Corpus analytics
Weighted toward what can be wrong. A corpus that publishes a false claim is worse than one that publishes less, so coverage and defect rates are shown together.
Trim scoping by brand
Rows whose trim is Unspecified
cannot be filtered by trim, which is most of the product. Extraction fixes only reach
these numbers after a re-ingest.
Availability mix by brand
standard
available
not offered
unstated - no claim made
A brand with no not_offered rows
is not necessarily better documented - Subaru's bullet lists never assert absence at
all, so everything it omits is unstated.
Coverage by model year
Catalog cluster sizes
Clusters sitting exactly at the 20-alias cap are the signal that merging is still too eager - the cap is hiding how large they would otherwise be.