How do you prove where a number came from?

For a business figure produced from data, by keeping four links for it, made when the number is produced. First, the rule that computed it, and which version of the rule. Second, the records it used. Third, for each record, the source document it was read from, such as the settlement PDF, scale ticket or invoice. Fourth, when it was run. With those four, anyone can follow a figure in a board pack or an AI answer back to paper. For figures that rest on judgement, such as accruals and estimates, the proof is different: the assumption, who approved it, and when.

What the usual answer says

Guides on data proof point to lineage tools, audit trails, documentation of sources and transformations, version control and governance. All help. Most describe how data moves between systems, and say less about what a complete proof of one number needs to contain.

Where auditors are heading

Auditors already test the information a business gives them. The body that sets international auditing standards, the IAASB, is revising the audit evidence standard in an Audit Evidence and Risk Response project. It builds on work begun in March 2019, and the revised standard is not yet final. Among its aims: to "Enhance the application of professional judgment and professional skepticism exercised by auditors" when "making judgments about information intended to be used as audit evidence", and to "encourage, auditors' use of technology in obtaining audit evidence and evaluating its sufficiency and appropriateness".

For a business, a figure with its links attached is quicker for an auditor to test than one that has to be sampled back to source by hand.

What AI guidance says

Plain-language AI answers raise the stakes. A manager who asks what did we pay for stumpage on this tract last quarter? gets a figure with no visible working. A 2024 federal generative AI risk profile asks organisations to record, for the AI systems they run, "data provenance information (e.g., source, signatures, versioning, watermarks)", and to "Maintain records of changes to content made by third parties to promote content provenance, including sources, timestamps, metadata". That guidance is about systems and content. Our extension to single answers is that a figure an AI gives about stumpage, volume or margin should come with the records it summed and the documents behind them.

Trust and governance

In a survey of 505 data and analytics leaders by Drexel University's LeBow College and Precisely, which sells data integrity software, published in January 2026, 71% of organisations with governance programmes reported high trust in their data, against 50% without. That is trust as reported, not proof, and the report itself warns that its year-on-year comparisons are affected by a change in who answered. Links are what let anyone check whether the trust is earned.

The four links, in practice

The rule and version is the definition used: which settlements count as stumpage, which dates, which unit. When it changes, old figures keep the version they were built with. The records are the rows used: settlement lines, ticket numbers, ledger entries. They can be stored as identifiers. Or the figure can be made reproducible instead, by keeping an unchanged snapshot of the source data and rerunning the rule against it, which proves the same thing without storing row lists. The documents are links from each record read from paper or PDF to the page it came from. This link fails when paperwork is keyed by hand and the image isn't kept. The run time explains why a later figure differs.

A simple test

Pick any figure from last month's management pack and ask for its rule, its records and the documents behind them. If the answer comes back as links, the business can prove its numbers. If it comes back as a spreadsheet someone rebuilt, it can't yet. Run the same test on an AI answer whose value you already know, and check whether it names the records it used.

Judgement figures

Accruals, estimates and allocations are where most disputes about numbers start, and records can't prove them. For those, keep the assumption in writing, the evidence it was based on, who approved it and when, and link the figure to that note the way a computed figure links to its records.

How Quarri works explains the platform as a layer over existing systems, not a migration.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.