Quote-to-Source Comparison Prompt Template

This prompt is for editorial and verification teams who need to compare an attributed quote against its original source material and produce a structured alignment report. It is designed for workflows where quote integrity is a high-visibility risk—newsrooms, publishing platforms, compliance review, and content moderation. The prompt expects you to provide both the attributed quote and the full original source passage. It returns a fidelity score, a list of discrepancies, and excerpted evidence so a human reviewer can make a final call.

This is not a prompt for verifying whether a public figure actually said something (use the Attributed Statement Verification Prompt for that). It is also not for checking paraphrases (use the Paraphrase Fidelity Scoring Prompt). This prompt assumes the quote and source are already in hand and focuses on the comparison itself. Do not use it when the source material is unavailable, when the quote is a non-verbatim summary, or when you need to search across multiple sources to find the original—those are separate verification sub-tasks with their own prompts in this pillar.

Before using this prompt, confirm that you have the complete, unredacted source passage and the exact quote as it appears in the draft content. The prompt's value depends on having both sides of the comparison. If the source is a transcript with disfluencies or corrections, consider the Quote-to-Transcript Alignment Prompt instead. For batch processing across many quotes, pair this prompt with the Quote Verification Batch Processing Prompt to manage throughput, cost, and partial-failure recovery.

This prompt template is the core instruction set for comparing an attributed quote against its original source material. It instructs the model to act as a meticulous editorial auditor, producing a structured JSON report that captures alignment, discrepancies, and evidence. The output is designed to be machine-readable for downstream routing, logging, or human review queues. Before using this prompt, ensure you have the exact quote text and the full original source passage available as inputs.

textCopy
You are an editorial verification auditor. Your task is to compare an attributed quote against its original source material and produce a structured alignment report.

## INPUT
[QUOTE]: The attributed quote to verify.
[SOURCE]: The full original source passage from which the quote was allegedly drawn.
[CONTEXT]: Optional context about the speaker, publication, or date.

## OUTPUT SCHEMA
Return a single JSON object with the following structure:
{
  "verification_id": "string",
  "quote": "string (the input quote)",
  "source_excerpt": "string (the relevant excerpt from the source)",
  "fidelity_score": number (0.0 to 1.0, where 1.0 is a perfect verbatim match),
  "match_type": "verbatim" | "paraphrase" | "partial" | "fabricated" | "unverifiable",
  "discrepancies": [
    {
      "type": "substitution" | "omission" | "insertion" | "reordering" | "context_stripping",
      "quote_fragment": "string",
      "source_fragment": "string",
      "severity": "minor" | "moderate" | "major",
      "explanation": "string"
    }
  ],
  "missing_context": "string or null (context present in the source but absent from the quote that could alter meaning)",
  "evidence": "string (a direct excerpt from the source supporting the assessment)",
  "confidence": number (0.0 to 1.0),
  "recommendation": "publish" | "revise" | "escalate"
}

## CONSTRAINTS
- If the quote cannot be located in the source, set match_type to "unverifiable" and fidelity_score to 0.0.
- Flag any omission of qualifying language (e.g., "but," "however," "in some cases") as a context_stripping discrepancy with a severity of at least "moderate."
- Do not hallucinate source text. Only use evidence explicitly present in [SOURCE].
- If the source is insufficient to make a determination, set confidence below 0.5 and recommend "escalate."
- For high-stakes domains (legal, medical, financial), always recommend "escalate" if any discrepancy is "major."

## RISK LEVEL
[RISK_LEVEL]: "low" | "medium" | "high"

To adapt this template, replace the square-bracket placeholders with your actual data. The [QUOTE] and [SOURCE] fields are required. The [CONTEXT] field is optional but recommended for disambiguating speakers or publications. The [RISK_LEVEL] parameter adjusts the model's conservatism: at "high," the model will recommend escalation for any moderate discrepancy. Integrate this prompt into your application by passing it as the system message, then providing the filled placeholders in the user message. Always validate the output JSON against the schema before acting on the recommendation, and log the raw output for audit trails.

The Quote-to-Source Comparison prompt is designed to operate as a single step within a larger verification pipeline, not as a standalone chat interaction. In production, you will typically call this prompt after claim extraction has identified an attributed quote and its alleged source, and after retrieval has fetched the original source material. The prompt expects a structured input containing the quote text, the source text, and any relevant metadata such as attribution context, publication dates, and speaker identity. The output is a structured alignment report with fidelity scores, discrepancy flags, and excerpted evidence—designed to be consumed by downstream systems rather than displayed raw to end users.

To wire this into an application, wrap the prompt in a function that accepts a QuoteVerificationRequest object with fields: [QUOTE_TEXT], [SOURCE_TEXT], [ATTRIBUTION_CONTEXT], and optional [METADATA] such as speaker name, date, and document title. The function should inject these into the prompt template, call the model with a low temperature (0.0–0.2) to maximize consistency, and parse the response against a strict JSON schema. The schema must validate: fidelity_score (0.0–1.0 float), discrepancy_flags (array of objects with type, severity, excerpt_quote, excerpt_source, explanation), alignment_evidence (array of matched passages), and recommendation (enum: accept, review, reject). Any response that fails schema validation should trigger a retry with the validation error message appended to the prompt as [PREVIOUS_ERROR]. After two retries, escalate to human review with the raw model output and validation failure log attached.

Model choice matters for this workflow. The comparison task requires strong instruction-following and precise excerpt matching, making Claude 3.5 Sonnet or GPT-4o suitable defaults. For cost-sensitive batch processing, consider using a smaller model with few-shot examples of correct alignment reports, but validate outputs more aggressively—smaller models are more prone to hallucinating source passages that do not exist in the provided text. Always include a post-processing check that verifies every excerpt_source string is a substring of [SOURCE_TEXT] (allowing for minor whitespace normalization). This substring check catches the most common failure mode: the model inventing plausible-sounding source text that is not actually present. Log every verification result with the model version, prompt template version, input hashes, and validation outcome for auditability and regression testing.

For high-stakes editorial or legal workflows, route all review and reject recommendations to a human review queue. Package the verification output alongside the original quote, source, and a diff view highlighting discrepancies. The human reviewer should be able to confirm, override, or add notes to each discrepancy flag. Store these human decisions as labeled data for future model evaluation and potential fine-tuning. Avoid using this prompt for real-time publishing gates without human-in-the-loop confirmation—quote integrity failures carry reputational and legal risk that automated systems alone should not own. For batch processing of large document collections, implement rate limiting, cost tracking per batch, and a dead-letter queue for requests that exhaust retries, ensuring no verification result is silently dropped.