Quote Fabrication Risk Assessment Prompt

This prompt is a triage instrument for editorial and investigative teams who need to evaluate the likelihood that an attributed quote is fabricated. It is not a final verdict. The ideal user is a researcher, journalist, or fact-checker who has encountered a quote in a draft article, social media post, or third-party report and needs a structured, repeatable way to assess its authenticity before investing hours in manual investigation. The prompt requires three pieces of context to function: the quote itself, information about the attributed speaker (such as their known speech patterns, public role, and typical venues), and any available provenance data (such as the publication where the quote appeared, a claimed date, or an archival reference). Without these inputs, the model cannot produce a calibrated risk score.

The prompt analyzes the quote against the provided context to produce a risk score and supporting indicators. These indicators include stylistic mismatch (does the quote sound like the attributed speaker?), absence from known archives (has the quote appeared in any verifiable record?), and provenance gaps (is the chain of custody for the quote broken or suspicious?). The output is designed to be consumed by a human decision-maker or routed into a downstream verification pipeline. High-risk items should always route to a human investigator. This prompt does not perform source-to-text comparison, which is handled by sibling prompts in the Quote Accuracy and Paraphrase Fidelity Verification group, such as the Quote-to-Source Comparison Prompt Template. It also does not search the web or query external databases; it reasons only over the context you provide.

Do not use this prompt when you need a definitive yes/no answer on fabrication. It is a risk assessment tool, not a lie detector. Do not use it when the attributed speaker is unknown or the provenance information is entirely absent—the model will have no signal to work with and may produce a misleadingly confident score. Do not use it as a replacement for source-to-text comparison when the original source is available; use the Quote-to-Source Comparison Prompt instead. For production systems, always pair this prompt with a human review step for any output above a 'medium' risk threshold, and log every assessment with the input context, model version, and risk score for auditability.

The prompt below is designed to be pasted directly into your system prompt or user message field. It instructs the model to act as a fabrication-risk triage analyst, evaluating a single attributed quote against a set of indicators and producing a structured risk score. The output is intended to route high-risk items to human investigators, not to make a final determination of fabrication. Before using this prompt, ensure you have the quote text, the claimed speaker, the publication or context where it appeared, and any available provenance metadata.

textCopy
You are a quote-fabrication risk analyst. Your job is to triage attributed quotes by estimating the likelihood that they were fabricated, partially fabricated, or misattributed. You do not make final determinations. You produce a structured risk assessment that routes high-risk items to human investigators.

## INPUT
- Quote: [QUOTE_TEXT]
- Claimed Speaker: [SPEAKER_NAME]
- Source Context: [SOURCE_DESCRIPTION]
- Publication Date: [PUBLICATION_DATE]
- Available Provenance: [PROVENANCE_NOTES]

## OUTPUT SCHEMA
Return a single JSON object with the following fields:
- "fabrication_risk_score": number between 0.0 and 1.0, where 0.0 is almost certainly authentic and 1.0 is almost certainly fabricated.
- "risk_category": one of "LOW", "MEDIUM", "HIGH", "CRITICAL".
- "primary_indicators": array of strings naming the strongest indicators that raised the risk score.
- "indicator_details": array of objects, each with "indicator" (string), "finding" (string), and "weight" (number 0.0-1.0).
- "recommended_action": one of "CLEAR", "REVIEW", "ESCALATE", "QUARANTINE".
- "investigator_brief": string summarizing what a human investigator should examine first.
- "abstention": boolean. Set to true if you cannot produce a reliable assessment (e.g., insufficient information).
- "abstention_reason": string explaining why you abstained, if applicable.

## INDICATORS TO EVALUATE
1. **Provenance Absence**: No record of the quote in archives, transcripts, or prior publications.
2. **Stylistic Mismatch**: Vocabulary, syntax, or phrasing inconsistent with the claimed speaker's known style.
3. **Temporal Implausibility**: Quote references events, terminology, or knowledge unavailable at the claimed time.
4. **Source Chain Break**: The quote appears only in a single unverifiable source with no upstream attribution.
5. **Convenience Bias**: The quote aligns too perfectly with the quoter's argument while contradicting the speaker's known positions.
6. **Anomalous Specificity**: Unusual precision or detail that is absent from the speaker's verified statements.
7. **Duplicate or Recycled Content**: Quote matches or closely mirrors a known quote from a different speaker or context.

## CONSTRAINTS
- Do not assume fabrication from a single weak indicator. Weigh multiple indicators together.
- If provenance information is missing, note it as an indicator rather than abstaining immediately.
- Distinguish between "likely fabricated" and "unverifiable." Use the abstention field when evidence is too thin to score.
- Do not output markdown. Return only the JSON object.

To adapt this prompt, replace each square-bracket placeholder with real values before sending. The [QUOTE_TEXT] should be the exact text as published. [SPEAKER_NAME] is the person to whom the quote is attributed. [SOURCE_DESCRIPTION] should include the publication name, article title, and any surrounding context. [PUBLICATION_DATE] helps the model assess temporal plausibility. [PROVENANCE_NOTES] should capture what you already know about the quote's origin—such as whether it appears in transcripts, archives, or prior reporting—or state 'No provenance information available.' The indicator list can be extended or pruned based on your domain. For high-stakes domains such as legal or political fact-checking, add domain-specific indicators such as 'Legal Terminology Misuse' or 'Policy Position Contradiction.' The output schema is designed for direct ingestion by a triage router, so keep field names stable across adaptations.

The Quote Fabrication Risk Assessment Prompt is designed as a triage classifier, not a final arbiter. Its job is to score the likelihood that an attributed quote was fabricated and route high-risk items to human investigators. In a production harness, this prompt sits between quote extraction and human review. It should never auto-publish a fabrication verdict. The harness must enforce that outputs above a configurable risk threshold are always escalated, and even low-risk scores should be logged for audit and trend analysis.

Wire the prompt into an application with a strict JSON output contract and a post-processing validation layer. The model must return a structured object containing at minimum: fabrication_risk_score (0.0–1.0), risk_indicators (a list of named indicators with confidence and evidence excerpts), provenance_status (e.g., verified_source, unverifiable, no_record), and a recommendation field constrained to auto_pass, review, or escalate. Use a JSON schema validator in your application layer—not just the prompt—to reject malformed responses and trigger a retry with the error message appended to the next request. For high-stakes editorial workflows, add a human-in-the-loop gate: any output with fabrication_risk_score >= 0.6 or recommendation == 'escalate' must be queued for review with the full prompt input, model output, and retrieved evidence packaged into a review packet.

The prompt works best when it has access to external evidence. Integrate it with a search or archive retrieval tool (e.g., a news database API, web search, or internal transcript store) before calling the model. The harness should: (1) extract the quote and attribution from the source content, (2) query external sources for the quote string and speaker-context pairs, (3) inject retrieved snippets into the [EVIDENCE] placeholder, and (4) call the model. If no evidence is found, the prompt's provenance_status will naturally flag no_record, which is a strong fabrication indicator. Log every step—extraction, retrieval queries, retrieved documents, model input, model output, and final routing decision—for downstream audit and prompt regression testing. Avoid calling the model without evidence; the prompt's value degrades significantly when it can only reason about stylistic mismatch without source-grounding.

For model choice, prefer models with strong reasoning and structured output capabilities. The prompt requires multi-factor analysis (provenance, style, temporal consistency, archive presence) and calibrated uncertainty, which smaller or older models may handle inconsistently. If using a model that doesn't support strict JSON mode natively, implement a retry loop with format-correction prompts from the Output Repair and Validation pillar. Set a maximum of 2 retries before logging the failure and routing to human review with a model_failure flag. For batch processing across many quotes, implement rate limiting, cost tracking per quote, and a dead-letter queue for items that exhaust retries.

Failure modes to monitor in production: (1) the model overweights stylistic mismatch for non-native speakers or translated quotes, producing false-positive fabrication flags; (2) the model underweights fabrication risk when a quote appears in one low-authority source that was indexed but not verified; (3) provenance checks fail silently when the retrieval system returns empty results due to query formulation errors rather than genuine absence. Mitigate these by running a weekly calibration set of known-fabricated and known-genuine quotes through the harness and tracking precision/recall at your chosen threshold. If the fabrication risk score distribution drifts (e.g., sudden spike in high-risk flags), investigate retrieval pipeline health before adjusting the prompt.

Next steps after implementation: build a review dashboard that shows investigators the quote, the model's risk score and indicators, the retrieved evidence, and a quick accept/override interface. Log every human decision to create a feedback dataset for future fine-tuning or prompt improvement. Do not use the prompt's output as a training label without human verification—the prompt is a triage tool, not ground truth.