Which version of the number should anyone trust?

Which version of the number should anyone trust?

When two systems report different values, the problem is not always bad data. The numbers may use different populations, event dates, currencies, allocation rules or revision policies while sharing the same label. Trust begins by making those semantics visible.

Create a metric contract at the level where decisions occur. Define the business concept, grain, source events, formula, inclusion rules, effective time, owner and permitted uses. Version material changes and state whether history will be restated. The objective is not one physical database; it is one governed meaning that every implementation can test.

Then preserve lineage from reported value to source. Record datasets, jobs, runs and code versions; reconcile critical totals across boundaries; and publish freshness and quality status beside the number. OpenLineage shows how standard metadata can expose production and usage paths across heterogeneous tools, making impact analysis practical before a change reaches a dashboard.

Certify outputs by risk. A regulatory figure may require formal approval and immutable evidence, while an exploratory analysis can remain provisional if clearly labelled. ISO/IEC 25012, reconfirmed in 2025, frames data quality through characteristics whose importance varies by stakeholder�an important reminder that fitness for use is contextual.

Measure definition exceptions, reconciliation breaks, stale consumption, unresolved ownership and time spent debating numbers. When disagreement occurs, resolve the contract rather than selecting the most convenient dashboard. The trusted version is the one whose meaning, provenance and quality are sufficient for the decision being made.