A supplier table has 12% of rows with a blank tax identifier and 3% where the identifier is present but does not match the registry. Both are called data quality issues in the same report. Why does separating them change who is asked to fix them?

The mechanism

A quality dimension is a hypothesis about where the defect was created, and that is its whole value. Reporting one aggregate "quality score" destroys exactly the information needed to act.

Blank means the value was never captured. The fix is upstream of the data entirely: a required field in the onboarding form, a validation at the point of entry, a supplier who was never asked. That work belongs to whoever owns the collection process — procurement, in this example — and no amount of pipeline engineering creates a value that was never provided.

Present-but-wrong means the value was captured and disagrees with an authority. The fix is a reconciliation process against the registry, a decision about which source wins, and a workflow for the exceptions. That belongs to a data steward with the authority to say which system is right.

A third dimension hides inside the second. If the identifier matches a registry entry but the registry entry belongs to a different company, the value is accurate and not valid for this record. That is a matching problem, and it goes to whoever owns the match rules.

Why the other options fail

• "Accuracy is always more severe." Severity comes from use, not from dimension. A blank tax identifier blocks a payment today; a wrong one fails an audit next year. Which matters more is a business question.
• "Completeness can be fixed automatically." Filling blanks automatically is usually the worst option available, because an inferred value is indistinguishable downstream from a captured one. A default that looks like data is worse than a null.
• "Both belong to the platform team." This is the failure mode the dimensions exist to prevent. A platform team can measure both and fix neither, and a quality programme that routes everything to them produces dashboards and no improvement.

When this is the wrong level of rigour

For a dataset with one consumer and one producer who sit in the same team, the taxonomy is overhead: they will say "12% blank" and fix it. The dimensions start paying when the producer and the consumer are different teams, because then the argument about whose problem it is happens whether or not you have the vocabulary — and the team with the vocabulary wins it in a meeting rather than over a quarter.