Domain-Specific Quote Verification Prompt Template

Use this prompt when you are building or operating a quote-verification product that must handle specialized terminology, evidence standards, and failure modes unique to a single domain—such as legal contracts, medical literature, financial disclosures, scientific papers, or news reporting. The template is designed for verification product teams who already have a working general-purpose quote-checking flow but find that domain-agnostic prompts produce false positives on technical language, miss context-stripping patterns specific to the domain, or fail to calibrate uncertainty in ways that satisfy regulatory or editorial reviewers. The core job-to-be-done is adapting the verification logic so that a quoted statement is evaluated against the evidence rules that domain experts actually use, not against a generic notion of 'accurate wording.'

The ideal user is an AI engineer or technical product manager who controls the prompt layer of a verification pipeline and needs a modular starting point that accepts domain-specific [EVIDENCE_STANDARDS], [TERMINOLOGY_GLOSSARY], [UNCERTAINTY_RULES], and [FAILURE_MODE_EXAMPLES] as explicit inputs. You should have access to original source material and the quoted or paraphrased text under review. This prompt is not a replacement for retrieval—it assumes evidence has already been provided in [SOURCE_MATERIAL] and [QUOTED_TEXT]. It is also not a replacement for human editorial judgment in high-stakes domains. When the cost of a false negative includes patient harm, legal liability, or market-moving misinformation, the prompt output must route to human review, not directly to publication.

Do not use this template when your verification task spans multiple domains simultaneously, when you lack a clear domain-specific evidence standard to encode, or when the quoted material is so short or ambiguous that even a domain expert would disagree on its fidelity. In those cases, a general-purpose quote-verification prompt with a human-in-the-loop fallback is safer than forcing domain adaptation onto insufficient signal. After reading this section, decide which domain you are adapting for, gather your domain-specific inputs, and proceed to the prompt template section to begin customization.

This template is the core engine for a domain-specific quote verification system. It is designed to be adapted by verification product teams working in legal, medical, financial, scientific, or news domains. Unlike a general-purpose quote checker, this prompt forces the model to apply the domain's specific evidence standards, handle specialized terminology, and calibrate its uncertainty based on the risk level you define. The square-bracket placeholders must be populated by your application harness before each request; do not use this template as a static system prompt without dynamic variable injection.

codeCopy
You are an expert quote verification analyst specializing in [DOMAIN] content. Your task is to compare an attributed quote against provided source material and determine its accuracy, context, and fidelity according to [DOMAIN]-specific standards.

## INPUT
- **Attributed Quote:** [QUOTE_TEXT]
- **Attributed Speaker/Entity:** [SPEAKER_NAME]
- **Source Material:** [SOURCE_TEXT]
- **Source Metadata:** [SOURCE_METADATA]

## DOMAIN-SPECIFIC EVIDENCE STANDARDS
[DOMAIN_STANDARDS]

## TERMINOLOGY HANDLING
[TERMINOLOGY_RULES]

## OUTPUT SCHEMA
You must respond with a single JSON object conforming to this schema:
{
  "verification_status": "EXACT_MATCH" | "SUBSTANTIAL_MATCH" | "PARAPHRASE_MATCH" | "MISQUOTATION" | "FABRICATED" | "INSUFFICIENT_EVIDENCE",
  "fidelity_score": 0.0-1.0,
  "discrepancies": [
    {
      "type": "WORD_SUBSTITUTION" | "OMISSION" | "ADDITION" | "CONTEXT_STRIPPING" | "MEANING_DISTORTION" | "TERMINOLOGY_MISUSE",
      "quoted_text": "string",
      "source_text": "string",
      "severity": "CRITICAL" | "MODERATE" | "MINOR",
      "explanation": "string"
    }
  ],
  "evidence_excerpts": ["string"],
  "confidence": 0.0-1.0,
  "confidence_rationale": "string",
  "domain_specific_notes": "string",
  "requires_human_review": true/false
}

## CONSTRAINTS
[CONSTRAINTS]

## UNCERTAINTY CALIBRATION
- If the source material is incomplete, fragmented, or ambiguous, set verification_status to INSUFFICIENT_EVIDENCE and confidence below 0.5.
- For [DOMAIN] content, treat terminology mismatches as potentially CRITICAL if they alter [DOMAIN_SPECIFIC_MEANING_FACTOR].
- If the quote contains [DOMAIN]-specific numerical claims, verify them against source data with a tolerance of [NUMERICAL_TOLERANCE].

## FEW-SHOT EXAMPLES
[EXAMPLES]

## RISK LEVEL
Current risk level: [RISK_LEVEL]
- LOW: Auto-approve SUBSTANTIAL_MATCH and above with confidence >= 0.8.
- MEDIUM: Require human review for anything below EXACT_MATCH.
- HIGH: Require human review for all outputs; this prompt is for triage only.

Now, verify the attributed quote against the source material and return only the JSON object.

To adapt this template, start by defining your domain's evidence standards in the [DOMAIN_STANDARDS] block. For legal domains, this might include rules about verbatim transcription from court records and handling of procedural terminology. For medical domains, it should specify tolerance for generic drug name substitutions and the criticality of dosage accuracy. The [TERMINOLOGY_RULES] block should define how synonyms, acronyms, and domain jargon are treated—for example, whether 'MI' and 'myocardial infarction' are considered an exact match or a paraphrase. The [CONSTRAINTS] block is where you inject output length limits, forbidden inference rules, and domain-specific red lines such as 'never infer a speaker's intent from tone alone.' Always populate [EXAMPLES] with at least two few-shot demonstrations showing both a clean match and a subtle misquotation that is domain-critical. The [RISK_LEVEL] parameter should be set dynamically by your application based on the content's downstream use—a quote in a peer-reviewed medical journal requires HIGH risk, while an internal draft memo might operate at MEDIUM. After copying this template, test it against a golden dataset of known quote-source pairs in your domain before integrating it into any production pipeline.

The Domain-Specific Quote Verification Prompt Template is designed to operate as a modular component within a larger verification pipeline, not as a standalone chat interaction. In production, this prompt typically sits between a claim extraction stage and a human review queue. The application layer is responsible for assembling the prompt with the correct domain configuration, source material, and the quote to verify. Because different domains (legal, medical, financial, scientific, news) have distinct evidence standards, terminology handling, and failure modes, the harness must select the appropriate domain adapter before invoking the model. A common pattern is to maintain a domain configuration registry that maps domain labels to evidence standards, acceptable source types, terminology glossaries, and calibrated uncertainty thresholds. The application fetches the correct configuration, injects it into the [DOMAIN_CONFIG] placeholder, and appends the quote and source evidence before sending the request.

Validation is the first line of defense after receiving the model's response. The prompt template includes a structured output contract—typically JSON with fields for verification status, confidence score, evidence alignment, discrepancy flags, and a rationale. The harness should validate this output against a schema before any downstream consumption. If the model returns malformed JSON, missing required fields, or values outside expected ranges (e.g., a confidence score not between 0 and 1), the harness should trigger a retry with the same prompt plus a repair instruction that includes the validation error message. A maximum of two retries is recommended before escalating to a human review queue. For high-risk domains such as legal and medical, even successfully validated outputs with confidence below a configurable threshold (e.g., 0.85) should be routed to human review with the full verification payload attached. Log every invocation—including the domain, quote, source excerpts, model response, validation result, and retry count—for audit trails and prompt performance analysis.

Model choice and tool integration significantly affect harness design. For domains requiring retrieval of external evidence (e.g., checking a quote against a public figure's recorded statements), the harness should integrate a retrieval step before invoking the verification prompt. Use a tool-calling pattern where the model can request a search or document lookup, or pre-fetch evidence and inject it into the [SOURCE_EVIDENCE] placeholder. When using models with smaller context windows, be deliberate about truncating source material: prioritize the passages most likely to contain the quoted statement, and include surrounding context to enable context-stripping detection. For batch processing scenarios, implement rate limiting, partial-failure recovery, and cost tracking per domain. Avoid deploying this prompt without eval guardrails—maintain a golden dataset of known quote-source pairs with expected verification outcomes, and run regression tests whenever the prompt template, domain configuration, or model version changes.