taken. A date close to a monument, and consistent with what is otherwise known of it, gives that monument a measured range.
What it is used forA measured range and its method
LicenceCC0 1.0
- Natural Earth
The country table. Which country a monument is in, and the area each of the map queries covers, come from Natural Earth's 1:50m countries. Where the record does not name a country, the table supplies one.
What it is used forThe country a record is in, and the areas the map queries cover
LicencePublic domain
- OpenFreeMap
The tiles the map is drawn from. OpenFreeMap serves OpenStreetMap's coastlines, roads, water and place names as vector tiles the browser draws itself, with no key and no account.
What it is used forThe map under the globe and on a site page
LicenceODbL 1.0
- NASA Blue Marble
The earth from orbit: the ground and the sea under the globe, before the map's own lines are drawn over it. NASA releases it for anyone to use.
What it is used forThe ground under the globe
LicencePublic domain
- UNESCO World Heritage
Linked, never copied. Where a monument is on the World Heritage List, the atlas links the listing so a reader can read it there. The listing's own text, data and photographs are never republished.
What it is used forA link on the site page
LicenceNot republished
- Excavation reports and published papers
Cited, never copied. A date that comes from a dig is read from the report's own words: the year or the range, the method that produced it, and the report itself, linked from the page. The text of the report is not republished.
What it is used forA dated range and its method, with the report that gives it
LicenceCited, not copied
Matching
Telling one monument from another
The same monument is described in several databases under several names. The corpus is built to fold those into one page rather than to publish the same stone three times.
One Wikidata item is one site
A monument that has a Wikidata item is one record, however many names it answers to. Alternative names are kept as alternative names.
A map record that names the item is the same site
An OpenStreetMap element that carries the same Wikidata identifier is merged into that record rather than published beside it.
Otherwise, near enough and similar enough
A map element within 75 metres of a record, with a similar name and the same type of monument, is treated as the same site. The threshold is deliberately generous, because a duplicated page is worse than a merged one.
The first records keep their addresses
The sites the atlas started with keep the identifiers and the addresses they have always had, so no link that ever worked stops working.
The age gate
Old enough to be a megalith
The atlas holds megaliths, so a site has to be old. A record is kept out when it is plainly a modern object, and the exclusion is written down with its reason so it can be undone.
A name that marks a modern object
A war memorial, a statue of a named person, a replica, a boundary stone, a wayside cross: a name that marks the object as modern keeps the record out, whatever the map says about it.
A date that begins too late
A site whose own date begins after 1026 CE, and for which no other sourced claim reaches back before then, is kept out.
A rescue by an older claim
A date from excavation, from a written record, from the encyclopaedia or from an alternative claim all count. If any of them reaches back far enough, the site is readmitted.
Every exclusion is recorded
Each excluded record is written to a file with the reason it was excluded, and a reserve of previously excluded records is kept so that a later claim can bring a site back.
The text
How a summary is written, and what stops it going wrong
A site with nothing written about it still has a page: the facts, the dates and the sources. A site with something written about it may also have a summary, and a summary is the one part of a page that a language model helps with. This is how that is done, and where the line is.
The model is given the source, and nothing else
The draft is written from the text of the sources cited on that page, and from nothing else. It has no memory of the monument, no other sources, and no access to the wider world.
The rules it is given
Use only what the source says. Write in its own words rather than copying. Never write a number that is not in the source: every year, height and count must be copied across. A date is always a range with a method, never a single invented year. An alternative claim goes in its own block with the source that makes it and how it stands, and never stands in for the mainstream account.
An automatic check, before anything is published
Every draft is checked twice. First, that it fits the record: an allowed dating method, an allowed status on a claim, a label and a value and a source on every fact, a summary of a sensible length. Second, and more strictly, that it is grounded: every number and every year in the draft must already appear in the source text, and the wording must not copy a run of the source's words. A number that is not in the source is a fabrication, and a draft that contains one is rejected whole.
A rejection is a rejection
A draft that fails the check is not published. It goes back once with the reason it failed, and if it fails again the site keeps no prose at all. The structured facts — type, place, country, dates, sources — are still published, because those come from the records and not from the model.
The fallback
A site with no surviving draft shows the facts, the dates and the sources, and says plainly that no summary has been written. That is the honest state of most of the catalogue, and it is a better page than a plausible story.
The 3D view
An interpretation, drawn from the type
Every site page can be opened in three dimensions. What it shows is a reconstruction, and it is labelled Interpretive 3D wherever it appears.
Built from the type, not from the stone
The models are parametric: a chambered tomb, a stone circle and a carved figure each have a procedure that builds them from the proportions their type is known by. They are placed on the real terrain at the real point. They are not surveys of the monument in front of you, and they do not pretend to be.
Honest about the hand
A model that has been placed by hand rather than by procedure says so, and says how far its position may be out.
Never a photograph
A generated reconstruction is never shown as though it were a photograph of the site, and it is never used as the evidence for a fact about the monument.
The lines we do not cross
What the atlas will not do
- Publish a date for a monument no source dates.
- Present an alternative or legendary claim as the established age of a site.
- Let a model assign a date, a grade or a claim of its own.
- Show a photograph of human remains as an illustration of a site.
- Copy a text or a photograph whose licence does not allow it.
- Take records from a partner's database, or from a source that does not publish for reuse.
- Ask a reader for an account to see a public record.
The four levels of confidence behind every date range.