How to validate every statistic in a proposal

Inventory

Find the numbers that ordinary proofreading misses

Inventory more than digits. Search for written numbers, percentages, ratios, ranges, dates, durations, currencies, units, ranks and sample sizes. Include quantified adjectives such as “half,” “double,” “majority,” “top three,” “fivefold,” “near zero” and “industry-leading.” Inspect axes, legends, icons, callouts, alt text, footnotes and filenames. A chart can carry a material result that never appears in body copy. A section heading can say “30 percent faster” after the paragraph was corrected to 18 percent.

Classify each figure by origin and consequence. Published third-party statistics need authority, date, applicability and citation. Internal operating metrics need a stable definition, source system and approved extraction. Client case-study results need measurement detail, contractual or client disclosure permission and contextual limits. Derived figures need a reproducible calculation. Estimates and forecasts need assumptions and uncertainty. Targets need an authorized future commitment and measurement plan. Buyer-provided values should be attributed and not silently recast as bidder facts.

Assign one statistic ID across all appearances. The record lists source wording, approved proposal wording, display value, full-precision calculation if relevant, owner, approver, expiry and every response location. This prevents an executive editor from updating prose while the old chart survives. It also separates two superficially similar values, such as annual transactions for the whole service and eligible transactions in the proposed automation scope. Shared units do not make quantities interchangeable.

Claim record

Reconstruct the population, denominator, method and time period

Open the full source at the exact table, page, query or approved record. Capture what was measured, the eligible population, observed sample, geography, baseline and comparison periods, unit, statistic, numerator, denominator, missing-data treatment and exclusions. Record whether the result is raw, seasonally adjusted, weighted, normalized or modeled. A “4 percent error rate” is not interpretable until the reader knows whether the denominator is documents, fields, transactions or audited samples and whether repeated errors within one item count once or many times.

Check time semantics. Publication date is not observation period. A 2026 report may present 2024 data. A twelve-month total is not a monthly run rate, and annualizing one high-volume week needs an explicit seasonality assumption. Check geography and entity. A regional subsidiary’s performance cannot automatically represent the group. Check system boundaries. A platform uptime figure may exclude scheduled maintenance, third-party services or customer-controlled components. Those definitions determine what the proposal may say.

Use the source most capable of supporting the claim. A search result, press article or vendor summary can lead to evidence but should not replace an available primary publication. Inspect revisions and footnotes. Statistical series may be rebased; case-study figures may be superseded; internal dashboards may apply current logic retrospectively. The Office for National Statistics defines statistical quality in terms that transfer well to bid review: relevance, accuracy and reliability, timeliness, clarity, coherence and comparability. A number can be arithmetically correct and still fail the intended comparison.

• Capture observation period separately from publication date.
• Name population, sample, geography, entity and system boundary.
• Preserve numerator, denominator, unit and missing-data rule.
• Record adjustments, weighting, exclusions and uncertainty.
• Check whether a newer source revised the value or definition.

Verification

Recalculate the number and test every word around it

Reproduce each derived value from controlled inputs. Check signs, units, order of operations, filters, grouping, duplicates and missing values. For a rate, recompute numerator divided by the correct denominator. For change, calculate both absolute and relative forms and choose the one stated. A rise from 20 percent to 25 percent is 5 percentage points or 25 percent relative growth, not “a 5 percent increase.” For an average, identify mean, median or another measure. For a range, confirm whether bounds are observed minima and maxima, confidence limits, percentiles or scenarios.

Preserve full precision in the calculation record and approve one display rule. Round once at presentation, not at every intermediate step. Check whether comparison values were rounded under the same rule. Avoid a sentence whose arithmetic appears false because chart labels and narrative use inconsistent precision. Conversions require source unit, target unit, factor, factor date and owner. Currency conversion also needs the chosen rate basis and date, but validating tender price itself remains part of the commercial pricing control.

Test the grammar. “Of respondents,” “of all customers” and “of completed cases” name different populations. “Up to” describes a bound, not a typical result. “Average saving” may falsely suggest every client achieved it. “Proven” can overstate an observational association. “Independent” requires an independent source or method. “Benchmark” needs a comparison population. Check the claim-support direction: the source must entail the proposal sentence, not merely discuss the same topic.

Apply consequence-based assurance. An analyst may verify an ordinary contextual market statistic. The authorized service owner should approve an operational performance claim. A named client result needs the case-study owner and disclosure authority. A safety, regulatory, financial-capacity or contractual service-level number may need specialist and executive review. The updated AQuA Book distinguishes verification of correct execution from validation that analysis is fit for its intended use. Both are necessary when an evaluator may rely on the result.

Release

Keep the approved meaning intact across text, charts and submission files

Validate the visual claim, not only the underlying cell. Check axis origin and scale, intervals, category ordering, area and icon proportionality, labels, color meaning, sample size and uncertainty. A truncated axis can magnify a small difference; a three-dimensional shape can make area look like quantity; a cumulative line can be mistaken for a period rate. The caption should state measure, population, period and source succinctly. If a chart compares unlike methods or periods, explain the break rather than smoothing it away.

Provide citations that let the evaluator find the evidence. Name publisher or owner, title or record, publication or extraction date, exact table, page or section, and a durable link or controlled attachment where permitted. Confirm that the evaluator can access the source and that the cited location still contains the value. Do not cite a home page for a specific statistic. For confidential evidence, follow the disclosure rules and give the buyer the authorized route instead of exposing the source in a public appendix.

Edition 3.0 of the UK Code of Practice for Statistics emphasizes suitable sources and methods, assured production, openness about strengths, limits and uncertainty, and clear presentation. Proposal teams are not producers of official statistics merely because they cite the Code. The principles are useful as a release challenge: is the figure relevant, based on an appropriate source, checked, explained honestly and presented so it cannot readily be misread? Avoid cherry-picking one attractive period or subgroup while concealing a materially different full result.

Freeze statistics by ID, then verify every occurrence in the final rendered package. Search again after executive edits, translation, pagination and chart export. Check accessibility text and linked spreadsheet cells. Record approval date, approved wording, permitted buyers or channels and expiry or next-review trigger. If a late amendment changes scope, volume or dates, reopen affected statistics. The end state is not “all numbers checked” as a status label. It is a set of reproducible claims whose meaning survives every place the evaluator encounters them.

• Check chart design for exaggerated or obscured comparisons.
• Use a citation that resolves to the exact supporting location.
• Communicate material uncertainty and limitations beside the claim.
• Reconcile every repeated value after final edits and translation.
• Reopen approval when scope, source, permission or currentness changes.