Financial Metric and Ratio Verification Prompt

This prompt is designed for finance, audit, and compliance teams who need to verify a specific financial metric or ratio claim against provided financial data. The job-to-be-done is not general financial analysis; it is a targeted, auditable verification of a single quantitative assertion—such as a reported gross margin of 42%, an EBITDA of $5.2M, or a debt-to-equity ratio of 1.8. The ideal user is an analyst or auditor who already has the source financial statements, trial balances, or data extracts and needs the model to perform the calculation, check period alignment, and flag any GAAP/IFRS inconsistencies or non-GAAP reconciliation gaps. The prompt assumes the claim is already extracted and the evidence is already gathered; it does not perform retrieval or claim extraction.

Do not use this prompt when the underlying data is missing, when the claim is purely qualitative (e.g., 'revenue grew significantly'), or when the metric definition is ambiguous and cannot be resolved from the provided context. This prompt is also inappropriate for forward-looking statements, pro-forma projections without reconciliation to actuals, or market-based valuations that require external pricing data. It is a calculation verification tool, not a financial analyst. If the metric requires subjective allocation (e.g., shared cost attribution across segments), the prompt will flag the assumption dependency rather than silently accepting one allocation method.

Before using this prompt, ensure you have the exact claim text, the relevant financial data with period labels, the applicable accounting standard (GAAP, IFRS, or other), and any known non-GAAP reconciliation items. The prompt includes a harness for formula correctness, unit normalization, and tolerance-window comparison. For high-stakes audit use, always route outputs to a human reviewer and log the verification result with the evidence snapshot used. The next step after this section is to copy the prompt template, populate the placeholders, and wire it into your verification pipeline with the validation and retry logic described in the implementation harness.

This section provides a copy-ready prompt template that you can drop into your verification pipeline. The template is designed to accept a financial claim, supporting data, and reference standards, then produce a structured verification result with calculation checks, period matching validation, and accounting standard alignment flags. Every placeholder uses square-bracket notation so you can wire it into your application's variable substitution system without ambiguity.

textCopy
You are a financial verification analyst. Your task is to verify a financial metric or ratio claim against provided source data. You must check calculation correctness, period alignment, accounting standard compliance, and reconciliation to GAAP or IFRS as specified.

## INPUT
**Claim:** [CLAIM_TEXT]
**Source Financial Data:** [SOURCE_DATA]
**Reference Period:** [REFERENCE_PERIOD]
**Applicable Standard:** [ACCOUNTING_STANDARD]
**Non-GAAP Reconciliation (if applicable):** [NON_GAAP_RECONCILIATION]

## CONSTRAINTS
- Recalculate the claimed metric from source data before comparing to the stated claim.
- Flag any period mismatch between the claim and the source data.
- Identify non-GAAP adjustments and verify reconciliation to the nearest GAAP equivalent.
- Apply a materiality threshold of [MATERIALITY_THRESHOLD] for numerical differences.
- If source data is insufficient to verify the claim, state what is missing rather than guessing.
- Do not invent financial data, assumptions, or adjustments.

## OUTPUT SCHEMA
Return a JSON object with this exact structure:
{
  "claim_summary": "string",
  "verification_result": "VERIFIED | DISCREPANCY_FOUND | INSUFFICIENT_DATA",
  "recalculated_value": number | null,
  "claimed_value": number | null,
  "absolute_difference": number | null,
  "percentage_difference": number | null,
  "within_materiality_threshold": boolean | null,
  "period_match": "MATCH | MISMATCH | UNKNOWN",
  "period_mismatch_detail": "string | null",
  "gaap_ifrs_alignment": "ALIGNED | NON_GAAP_RECONCILED | NON_GAAP_UNRECONCILED | NOT_APPLICABLE",
  "reconciliation_gap": "string | null",
  "calculation_steps": ["string"],
  "missing_data": ["string"],
  "confidence": "HIGH | MEDIUM | LOW",
  "requires_human_review": boolean,
  "review_reason": "string | null"
}

## EXAMPLES
[EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]

Adaptation guidance: Replace [CLAIM_TEXT] with the exact financial assertion being checked, such as 'Q3 2024 gross margin was 62.3%.' Populate [SOURCE_DATA] with structured financial statements, trial balances, or ledger extracts. Set [MATERIALITY_THRESHOLD] to a percentage like 0.5 or an absolute dollar amount. The [EXAMPLES] placeholder should contain one verified and one discrepant example to anchor model behavior. For high-risk audit workflows, set [RISK_LEVEL] to HIGH and ensure the output's requires_human_review field gates any automated downstream action. If your verification pipeline uses a non-GAAP reconciliation step, populate [NON_GAAP_RECONCILIATION] with the company's reconciliation schedule; otherwise, set it to NONE.

What to do next: Copy this template into your prompt management system and replace each placeholder with values from your verification harness. Before shipping, run at least five test cases through the prompt: two with clean matches, one with a deliberate calculation error, one with a period mismatch, and one with insufficient source data. Confirm that recalculated_value matches your independent calculation and that requires_human_review triggers correctly on discrepancies above your materiality threshold. If the model hallucinates reconciliation details when [NON_GAAP_RECONCILIATION] is NONE, add an explicit constraint: 'If no reconciliation is provided, do not invent one.'

This prompt is designed to operate as a verification microservice within a larger financial review pipeline. It expects a structured input payload containing the claim to be verified, the source financial data (balance sheet, income statement, cash flow statement, or a structured data object), and the applicable accounting framework (GAAP, IFRS, or a custom policy). The application layer should assemble this context before calling the model, ensuring that the period under review in the claim exactly matches the period of the provided financial data to prevent temporal mismatches.

The implementation harness must enforce a strict schema contract. Before the prompt is called, validate that the [INPUT] JSON includes required fields: claim.text, claim.metric_type (e.g., 'gross_margin', 'ebitda', 'current_ratio'), claim.stated_value, claim.fiscal_period, and source_data containing the relevant line items. After the model returns its verification result, a post-processing validator should parse the output against a defined [OUTPUT_SCHEMA] that includes verification_status (enum: 'VERIFIED', 'DISCREPANCY_FOUND', 'UNABLE_TO_VERIFY'), calculated_value, tolerance_delta, gaap_ifrs_adjustments (array of objects), and evidence_citations. If the output fails schema validation, implement a retry loop with the error message fed back into the prompt's [CONSTRAINTS] section, up to a maximum of two retries before escalating to a human reviewer.

For high-risk financial workflows, never allow this prompt's output to be the final word without a human-in-the-loop gate. Log every verification attempt, including the raw input, model output, validation status, and any retry attempts, to an immutable audit trail. When a discrepancy is found, route the result to a review queue with a pre-formatted summary of the claim, the model's calculated value, and the specific line items used as evidence. For non-GAAP metric claims, the harness should flag the output for mandatory human reconciliation, as these often involve company-specific adjustments that a general-purpose model cannot reliably adjudicate. Model choice matters: prefer models with strong numerical reasoning and structured output capabilities, and set the temperature to 0 to maximize deterministic, repeatable calculations.