Verifiable evidence chains for large-scale ecological modelling.
The Protect Baltic WP3 species distribution models are published, versioned and citable — 630 species, 7 product families, 4 410 raster products across fish, macrophytes and invertebrates. A DOI identifies the dataset. It does not identify a claim derived from it. This showcase follows one 250 m cell of the Baltic down to a single value and shows which evidence travels with it, and which does not.
One location, held constant. Change the species and the answer changes, not the place. Every figure comes from files retrieved from the published archive; showcase declarations are marked as constructed.
Compare the same location across species.
Row 3387, column 1926 — centre 4691222.4269 / 3965570.1766 in EPSG:3035, which is 58.642°N 16.374°E, Bråviken on the Swedish east coast.
Two published predictor stacks are available on the same 250 m grid as the species raster. Following one row across the species raster and both predictor stacks shows where the spatial evidence chain can be bound and where it cannot.
Same grid, same row, same publication chain. On this row 1 801 cells carry a species value; 1 709 of them also carry all 27 predictors, and 92 carry none. No cell carries predictors without a species value. Why the environmental stacks are NoData where the species raster is not is not represented in the artefacts examined.
| identity | VERIFIED | artefacts bound by DOI, size and publisher md5 |
| bytes | VERIFIABLE | retrieved by HTTP range, CRC32 checked after inflate |
| cell | VERIFIED | identical CRS, dimensions, pixel size and origin across all three artefacts |
| value | VERIFIED | reproduced by two independent runs on two machines |
| name | VERIFIED | all 27 predictor names read from the artefacts' own metadata |
| meaning | NOT ESTABLISHED | see below |
depth = 3852.9119
Unit and semantic definition not established from the identified published evidence examined. The name is machine-readable. The value is machine-readable. What the value means is recorded in no artefact reached from the publication metadata: not in the raster, not in the two deliverables titled modelling code, and not in the referenced repository, where the predictor appears as an input raster rather than as code that defines it.
This is not a claim that no definition exists. It is a measured statement about which evidence was examined and what it did not contain, anchored to the repository commit at the time of measurement.
predictor_list.csv in the published repository identifies
depth as a predictor for all three species groups represented in
that file.
EVE was not used to generate the ecological models, and it does not determine whether they are scientifically correct. It was used as the verification framework for separating and binding the parts of the evidence chain.
Keeping those layers apart surfaced four things that are easy to lose in ordinary analysis:
A cited cell value carries no trace of the expert review recorded for that species — including notes the modelling team wrote themselves.
The binary threshold is derived per species, so a higher raster value can carry a worse suitability class.
Where two product families hold the same content, a content hash gives both the same identity and cannot establish which role an artefact played.
A predictor's name and value can be read and reproduced from the artefact, while what the value means is recorded in none of the published evidence examined.
EVE separates these layers so that one cannot silently stand in for another.
All data used here is publicly available under CC-BY-4.0 and was produced by the Protect Baltic project, funded by the European Union's Horizon Europe programme.