Journalism Source Attribution Verification Prompt Template

This prompt is designed for newsroom fact-checking desks and editorial verification teams that need to evaluate whether claims attributed to sources in a news article match the original source material. The core job-to-be-done is detecting quote distortion, context-stripping, paraphrasing that changes meaning, and gaps in anonymous source credibility. Use this prompt when you have both the published article text and the original source transcript, document, or recording available for comparison. The prompt assumes you have already retrieved or uploaded the relevant source evidence—it does not perform retrieval itself.

The ideal workflow places this prompt after evidence gathering and before publication or correction. Feed it the article text as [ARTICLE_TEXT] and the source material as [SOURCE_MATERIAL], along with any [ATTRIBUTION_RULES] your newsroom enforces (e.g., rules for anonymous sources, paraphrasing thresholds, or acceptable context trimming). The prompt outputs a structured attribution fidelity report with per-claim scores, evidence excerpts, and specific flags for distortion, omission, or credibility gaps. This output can feed directly into an editorial review queue or a CMS verification plugin.

Do not use this prompt for general fact-checking of unsourced claims, real-time breaking news verification where source material is unavailable, or evaluating the truth of claims independent of their source attribution. It is not a substitute for verifying whether a source's statements are factually correct—only whether the article faithfully represents what the source said. If source material is missing, the prompt will correctly flag claims as unverifiable rather than hallucinating an assessment. For high-risk stories involving legal exposure or safety-critical information, always route the prompt's output through human editorial review before publication.

This template is designed to be pasted directly into your orchestration layer. It evaluates a single claim-attribution pair by comparing the claim as reported against the provided source material excerpt. The prompt instructs the model to act as a strict editorial auditor, focusing on semantic fidelity, quote accuracy, and context preservation. Before each inference call, replace the square-bracket placeholders with live data from your claim extraction pipeline and source retrieval system. The output is a structured JSON object suitable for downstream scoring, routing, and audit logging.

textCopy
You are an editorial attribution auditor. Your task is to compare a reported claim against its original source material and determine whether the attribution is faithful.

## INPUT
Reported Claim: [CLAIM_TEXT]
Attributed Source: [SOURCE_NAME]
Source Material Excerpt: [SOURCE_EXCERPT]
Original Context (if available): [ORIGINAL_CONTEXT]

## CONSTRAINTS
- Compare the semantic meaning, not just exact wording.
- Flag any quote that is altered, truncated, or recombined in a way that changes meaning.
- Detect context-stripping: when a quote is technically accurate but misleading because surrounding context is omitted.
- For anonymous sources, note the absence of verifiable attribution without judging the claim's truth.
- If the source excerpt does not contain enough information to verify the claim, mark it as "insufficient evidence" rather than guessing.

## OUTPUT_SCHEMA
Return a single JSON object with the following fields:
{
  "attribution_fidelity": "faithful" | "distorted" | "context_stripped" | "unverifiable" | "misattributed",
  "confidence": 0.0-1.0,
  "quote_accuracy": "exact" | "minor_variance" | "paraphrased" | "altered" | "fabricated" | "not_applicable",
  "context_preserved": true | false | "partial",
  "evidence_excerpts": ["string of relevant source text"],
  "explanation": "Concise explanation of the fidelity determination with specific reference to the source material.",
  "distortion_type": "quote_alteration" | "context_omission" | "meaning_inversion" | "false_attribution" | "anonymous_unverifiable" | "none" | null,
  "requires_human_review": true | false
}

## EXAMPLES
Example 1:
Reported Claim: "The CEO said profits were 'the highest in company history.'"
Source Excerpt: "Our profits this quarter were strong, though not the highest in company history. We remain optimistic about future growth."
Output: {"attribution_fidelity": "distorted", "confidence": 0.95, "quote_accuracy": "altered", "context_preserved": false, "evidence_excerpts": ["though not the highest in company history"], "explanation": "The source explicitly negates the claim. The quote is fabricated by omitting the negation.", "distortion_type": "meaning_inversion", "requires_human_review": true}

Example 2:
Reported Claim: "An anonymous official confirmed the deal is imminent."
Source Excerpt: [No source excerpt provided for anonymous source]
Output: {"attribution_fidelity": "unverifiable", "confidence": 0.9, "quote_accuracy": "not_applicable", "context_preserved": false, "evidence_excerpts": [], "explanation": "Claim relies on an anonymous source with no verifiable material provided. Cannot assess fidelity.", "distortion_type": "anonymous_unverifiable", "requires_human_review": true}

## RISK_LEVEL
[RISK_LEVEL]

To adapt this template for your newsroom or verification pipeline, start by mapping your internal claim and source data structures to the [CLAIM_TEXT], [SOURCE_NAME], and [SOURCE_EXCERPT] placeholders. The [ORIGINAL_CONTEXT] field is optional but strongly recommended for detecting context-stripping; populate it with the paragraph or section surrounding the quoted material. The [RISK_LEVEL] placeholder should be set to "high" for public-facing editorial content, which will cause the model to lower its threshold for flagging issues and recommending human review. For lower-stakes internal drafts, set it to "standard". Always validate the output JSON against the schema before accepting the result, and route any response with requires_human_review: true or confidence below your operational threshold (typically 0.85) to an editor. Never use this prompt to make automated publication decisions without human oversight on flagged items.

This prompt is designed to be a single step inside a larger verification pipeline, not a standalone chat interaction. The application layer should handle source retrieval, text extraction, and evidence segmentation before the prompt is called. The model receives a single claim, the original article context, and the source material excerpt, and returns a structured attribution fidelity assessment. The application is responsible for iterating over claims, calling this prompt for each one, and aggregating results into a verification report. Do not send an entire article and all sources in one call—batch at the claim level to keep context focused and reduce hallucination risk.

Input assembly: Before calling the prompt, extract the exact claim text and the surrounding paragraph from the article. Retrieve the cited source material—ideally the specific passage, timestamp, or document section the article references. If the source is a person, include the full quote or transcript segment. If the source is anonymous, include whatever attribution context exists (e.g., 'senior administration official'). The [CLAIM] placeholder receives the verbatim claim. [ARTICLE_CONTEXT] receives the paragraph containing the claim plus one paragraph before and after. [SOURCE_MATERIAL] receives the original source excerpt. [SOURCE_TYPE] is one of: named_document, named_person, anonymous_person, press_release, transcript, dataset, or other. [ATTRIBUTION_TEXT] is the exact attribution phrase from the article (e.g., 'according to the report,' 'said in an interview').

Validation and retries: Parse the model's JSON output and validate that attribution_fidelity_score is a float between 0.0 and 1.0, distortion_type is one of the allowed enum values, and evidence_excerpts is a non-empty array with character offsets that fall within the provided source material length. If validation fails, retry once with the same prompt plus the validation error message appended as a [PREVIOUS_ERROR] field. If the second attempt also fails, flag the claim for human review and log the raw output. For attribution_fidelity_score below 0.7, route to a human reviewer with the full prompt context and model output pre-filled. For scores between 0.7 and 0.9, surface in a review queue but allow the pipeline to continue. Scores above 0.9 can auto-pass unless the claim involves an anonymous source—anonymous source claims should always route to human review regardless of score.

Model choice and latency: Use a model with strong instruction-following and JSON output reliability, such as Claude 3.5 Sonnet or GPT-4o, with response_format set to json_object and the JSON schema provided in the system instructions. Expect 2–5 seconds per claim. For high-throughput newsroom pipelines processing hundreds of claims per article, consider batching claims into groups of 5–10 per call with an array output schema, but only if you validate that batch quality doesn't degrade versus single-claim calls. Log every call with the claim ID, model version, prompt template version, latency, and validation outcome for audit trails.

What to avoid: Do not use this prompt for real-time breaking news verification without human-in-the-loop review. Do not treat the model's output as a final editorial decision—it is a triage and evidence-surfacing tool. Do not skip source retrieval and expect the model to verify claims from its training data; the prompt requires explicit source material in the [SOURCE_MATERIAL] field. If source material is unavailable, route the claim to a 'source needed' queue rather than calling the prompt with an empty field. Finally, never expose raw model outputs to readers without editorial review—the output is an internal verification artifact, not publishable content.