Documentation drift: 6 documented columns absent from the catalog

Hi — we surveyed public dbt projects, comparing each one's manifest.json (what the YAML documents) against its catalog.json (what the warehouse reported at dbt docs generate time). Your published docs site was one we could reach.

The comparison flagged 6 documented columns across 5 models that do not appear in the catalog. Small list, so it should be quick to confirm or dismiss.

We hand-checked 4 of the 6 against your model SQL; all 4 were real disagreements.

How to reproduce

dbt docs generate
npx clarilayer dbt-check

It reads target/manifest.json and target/catalog.json locally, sends nothing anywhere, and needs no account. Full disclosure: we build that tool, and this issue is the result of running it across public dbt projects.

Caveats worth reading before anyone spends time

• These are disagreements between two files, not bugs we have confirmed in your pipeline.
• dbt docs generate does not rebuild models. If your docs job runs without a preceding dbt run, the catalog can describe an older deployment, in which case some of these reflect undeployed changes rather than documentation errors. Our check cannot tell those apart — a known blind spot on our side.
• Some may be intentional: a column deliberately not propagated, a model you know is not built in the docs target.

Data and method

Corpus, per-project results, and the full false-positive audit: https://github.com/clarilayer/clarilayer/tree/main/survey — your project's record is in per_project.json. Write-up: https://clarilayer.com/blog/dbt-docs-warehouse-drift-survey

No action needed on our account. If any of these are wrong, that is the more useful outcome for us and we would genuinely like to hear it.

Start by reviewing the project's survey record in per_project.json, then run dbt docs generate and npx clarilayer dbt-check to compare target/manifest.json with target/catalog.json. Inspect the five referenced models and their SQL to confirm or dismiss the six reported column differences, while checking whether an older deployment explains them.