Review source quality
GraphArch preserves RDF statements even when sources describe the same resource differently. Multiple labels or definitions are not automatically RDF errors. They can represent alternative names, editorial disagreements, or different perspectives that deserve review.
Publication planning identifies distinct values for the same resource, annotation predicate, language tag, and datatype. It checks labels, preferred labels, titles, definitions, descriptions, and Markdown descriptions. A French translation alongside an English label does not trigger this finding. Repeated identical statements from several sources do not trigger it either.
A finding retains every competing value and its source. Git sources identify
the exact commit and file. The KG stores these as warning-level ga:Diagnostic
resources, with diagnostic fragments, the affected resource, source provenance,
and the observed revision. Existing documentor warnings, including missing
labels and definitions, are retained with their immutable documentation
contribution and included in the same report.
What reviewers see
An affected Markdown resource description receives a warning callout containing the competing statements. Specifications and PDFs include Quality issues as a separate numbered appendix. If Acknowledgements is appendix A, the quality report becomes appendix B and has its own entry in the contents.
The report groups findings by warning type, such as No description found, and lists the affected resources underneath each heading. For competing annotations, each resource also retains the alternative statements and their sources. This lets reviewers address one kind of issue across the publication without reading the same warning repeatedly.
Missing-annotation warnings refer to the collected annotations selected for that resource, not to one unspecified file. For example: No English or language-neutral label or description found in the collected sources. A label without a language tag is accepted, as is an English label. A Dutch-only label does not satisfy an English publication. Missing labels and descriptions remain separate machine-readable diagnostic codes, while a resource missing both gets one concise inline warning. These findings do not mean the RDF is invalid.
Source references recognise rdfs:label, skos:prefLabel, and sh:name as
names; dct:title or the older dc:title supplies a name when no label is
available. Preferred SKOS labels take precedence over general labels. Plain
source descriptions can use dct:description, dc:description, sh:description,
skos:definition, prov:definition, or rdfs:comment; the EKGF
markdownDescription annotation supplies rich explanatory documentation.
Prefix spellings such as dct: and dcterms: make no difference when they
expand to the same predicate IRI.
Companion websites provide a separate report linked from their footers. Paginated specifications link to the report on their overview page. Rich descriptions and different versions of the truth are all retained; GraphArch does not choose one as the truth.
A publication renders the findings pinned to its plan. Exporting the KG, removing the source files, and rendering the restored graph retains the report. A later source edit cannot change an earlier publication's findings.
Use the report in CI
Warnings appear in the publication manifest as well as the KG. Warning-only
publications succeed and are recorded as published-with-warnings. The same
rule applies to build-family, build-site and publication render: each
reports every diagnostic on standard error and exits nonzero only when one is an
error. Rendering an exported KG therefore reaches the same outcome as the build
that produced it.
--fail-fast and --keep-going continue to govern actual execution errors;
unsupported markup, malformed input, and missing required release content
remain errors.
Teams can review the report on each CI publication and resolve disagreements in the source repositories. If several perspectives are intentional, the report keeps them visible. Consumers that enforce stricter editorial rules can inspect the manifest's diagnostic codes and severities in their own CI policy.
Reports are bounded to 1,024 findings per target and 128 value/source records per competing-annotation finding. Exceeding a budget fails explicitly instead of publishing an incomplete report.