What a customer data platform reporting dashboard should show about evidence health and corrections

A customer data platform reporting dashboard should establish whether its evidence supports a decision before it presents a headline measure. If the view is still being built, leave the evidence panel blank and add one decision card for a defined England journey. Do not enter sample figures that might be mistaken for results.

If an evidence-health field is unknown

Withhold the headline whenever a source, identity, correction or suppression field is unknown. Start with an empty evidence-health panel; it is a decision gate, not a score. Give every field an owner and record its last accepted run before using any result.

The panel should show source freshness, schema and query versions, identity-rule version, failed ingestion, duplicate events, unmatched records, unresolved identity candidates, open corrections, suppression not propagated and destination rejection. These fields expose whether the evidence chain is current and complete enough for the decision.

The ONS describes quality as fitness for purpose in its quality framework. That framework concerns official statistics, but relevance, accuracy, timeliness and comparability remain useful questions for an internal buyer report. Use them to judge whether the evidence suits the decision, rather than to imply that the dashboard is an official statistic.

• Rule of thumb: No evidence status, no headline.

If a measure has no frozen definition

Do not display a measure until its decision and calculation are fixed in a versioned definition. The visible value must point to that exact definition, rather than a similarly named supplier-interface field, so that a reader can establish what was counted and for which period.

Record the decision, event, eligible population, numerator, denominator or unit, Europe/London period, exclusions, deduplication, treatment of missing and late data, correction state, uncertainty and threshold source. Keep these details beside the view with the frozen definition and the decision-card template.

The Government Analysis Function reproducible analytical pipelines strategy promotes version control, documented manual steps, testing and audit trails for government analysis. An England business can borrow those techniques without presenting its dashboard as an official statistic. Keep the decision card and definition together.

• Rule of thumb: Display only what its frozen definition can explain.

If the rows mix evidence types

Keep raw source events, identity decisions, rights states, audience delivery, supplier-attributed credit, reconciled finance and any causal estimate in separate rows. These are different evidence layers, so a single total can hide how a value was produced and what the team is entitled to conclude.

Never add supplier credits from several destinations into a cross-channel total: their attribution windows, identity graphs and model versions may overlap. Keeping each credit in its own row prevents those overlapping windows, graphs and versions from being mistaken for separate outcomes or a combined result.

HM Treasury's Magenta Book distinguishes process monitoring from impact evaluation. A dashboard observation should remain observational unless a qualified evaluator approves the counterfactual and assumptions. Keep any causal estimate in its own row, and do not let a supplier-attributed result imply an impact finding.

• Rule of thumb: Separate what happened from what a model credits or a design estimates.

If a source, query or rule changes

Treat a change to a source, schema, query, model or identity rule as a series break. Mark where it occurs, retain the earlier definition and state whether the new figure is comparable. If a released value is corrected, show both versions, the reason, the owner and the date.

A silent overwrite destroys the audit trail. The NCSC SaaS security guidance covers audit information, service identities, incident detection and recovery. Keep access narrow, and retain logs that can explain who changed a query or export. An observable reporting workflow helps preserve the route through which a figure was produced or corrected.

• Rule of thumb: Preserve the old value and explain every change.

If freshness clocks disagree or one is absent

Record when the source event occurred, when it arrived, when the accepted query ran and when the view refreshed.

Separate clocks reveal ingestion delay; a recently refreshed tile does not prove that its events are recent. If any required clock is missing, mark the measure unavailable and route the feed to its owner.

Never carry forward a previous value without an explicit stale-data marker. If suppression remains unconfirmed downstream, keep that exception visible and assign it to the destination owner. Use a consent-trail audit to examine the related rights state rather than assuming suppression has propagated.

• Rule of thumb: Missing clocks mean unavailable, not current by default.

If the evidence passes but the view may still fail its users

Do not release on data checks alone. The Government Analysis Function dashboard software guidance discusses accessibility and security in a government context. The buyer must test the configured dashboard itself with appropriate users and methods before publication, rather than assuming a technically complete view is usable.

Test table navigation, labels, correction notices and exported evidence. Also confirm each displayed field has an owner and last accepted run, each definition is linked, and previous and corrected values remain visible. Use the quality checks to record the outcome before release.