Metric definitions will change as the factory learns, customer requirements evolve and source capture improves. The risk is not change itself. The risk is an untraceable change that makes a new result look like an improvement or decline against an incompatible history. Data governance is a process of setting and enforcing priorities for managing and using data, which gives definition change an explicit decision right rather than leaving it to whoever edits a spreadsheet first.[S22]
A change request begins with the reason and the decision affected. It distinguishes correction of an implementation error from a deliberate change in operational meaning. Fixing a denominator mapping in code is not the same as deciding that repaired garments now count as first-pass good. The first restores the approved contract. The second creates a new measure version and may invalidate comparisons, incentives and model features built on the old one.
The request records the current definition, proposed definition, source-event implications, owner, approvers and effective date. It identifies every affected surface: source sheet, validation rule, line board, plant report, export, sustainability calculation, incentive rule and analytical dataset. A small factory can keep this register in a controlled workbook. What matters is completeness and accountability, not the sophistication of the tool.
Back-calculation requires evidence, not preference. If the retained raw events contain enough information to calculate both versions, the steward may rebuild history and reconcile the result to the original records. The back-calculation keeps its own version and test evidence. If the old events do not contain the new distinction, history must not be fabricated. The report carries a break, explains the difference and, where useful, presents overlapping old and new measures during a controlled transition.
Parallel operation provides a practical bridge. Both definitions run on the same events for a suitable review window, with differences investigated and explained. The purpose is not to choose the more flattering series. It is to prove that the new rule captures the intended population, that source systems implement it consistently and that users understand the new comparison basis. The transition ends through the same approval route that authorised it.
Migration acceptance uses concrete reconciliation cases. The worked example in the metric contract becomes a test fixture, joined by boundary cases such as partial bundles, repairs, late entries, changed routings and duplicate events. Expected outputs are agreed before release. Source totals reconcile to KPI inputs, KPI outputs reconcile to every report and corrections preserve their history. A visual check of the new dashboard cannot replace these tests.
Historical extracts carry the definition version with the values. This prevents a later analyst from treating differently defined observations as one training population. Models and rules that consume the KPI receive an impact review before migration. If a threshold was learned under one inspection or counting regime, it remains provisional until validated under the new one. The same caution applies to external benchmarks and target commitments.
Retirement is also controlled. When a measure no longer supports a decision, the owner records why it is withdrawn, which reports lose it and how its history remains accessible. Removing an inconvenient KPI without a record damages institutional memory just as surely as redefining it in secret. A useful data dictionary therefore contains active, superseded and retired versions, with their effective intervals and relationships.
GFT's practical rule is that the history of meaning is part of the operating record. The factory does not need a heavy governance bureau to preserve it. It needs one accountable owner, one visible change register, reproducible calculations and the discipline to leave a discontinuity visible when evidence cannot bridge it. That discipline protects every later comparison, from a shift review to an investment model.