Fabricated Financial Figure Flagging Prompt for Audit Reports

This prompt is built for a single high-stakes job: detecting monetary amounts, financial ratios, and accounting figures in a model-generated audit report that cannot be traced back to the source financial statements. The ideal user is an audit platform engineer, a financial compliance developer, or a product team embedding AI into an accounting workflow where a hallucinated number is not a minor error—it is a potential misstatement. The prompt assumes you already have a draft report and the underlying source statements. It does not generate the report; it audits the output for fabrication.

Use this prompt when the cost of a fabricated figure is high and human review is mandatory before any number reaches a client, regulator, or database. It is designed for post-generation validation, not for real-time streaming or low-risk narrative text. The prompt requires two concrete inputs: the model-generated report text and the source financial statements. Without both, the flagging logic has no grounding. It also expects a materiality threshold so the model can distinguish between a rounding discrepancy and a potentially material fabrication. Do not use this prompt for general fact-checking, for detecting hallucinated prose that contains no figures, or for reports where source statements are unavailable or incomplete.

The output is not a final answer—it is a structured flagging report that must be routed to a human reviewer. Each flagged figure includes the claimed value, the source evidence (or lack thereof), a materiality assessment, and a recommended action. The prompt is designed to fail safe: when in doubt, it flags for review rather than silently passing a suspect figure. Before wiring this into production, pair it with a schema validator that confirms the JSON structure, a retry loop for malformed outputs, and a review queue that prevents any flagged figure from proceeding without human sign-off. If your workflow cannot support human-in-the-loop review, this prompt is not the right tool.

This prompt template is designed to be dropped into an audit workflow where a language model has generated a draft report or analysis based on provided financial statements. Its job is to cross-reference every monetary amount, ratio, and accounting figure in the draft against the source documents and flag any value that cannot be traced back to the original data. The template uses square-bracket placeholders so you can swap in your own source content, risk tolerance, and output format requirements without rewriting the core instruction. Use it as the second pass in a generate-then-verify pipeline, or as a standalone audit step before any model-generated financial content reaches a reviewer.

textCopy
You are an audit verification assistant. Your task is to review a draft financial report and identify every monetary amount, financial ratio, percentage, or accounting figure that cannot be traced to the provided source financial statements.

## INPUT

[DRAFT_REPORT]

## SOURCE DOCUMENTS

[SOURCE_FINANCIAL_STATEMENTS]

## MATERIALITY THRESHOLD

Flag any discrepancy or untraceable figure exceeding [MATERIALITY_THRESHOLD] in absolute value. For figures below this threshold, note them but mark them as immaterial.

## INSTRUCTIONS

1. Extract every discrete financial figure from the draft report (monetary amounts, ratios, percentages, growth rates, counts, and accounting metrics).
2. For each figure, attempt to locate the corresponding value in the source financial statements. A figure is traceable only if it appears explicitly or can be calculated directly from source values using standard arithmetic.
3. Classify each figure as one of:
   - TRACEABLE: Found in source documents with matching value.
   - DERIVED: Calculated from source values and the calculation is correct.
   - MISMATCHED: Found in source but with a different value.
   - FABRICATED: No corresponding value found in source documents.
   - ROUNDED_AMBIGUOUS: Approximate match but rounding prevents exact verification.
4. For each MISMATCHED or FABRICATED figure, assess materiality against the threshold.
5. Do not invent source references. If you cannot find a figure, classify it as FABRICATED.

## OUTPUT SCHEMA

Return a JSON object with this structure:
{
  "report_id": "string",
  "total_figures_extracted": number,
  "traceable_count": number,
  "derived_count": number,
  "mismatched_count": number,
  "fabricated_count": number,
  "rounded_ambiguous_count": number,
  "findings": [
    {
      "figure_text": "string (the exact text from the draft)",
      "figure_value": number,
      "figure_type": "monetary_amount | ratio | percentage | growth_rate | count | other",
      "classification": "TRACEABLE | DERIVED | MISMATCHED | FABRICATED | ROUNDED_AMBIGUOUS",
      "source_reference": "string or null (section, line, or statement where found)",
      "expected_value": number or null,
      "discrepancy": number or null,
      "materiality_assessment": "MATERIAL | IMMATERIAL | NOT_APPLICABLE",
      "requires_human_review": boolean,
      "notes": "string"
    }
  ],
  "overall_assessment": "CLEAN | MINOR_ISSUES | REQUIRES_REVIEW | UNRELIABLE",
  "human_review_required": boolean
}

## CONSTRAINTS

- Do not guess source locations. Use null for source_reference when untraceable.
- Flag all FABRICATED figures with requires_human_review: true regardless of materiality.
- For DERIVED figures, include the calculation in the notes field.
- If the draft contains no financial figures, return an empty findings array with total_figures_extracted: 0.
- Preserve the exact figure_text from the draft report for traceability.

To adapt this template, replace the bracketed placeholders with your actual inputs. The [MATERIALITY_THRESHOLD] should be set based on your audit standards—typically a percentage of revenue, net income, or total assets, or a fixed dollar amount. If your source documents are lengthy, consider pre-chunking them and running this verification per section, or use a retrieval step to supply only the relevant source passages. The output schema is designed to feed directly into a review queue: any finding with requires_human_review: true should block automated publication and route to an auditor. Before deploying, test this prompt against a golden set of reports with known fabrications, mismatches, and clean outputs to calibrate your materiality threshold and verify that the model does not hallucinate source references in the source_reference field.

This prompt is designed to operate as a post-generation validation step within an audit report pipeline, not as a standalone chat interface. The typical integration pattern places it after a model generates draft audit commentary or financial analysis, but before that content reaches a human reviewer or a final report. The harness must supply the model's output text and the source financial statements as structured inputs, then parse the flagging results into a downstream review queue. Because the domain is high-risk, the implementation must treat every flagged figure as requiring human verification, not automatic removal.

Wire the prompt into your application as a gated validation service. The service should accept two required inputs: the generated audit text and a structured representation of the source financial statements (JSON with line items, balances, ratios, and period metadata). Call the LLM with the prompt template, passing both inputs. Parse the response into a typed object—expect an array of flagged figures, each with a figure, claimed_value, source_traceability status, materiality_assessment, and recommended_action. Validate the response schema strictly: if the model returns malformed JSON, retry once with a repair prompt from the Output Repair and Validation pillar. If the retry also fails, escalate the entire output block for full human review rather than attempting partial flagging. Log every flagging result with the model version, input hashes, and timestamp for audit trail purposes.

Build a human review queue that surfaces only flagged figures above a configurable materiality threshold. For each flagged item, display the original generated text, the claimed figure, the source financial statement excerpt that should have grounded it, and the model's materiality assessment. The reviewer must explicitly confirm, correct, or dismiss each flag before the audit text can proceed. Do not auto-correct figures based on the model's suggestions—the model can identify potential fabrications but cannot be trusted to compute corrected values. After review, store the reviewer's decisions alongside the original flags for governance and model evaluation. This feedback loop also creates a dataset for fine-tuning or prompt improvement over time.