Reproducibility, versioning and file formats

Who this is for

Anyone checking a claim, reusing the data, or trying to reproduce a figure. None of it is needed to read the atlas: every layer’s historical argument is on its own page, and this is the machinery underneath.

Versions identify bytes, not dates

Every published layer carries a version string like 4440ae3946c6. It is the first twelve characters of the file’s SHA-256, so the version is a fact about the content, not a release number someone chose.

Two consequences worth knowing:

A given versioned URL always returns the same bytes. If the data changes, the version changes with it, and the old URL is still a different file.

There is no “accessed on” date in our citations, deliberately. An access date tells you when someone looked, which is a poor substitute for telling you what they looked at. Release ID and checksum pin it exactly, and two identical builds produce identical citations.

The release manifest

public/geo/manifest.json is the join between a layer and its file: format, CRS, geometry types, feature count, byte size, SHA-256, and the content-addressed URL. The whole set carries a release ID such as geo-015d45320b25c121, which changes whenever any asset does.

To verify a download, hash it and compare with the manifest entry. If they match, you have the bytes this atlas drew.

Formats, and why there are several

Format	Used for	Why
GeoJSON	Most layers	Plain text, readable in anything, fine at these sizes.
FlatGeobuf	Larger vector layers such as the Venetian family	Indexed binary, streams by bounding box rather than loading whole.
PMTiles	Reserved for tiled layers	Single-file tile archive served by range request, no tile server.

All are static files on a CDN. There is no database behind this atlas and no API to rate-limit: the delivery model is deliberately boring, and the reason it can stay boring is that catalogue organisation never required a backend.

Where the schemas are

Field definitions in reader’s language are in the data-fields pages for Dacia, the Venetian network and the Hanseatic world. The exact machine schemas, with column types and compile contracts, stay in the repository - they answer a different question, and copying them into public prose would mean two copies drifting apart.

Reproducing a layer

Every layer is compiled from committed source data by a script in the repository, and the compile is deterministic: the same inputs produce the same bytes, which is what makes content addressing meaningful. The publication guide in the repository documents the pipeline end to end.

What is not here

Governance and planning material - tickets, specifications, delivery decisions - lives in the project’s Confluence and Jira and is not part of any public reading path. Nothing on this site requires an account to follow. If you ever hit a login while trying to understand a layer, that is a defect.