01Start with the decision the data supports
Identify the user, the question and the action the answer should inform. Agree the reporting period, relevant population and exclusions. A metric definition should make it possible for two reviewers to reach the same interpretation.
02Reconcile identity across the sources
Map customer, product, location or transaction identifiers and record where matches are uncertain. Review missing values and differing levels of detail. Do not hide a many-to-many relationship behind a join that changes the apparent total. Working example: An event dictionary and delivery cohort definition.
03Make transformation and timing visible
Document source ownership, cleaning rules, accepted late records and refresh cadence. Distinguish event dates from the time a source was updated. Users should know when a report is complete enough for the decision they want to make. For remote-site work readiness, agree how field coordination, equipment specialists and site supervisors contribute where those functions are involved.
04Give quality issues an owner
Set checks for required fields, duplicate keys and meaningful reconciliation totals. Show which inputs failed and who can correct them. Dashboards should expose material limitations instead of presenting uncertain records as equally reliable. Review check: Analysts can reproduce the measure for a selected shipment group.