AGENTS.md — South American Rural Cadastral Boundaries

Guidance for AI agents and automated clients that work with this catalog.

One rule survives every edit to this file. Every claim here is quoted from a source or measured from the data. If you cannot point at where a fact came from, it does not belong in this file. An agent that acts on an invented join key or an invented column name gives a confident wrong answer, and nothing downstream catches it.

Every count below came from a query over all rows of the published file. The queries are in the per-collection agent guides.

What This Catalog Holds

Root: https://data.source.coop/tristangruppwri/cadastral/catalog.json

Three national rural cadastral parcel collections, each a single GeoParquet 1.1 file in EPSG:4326 with a bbox covering column and ZSTD compression.

brazil-car-area-imovel carries a second asset, suspect, with 2 rows. It is GeoParquet 1.0.0 in EPSG:4674, not EPSG:4326, and it has no bbox covering column. It is a quarantine sidecar, not part of the main dataset.

How To Read It

No credentials. Each file is one object over HTTPS with range support, so a query reads only the row groups it needs.

There is no partition:glob anywhere in this catalog. Every asset is a single file, so plain HTTPS is the whole access path and nothing here needs s3://.

Join Keys

No collection joins to another. The three cadastres are separate national registries with separate identifier schemes and no shared key.

Inside a collection, no identifier is unique. Measured over every row:

cod_imovel is the join key from Brazilian CAR property boundaries to the other CAR theme layers, which this catalog does not hold. It is the closest thing here to a primary key and it still repeats on 15,614 rows.

Quirks That Produce Silently Wrong Answers

No key is unique. See the table above. COUNT(*) counts geometries, not properties. A join on any of these columns fans out.

The Brazilian suspect sidecar is in EPSG:4674, not EPSG:4326. It also declares GeoParquet 1.0.0 while the main file declares 1.1.0. Reading both into one frame mixes coordinate systems. Its 2 geometries have deliberately impossible coordinates, reaching latitude 239,074,544.

Paraguay dpto is a letter code, not a name. The published values are single characters such as L and K. There is no department-name column.

Paraguay obs is uncontrolled free text copied from the source registry. 886,763 of 2,192,890 rows are not blank. Its content is not a controlled vocabulary and does not parse. The same text is repeated as the last pipe-delimited field of clave_comparacion in 822,579 rows, so a query that drops obs still returns the text through clave_comparacion.

Brazil des_condic has 16 distinct values, not 10. Documentation that lists ten statuses is incomplete. Five of the sixteen begin with "Cancelado", and together they cover 152,344 rows.

Brazil nom_tema is constant. All 8,453,552 rows carry Area do Imovel. It is not a filter.

Uruguay AREAHA is an int64, not a float. A parcel smaller than one hectare rounds. AREAMC, in square metres, is a double.

Every parquet object is served with Content-Type: binary/octet-stream. The collections declare application/vnd.apache.parquet. A client that dispatches on the served header, rather than on the STAC declaration, sees an opaque blob.

Licence

Every collection declares other, not an SPDX identifier, and carries a rel:license link. Read that link before you redistribute anything. Two of the three upstream sources publish no licence at all. This is an open question, not a settled value.

Structure

Assets and structural links resolve relative to the object that carries them. Catalogs here carry no self link, so a client tracks its own location.