Citation Accuracy Audit Prompt

This prompt is a post-generation auditor designed for teams shipping AI outputs that include citations. Its job is to verify three properties per citation: that the source actually exists, that any quoted material matches the source text, and that the source genuinely supports the claim it's attached to. Use it when your system produces cited summaries, research briefs, legal analyses, or documentation where a fabricated or misattributed citation carries reputational, compliance, or operational risk. The ideal user is an AI engineer, evaluation lead, or content verification specialist embedding this prompt into a validation layer before outputs reach end users, reviewers, or downstream systems.

This is an evaluation prompt, not a generation prompt. It belongs in a post-generation pipeline stage—after the model produces a cited output but before that output is published, stored, or acted upon. Wire it into a CI/CD gate for documentation pipelines, a human review queue where each citation receives a pass/fail verdict with evidence, or an automated monitoring system that samples production outputs for citation quality. The prompt expects a generated text and its cited sources as input. It returns a structured audit report with per-citation verdicts, evidence for each verdict, and an overall pass/fail summary. Do not use this prompt to generate cited content; it verifies existing citations and will hallucinate if asked to produce them.

Avoid this prompt when citations are decorative rather than evidentiary, when sources are not available for verification, or when the cost of verification exceeds the risk of error. It is not a substitute for human review in high-stakes domains like legal filings or medical documentation—use it to triage and prioritize human attention, not to replace it. For RAG systems where citations are generated alongside answers, pair this prompt with a retrieval-augmented generation pipeline that provides the source text for each citation. If your system cannot supply the full source content, this prompt cannot verify quote accuracy or claim support and should be scoped down to source existence checks only. Start by running this prompt on a sample of production outputs to calibrate its verdicts against human judgment before deploying it as an automated gate.

This prompt template performs a structured audit of every citation in a provided text. It verifies that each referenced source exists, is quoted accurately, and genuinely supports the associated claim. The output is a pass/fail report per citation, suitable for integration into documentation review pipelines, research QA workflows, and RAG system evaluation harnesses. Use this prompt when you need to validate that cited evidence actually backs the claims being made, not just that a source was mentioned.

textCopy
You are a citation accuracy auditor. Your task is to verify every citation in the provided text against the provided source documents. For each citation, you must determine whether the source exists, whether any quoted material matches the source exactly, and whether the source genuinely supports the claim it is attached to.

## INPUT

### Text to Audit
[TEXT_TO_AUDIT]

### Source Documents
[SOURCE_DOCUMENTS]

## OUTPUT SCHEMA

Return a JSON object with the following structure:
{
  "audit_summary": {
    "total_citations": <integer>,
    "passed": <integer>,
    "failed": <integer>,
    "unverifiable": <integer>
  },
  "citation_results": [
    {
      "citation_id": "<string identifier for the citation>",
      "citation_text": "<the citation as it appears in the text>",
      "claimed_source": "<the source being cited>",
      "associated_claim": "<the claim this citation is attached to>",
      "source_exists": <true | false | "cannot_verify">,
      "quote_accuracy": "<exact_match | minor_deviation | major_deviation | fabricated | no_quote_present>",
      "claim_support": "<fully_supported | partially_supported | unsupported | contradicted | cannot_assess>",
      "overall_result": "<pass | fail | unverifiable>",
      "failure_reason": "<explanation if failed or unverifiable, otherwise null>",
      "evidence_excerpt": "<relevant excerpt from source document, or null if not found>"
    }
  ]
}

## CONSTRAINTS

- Do not mark a citation as passed unless you can locate the source document and confirm the claim is supported.
- If a source document is not provided or cannot be located, mark the citation as unverifiable, not failed.
- For quote accuracy, compare the quoted text character-by-character against the source. Minor punctuation differences count as minor deviation. Missing or added words count as major deviation.
- If a citation makes multiple claims, assess whether the source supports all of them. Partial support is a fail.
- Do not infer support. If the source does not explicitly state what the claim asserts, mark it as unsupported.
- Flag any citation that references a source not present in the provided source documents.

## RISK LEVEL
[HIGH] This audit may be used in regulated, legal, or safety-critical contexts. All failures must be escalated for human review. Do not make assumptions about source content beyond what is explicitly provided.

To adapt this template, replace [TEXT_TO_AUDIT] with the document or response containing citations, and [SOURCE_DOCUMENTS] with the full text of every referenced source. If sources are large, pre-chunk them and include only the relevant passages. For production use, wrap this prompt in a validation layer that checks the output JSON schema before accepting results. Always route failed and unverifiable citations to a human review queue, especially when the [RISK_LEVEL] is high. Pair this prompt with the Source Attribution Verification Prompt for a defense-in-depth approach to citation integrity.

The Citation Accuracy Audit Prompt is designed to operate as a structured evaluation step within a larger content verification or documentation pipeline. It is not a real-time, user-facing chat prompt. Instead, it should be called programmatically after a document has been drafted and its citations extracted. The prompt expects a clean mapping of claims to their cited sources, and it returns a structured audit report. The primary integration points are: a pre-processing step that extracts citation pairs from the target document, the model call itself, a post-processing validation layer, and a conditional human review queue for any citations that fail the audit.

To wire this into an application, begin by extracting the input payload. You need a list of objects, each containing a claim, the cited_source_text, and a source_identifier. This extraction can be done with a separate, simpler prompt or a deterministic parser if your citations follow a strict format. Once the payload is assembled, call the Citation Accuracy Audit Prompt using a model with strong instruction-following and structured output capabilities, such as gpt-4o or claude-3.5-sonnet. Configure the API call to request a JSON response matching the defined OUTPUT_SCHEMA. Implement a retry loop with exponential backoff for malformed JSON responses. If the model fails to return valid JSON after three attempts, log the raw output and escalate the entire audit batch for manual review. For high-stakes documentation, set temperature=0 to maximize determinism.

Post-processing is critical. Validate the model's output against the expected schema: ensure every citation from the input has a corresponding entry in the audit report, that pass/fail values are boolean or strictly controlled enums, and that audit_notes are present for every failure. Any citation marked as fail or unverifiable should be routed to a human review queue. This queue should present the original claim, the cited source text, and the model's audit notes side-by-side, allowing a human reviewer to confirm the finding or override it. Log every audit result, including the model version, prompt version, and input hash, to build an evidence trail for governance and to enable regression testing when the prompt is updated. Do not automatically rewrite or correct the source document based on the model's output; the audit report is a diagnostic tool, not an auto-fix agent.