Quality is a set of answerable questions about a record: is the field filled, is the value well formed, do the fields agree, is it one record, is it true of the world, is it current enough.
The order is the order in which they are usually fixed: fill the fields, then make them well formed, then make them agree, then merge the duplicates, then verify against the world, then keep them fresh.
| Dimension | The question | Measured by | A failing supplier record |
|---|---|---|---|
| Completeness | Is the field filled? | Presence of a value in the fields that apply to this record type and status | An active supplier with no tax id |
| Validity | Is the value well formed? | Format, list membership, range and checksum rules on one field | An IBAN that fails the checksum; a country code outside ISO 3166 |
| Consistency | Do the fields agree? | Row rules across fields of one record, and across records that refer to each other | Country GB with a VAT prefix of DE; a payment term absent from the terms list |
| Uniqueness | Is it one record? | Match rules on the domain keys; the cross-reference of source keys | Two supplier records with one tax id |
| Accuracy | Is it true of the world? | Verification against a registry, a postal service, a bank or a partner's own record | A registered name that the company registry spells differently |
| Timeliness | Is it current enough? | Age of the last verification or change against the freshness the consumer needs | Bank details last verified two years ago |
A score is a ratio, so it can be compared across sources and over time.
For one record, the score on a dimension is the number of checks passed divided by the number of checks that apply to it. A check applies according to the record's type and status: a blocked supplier is excused the bank checks, an inactive material is excused the planning-parameter checks. The overall score of the record is the same ratio across all its applicable checks.
A column's score is the share of records passing the checks on that column; a source's score is the share across its columns; a domain's score across its sources. The roll-up is arithmetic, so a change in one rule shows at every level the next time the agents run.
Two weights refine the plain ratio. A field weight says how much a failure costs: a wrong bank account weighs more than a missing fax number, and the weights are set per domain by the owner. A consumer weight says who depends on the field: a field read by planning every night weighs more than one read by a quarterly report. The profiling page shows the weighted scorecard.
One source's scorecard on the showcase estate: six dimensions, one number each. Consistency is the weakest, which points the first rules at the row checks. Showcase estate.
The catalogue entry for each domain names the checks that carry the most weight. Three examples.
| Domain | Dimension that carries most weight | Why | Sample check |
|---|---|---|---|
| Supplier | Accuracy and timeliness of bank and tax details | Payments and compliance depend on them; changes are a fraud vector | Bank account verified within twelve months; two-person rule on changes |
| Material and product | Completeness of planning and logistics attributes | A material with defaults is planned wrongly | Lead time, minimum order and planner set for active plant materials |
| Customer | Uniqueness and consistency | Exposure and revenue add up only to one record per counterparty | One record per tax id; billing and shipping countries consistent with the tax registration |
Six cover the questions the business asks; anything finer is a rule inside one of them. A lineage question (where did the value come from) is answered by provenance on every value, which is a property of the record itself.
Excusals are rules too, with an owner and a history, and the scorecard shows the excused count beside the score. A domain whose excusals rise is visible to the governance lead.
On every run of the rule agents, which is nightly by default and on change for the fields that carry the most weight. Trends are kept per day.
Bring one domain and two systems. In a week the sources are profiled, the ten costliest fields are under rules, the first corrections are written back with their acknowledgements and the publish gate is on.