Numeric provenance is the practice of tracing a reported figure back to its origin — the source system, the extraction date, the filters applied and the transformations performed — so that when two documents disagree about the same metric, the disagreement can be explained and resolved rather than argued about.

Why numbers diverge without anyone being wrong

Almost none of these disagreements are errors. They are undocumented differences in definition. Common causes: extraction timing, inclusion rules, currency handling, period boundaries (calendar vs fiscal quarter), status filters (invoiced vs recognised vs collected), manual adjustment, and restatement. Once you see the list, the surprising thing is not that numbers disagree — it is that they ever agree.

What provenance requires

Six pieces of context must travel with every figure: source (which system), as-at date (the moment the data represents), extracted date (when it was pulled), definition (inclusion and exclusion rules), transformations (conversions, adjustments, allocations), and author. A figure carrying all six can be reconciled against any other figure carrying all six, in minutes, by anyone.

What a provenance system does

Detection: sweep the reporting estate for figures claiming the same metric and period, surface divergences. Explanation: reconstruct the likely cause from extraction timestamps, filter settings, formula chains, version history. The goal is converting these disagree into these disagree because one was pulled before month-end close.

Why single source of truth usually fails

People continue producing figures outside the warehouse because it doesn't yet answer their specific question and their deadline is Thursday. Provenance is the pragmatic intermediate step — it doesn't require centralisation, just that any figure can explain itself.