A management report that accepts only approved figures

A management report often draws data from systems that do not communicate, creating hours of export work. It also carries a quieter risk: a number added in commentary, pasted from a draft, or invented looks exactly like a sourced figure in the finished document.

Model-generated commentary increases that risk. A draft can introduce a number that exists in no approved source and present it in polished, convincing prose.

A gate built into publication

In the management newsroom, commentary reaches publication only through a gate that knows the approved figures and their sources.

The controlled case uses generated data, contains no client material, and writes no operator decisions.

The report schema separates approved figures, each with its source, from editorial commentary. Reconciliation follows fixed code with an explicit tolerance, so the same data always yields the same verdict stored with the figure.

The system's heart is the renderer. It builds the set of allowed number representations exclusively from approved figures, in every format a number can take in text. Commentary containing a detected digit sequence outside that set is rejected before publication. There is no blocklist of forbidden values to maintain; there is an allowlist that follows from the data.

Discrepancies between sources beyond tolerance never disappear into reconciliation. They remain visible as annotations in the report, with provenance, for the data owners to explain.

Why publication control uses fixed rules

Reconciliation and publication control use no language model. Comparing figures against a tolerance, checking references, and rejecting an unsupported digit are deterministic tasks. A model may draft commentary upstream, but its text receives exactly the same digit and source checks as human-written commentary.

The allowlist therefore protects both drafting paths. The author can be a controller or a model; publication still depends on the approved figures and valid references.

Begin with the metric dictionary

A publication gate protects figures only after the organization agrees on what they mean. Revenue, margin, active customer, and plan attainment need an owner, formula, period, currency, adjustment rules, and source system. Without this contract, two correct queries may produce different answers.

A minimum metric record includes its business name, definition, query or calculation rule, refresh frequency, owner, and permitted display formats. Each report edition should retain the metric-definition version it used.

Implement in four stages

1. Inventory the sources, metrics, exports, and manual adjustments in the current report.
2. Reconcile by defining tolerances and owners for differences across ERP, CRM, and spreadsheets.
3. Control publication by separating approved figures from commentary and deriving permitted representations.
4. Run in parallel for several cycles, producing the current and new reports and reviewing differences with the controller.

Every stage should leave a reviewable artifact: metric definition, reconciliation verdict, source reference, and publication decision.

Edge cases a digit gate can miss

Numbers appear as percentages, negatives, ranges, dates, identifiers, abbreviations such as “1.2m,” and rounded values. The gate needs explicit equivalence rules and must distinguish narrative digits from measures. Currency and unit belong to the approved value.

A subtler failure uses an approved number in the wrong context. A set of allowed representations cannot prove that revenue was not labeled as margin. Commentary should therefore reference metric identifiers, and a controller should approve meaning. Digit control is a strong technical layer; it does not replace subject-matter review.

The measurement

The control test stopped all three prepared foreign digits and the missing figure reference. Reconciliation produced the expected verdict for all five prepared discrepancies. Across 2,000 warmed in-memory calls, the renderer's 95th-percentile time was 0.007125 ms. A separate small-model commentary draft using gpt-5-mini consumed 396 units of model text, took 2.00 seconds, and cost about PLN 0.0010 before passing through the gate.

These checks cover the renderer and prepared reconciliation cases. They do not measure source-system authenticity, every numerical notation, the quality of commentary, or the time required to assemble a client's full report. Those depend on agreed metric definitions, integrations, and the controller's workflow.

The cost of one commentary draft is usually small beside data collection and reconciliation. A pilot should track time from source close to publication, manual adjustments, discrepancies without owners, rejected commentary, and the effort required to defend a figure during management review.

Readiness checklist

• Every metric has a definition, owner, and version.
• Every figure leads to a source or executed query.
• Reconciliation tolerances are explicit and testable.
• Commentary references metric identifiers.
• Percentages, currencies, ranges, and rounding have test cases.
• Discrepancies remain visible until the data owner decides.
• Publication records the data, commentary, and control versions.
• The process has a rejection path and a manual fallback.

Details and screens on the case page. If your board report takes hours and the question of where a number came from can stop a meeting, a free process scan is a good first step.