KPI Validation and Reconciliation
A KPI is a promise — test the promise
KPIs are worth nothing un-validated. Tests, control totals, an approval workflow and change audit keep the number real — because a red-light KPI that's 'wrong' poisons the whole dashboard.
What you will learn
- Design validation: definition reviews, control totals, statistical taste
- Run reconciliation gates in CI
- Add approval + audit to KPI changes
What is this?
KPI validation proves the number means what it says: definition reviewed, formula reconciled to a control total, edge cases tested, drift gates in CI, and changes approved and audited. Unvalidated KPIs are vandalism on the board — a wrong red light makes the entire dashboard untrustworthy, so validation has to be part of the definition, not an afterthought.
Why do you need it?
- One wrong KPI silently poisons trust in the whole deck.
- Reconciliation catches errors a dashboard review sneezes past.
- An approval/audit trail turns 'KPIs change' into a governed act.
- Validation is what execs feel as faith in the red light.
Mental model
An airline's pre-flight checklist
KPI validation
Nobody 'approves' a flight by smelling the cabin. The checklist runs: instruments, fuel, control surfaces, margins — every flight, mechanically. KPIs fly on the same contract: definition, control total, edge tests, drift, signoff before the board sees it.
How it works
- Definition review: formula, grain, filters, direction documented and approved.
- Control total: reconcile the KPI to an independent aggregate within a delta.
- Edge cases: zero-dividers, empty periods, snaps of grain — tested.
- CI gates: drift thresholds, freshness and schema checks on the KPI query.
- Approval + audit: changes flow through review; every version logged.
Validated KPI from definition to board.
Examples
A validation telemetry row
kpi: activation_pctdefinition_reviewed: 2026-01-15 (vov)control_total: raw_activation table SUM matchesdelta_gate: <= 1% (pass)edge_cases: empty cohort -> NULL handled, 0->render '--'ci: freshness daily, drift weekly, schema tests greenapproved_by: platform-leads (2026-01-16)history: [v0 2025-11 init, v1 2026-01 filter fix]The full validation contract: defined, reconciled, edge-tested, approved, audited — so the red light can be believed.
In production
- A revenue KPI re-reconciles nightly against the control; a +2% drift blocks the board deck.
- An activation KPI edge-case rule (blank cohort → 'no data') stops zero-from-nobody silliness.
- KPI version audit logs answer 'why did revenue change' in a PR, not a guessing game.
Best practices
- Build the control total into the same pipeline so reconciliation is continuous.
- Gate KPI changes (approval + CI) like code changes.
- Test edge cases explicitly: zero, empty, grain snap, division guards.
- Ships changes with a versioned change log; audits forever.
Common mistakes
- Dashboard 'review' passes, control-total test never runs.
- Zero-from-empty cohorts rendering as red '0%' — a fake crisis.
- KPI drifted from its stated definition with no one approving.
- Only reviewing the happy path; denominators crash later.
Debugging tips
- KPI suddenly off by a hair → run the reconciliation gate; suspect grain/definition drift.
- Red light for an empty slice → the edge-case test should have blocked it.
- Everyone 'remembers' changing something → no audit trail; add versioning.
Interview checkpoint
Quick quiz
Which gate proves the KPI matches an independent aggregate?
Empty-period cohorts should render as a dramatic red zero.