Budget Variance Analysis Verification Checklist

Lock the review packet

Create a read-only review packet before testing. Include the approved budget version, actuals extract, mapping table, calculation workbook or query, variance narrative, forecast update, approval record, and a manifest containing file names, dates, owners, and hashes or immutable export identifiers. Record the reporting period, entity scope, currency, accounting basis, sign convention, and materiality threshold on a cover sheet.

The GAO Cost Estimating and Assessment Guide describes reliable cost work as defining purpose and scope, documenting a technical baseline and assumptions, collecting data, checking uncertainty, presenting results, and updating estimates with actual costs. Its federal programme context does not make it an accounting standard for every organization, but it supplies useful verification prompts. The 2025 OMB Circular A-11 separately addresses budget preparation, execution, reporting, and performance. Use that separation to test whether the packet labels each number correctly rather than blending an approved budget, an execution actual, and a later forecast.

Fail this gate if a reviewer cannot identify the exact baseline and actuals used without asking the author. A link to a live folder is not enough because its contents may change after review.

Sources: [1] [2]

Reperform the calculation controls

Run each check independently and record pass, fail, not applicable, reviewer, evidence reference, and timestamp:

Choose samples by risk, not convenience. Always include the largest favorable item, largest unfavorable item, every manual adjustment above materiality, every unusual sign, and at least one item from each driver type. Add a random sample only after these targeted items. AI may identify anomalies or compare formulas, but the reviewer must preserve the tested source rows and independently reproduce the result.

• Reconcile the total actuals in the analysis to the frozen source extract.
• Reconcile the total budget to the approved baseline and confirm no later forecast was substituted.
• Recalculate amount and percentage variances using the stated sign convention.
• Test roll-ups from transaction or account detail to every reported subtotal.
• Scan for omitted rows, duplicate rows, hidden filters, hard-coded totals, broken joins, and overwritten formulas.
• Reperform foreign-exchange, allocation, volume, rate, and timing calculations from source inputs.
• Confirm that one driver is not counted in two bridge categories.
• Tie all bridge components exactly to the headline variance or expose the remainder as unresolved.
• Recalculate any forecast consequence separately from the current-period variance.

Challenge explanations and claims

For every material variance, test a three-part chain: amount, cause, consequence. The amount needs a reproducible calculation. The cause needs a source such as an invoice, signed order, usage record, approved scope change, staffing record, or journal support. The consequence needs a bounded statement about timing, recurrence, or forecast impact.

Mark an explanation unsupported if it merely restates the number—for example, “software costs were higher than budget”—or if it assigns intent without evidence. Separate observed facts from analyst interpretation. A late annual invoice may support a timing classification; it does not by itself prove that the full-year forecast is unchanged. When two records conflict, capture both and state what additional evidence would resolve the difference.

Review the narrative for asymmetric skepticism. Favorable variances require the same evidence as unfavorable ones. Confirm that labels such as efficiency, savings, overspend, or underperformance follow from the defined comparison and do not overstate causation. Keep missing invoices, late postings, and incomplete populations in a residual-uncertainty section.

Worked review record

Assume an analysis reports a $70,000 unfavorable quarterly variance: $42,000 software timing, $18,000 service volume, and $10,000 contract rate. The reviewer freezes the workbook and source exports, then records the following evidence:

The checklist disposition is “return for correction,” not “mostly correct.” The exception log identifies the affected bridge line, source reference, $7,000 exposure, owner, due date, and required retest. After the author separates timing from scope and updates the forecast, the reviewer reruns the affected controls and signs the new artifact hash. No unrelated passed check needs to be repeated unless the correction changed a shared formula or source population.

• The budget ties to $540,000 and actuals tie to $610,000.
• The three bridge lines total $70,000 under the stated unfavorable-positive convention.
• The $42,000 invoice exists, but the contract service period shows that only $35,000 is a timing difference; $7,000 relates to an added module.
• Ticket counts multiplied by the contracted rate reproduce the $18,000 volume effect.
• The signed renewal reproduces the $10,000 rate effect.
• The forecast workbook removes all $42,000 as timing, so it understates the added-module cost by $7,000.

Approve, qualify, or stop release

Use three outcomes. Approve only when all critical controls pass and remaining noncritical observations do not change the decision. Qualify when a visible limitation is acceptable to the accountable owner and the artifact narrows its conclusion. Stop release when totals do not reconcile, a material cause lacks evidence, formulas cannot be reproduced, approval is missing, or residual uncertainty could reverse the proposed decision.

The final evidence bundle should contain the packet manifest, completed checklist, sampled source references, recomputation evidence, exception log, correction history, residual-uncertainty statement, reviewer identity, accountable approval, and release hash. It should also state when the analysis must be refreshed because new actuals, baseline changes, late invoices, or scope decisions can invalidate the result.

Two Prune can use a realistic budget-review task to observe how a participant tests evidence, uses AI support, records exceptions, and communicates a disposition. That assessment evidence is limited to observed performance in the simulation. It is not an accounting opinion, assurance engagement, control attestation, or employment decision.