Multimodal Quote Attribution Verification Prompt

This prompt is built for editorial teams, fact-checkers, and verification engineers who need to confirm that a quote attributed to a specific person in a published article, transcript, or social post actually matches what that person said in the original source material. The core job-to-be-done is not just string matching—it's verifying fidelity of meaning when the original source is audio or video, the quoted form is text, and the attribution chain crosses formats. You should use this prompt when you have a claimed quote, a candidate source recording or transcript, and you need a structured verdict on whether the attribution is accurate, paraphrased, context-stripped, or fabricated.

The ideal user is someone integrating this into a content verification pipeline: a newsroom tools engineer, a platform trust-and-safety developer, or an AI architect building a claims-verification harness. Required context includes the full attributed quote as it appears in the published content, the speaker's name as attributed, and the source evidence—ideally a timestamped transcript segment, an audio clip, or a video segment with speaker identification. The prompt is designed to work even when the evidence format differs from the quote format, but it assumes you have already performed speaker diarization or identity labeling on the source media. Do not use this prompt for general sentiment analysis, for verifying statistical claims, or for checking whether a quote is factually true—only whether it was said by the attributed person in a way that preserves the original meaning.

This prompt is not a replacement for human editorial judgment in high-stakes publishing decisions. It produces a structured fidelity assessment with flags for paraphrasing, context-stripping, and speaker mismatch, but the final decision to publish, correct, or retract a quote should involve human review. The prompt is also not designed for real-time live-stream verification, for quotes translated across languages without back-translation evidence, or for situations where the original source is unavailable and you are relying on second-hand accounts. In those cases, the prompt should abstain and flag the evidence gap rather than guess. Wire this prompt into a pipeline that includes upstream speaker diarization, downstream human review routing for low-confidence matches, and logging for audit trails.

This prompt template is designed to be the core instruction set for a multimodal quote attribution verification system. It accepts a target quote, the attributed speaker, the source context where the quote appears, and the original reference material. The model is instructed to produce a structured verification output, including a fidelity score, speaker identity notes, and flags for context-stripping or paraphrasing. The square-bracket placeholders allow you to inject dynamic content from your application's retrieval or user-input layer before each inference call.

textCopy
You are a meticulous editorial fact-checker specializing in quote attribution and fidelity across text, audio, and video sources.

Your task is to verify the accuracy and attribution of a target quote against provided original source material. The target quote and its context may come from one format (e.g., a written article, an audio transcript, a video description), while the original evidence may be in another. You must account for these format differences.

## INPUT

**Target Quote:** [TARGET_QUOTE]
**Attributed Speaker:** [ATTRIBUTED_SPEAKER]
**Source Context (where the target quote appears):** [SOURCE_CONTEXT]
**Original Reference Material (ground truth):** [ORIGINAL_REFERENCE]

## CONSTRAINTS

- Do not assume the target quote is accurate. Your job is to verify it.
- Distinguish between verbatim matches, defensible paraphrases, and meaning distortions.
- If the original reference is audio or video, consider transcription errors, tone, and non-verbal cues described in the reference.
- If speaker identity is ambiguous in the original reference, flag it rather than guessing.
- If the target quote strips critical context that changes its meaning, you must flag it as a context-stripping violation.

## OUTPUT_SCHEMA

Respond with a single JSON object conforming to this schema:
{
  "quote_fidelity_score": "EXACT_MATCH | FAITHFUL_PARAPHRASE | LOOSE_PARAPHRASE | MINOR_DISTORTION | MAJOR_DISTORTION | FABRICATED",
  "speaker_identity_verification": "CONFIRMED | PROBABLE | AMBIGUOUS | MISATTRIBUTED | UNVERIFIABLE",
  "context_stripping_flag": true or false,
  "context_stripping_explanation": "string, required if flag is true",
  "evidence_citation": "A direct quote or timestamp from the [ORIGINAL_REFERENCE] that supports your assessment",
  "verification_rationale": "A concise explanation of your comparison, noting any format-specific challenges (e.g., OCR errors, transcription ambiguity)",
  "requires_human_review": true or false
}

To adapt this template, replace the placeholders with data from your application. For example, [TARGET_QUOTE] might be a user-submitted string, while [ORIGINAL_REFERENCE] could be a text chunk retrieved via RAG or a transcript from an audio file. The OUTPUT_SCHEMA is designed to be machine-readable, allowing your harness to parse the JSON and route the result. If the requires_human_review field is true or the quote_fidelity_score is MAJOR_DISTORTION or FABRICATED, your implementation should automatically escalate to a human review queue. Before deploying, test this prompt with edge cases like overlapping speech in transcripts, OCR errors in image-based quotes, and paraphrases that change numerical values.

This prompt is designed to be a single step within a larger verification pipeline. It expects a structured input containing the attributed quote, the speaker identity, the source medium, and the reference evidence. The application layer is responsible for assembling this input from upstream extraction systems—such as an ASR pipeline for audio, an OCR module for screenshots, or a direct text parser for written articles. Do not pass raw, unprocessed files to this prompt; the model needs pre-extracted, text-based claims and evidence to perform reliable comparison.

The output is a structured JSON object with quote-fidelity scores, speaker-identity verification notes, and context-stripping flags. Implement a strict post-processing validator that checks for the presence of all required fields, enforces the score range (e.g., 0.0 to 1.0), and confirms that flag fields are boolean. If validation fails, use a retry loop with a maximum of two attempts, appending the specific validation error to the prompt context on retry. For cross-format speaker disambiguation, the harness should pre-process evidence to include a speaker_label field derived from diarization metadata or document bylines, which the prompt uses to match against the claimed speaker.

Given the high-risk nature of misattribution in editorial and compliance contexts, this workflow requires a human review step for any output where the fidelity_score falls below 0.85 or the context_stripping_flag is true. The application should route these low-confidence results to a review queue, presenting the original quote, the evidence snippet, and the model's structured notes side-by-side. Log every verification attempt, including the full prompt, raw model output, validation result, and final routed decision, to create an audit trail. Avoid using this prompt for real-time, low-latency applications; the model's reasoning over nuanced paraphrasing and speaker identity is better suited for asynchronous batch processing where accuracy is prioritized over speed.