Speaker Intent Preservation Evaluation Prompt

This prompt is for verification teams, editorial QA engineers, and AI pipeline builders who need to assess whether an attributed quote or paraphrase preserves the original speaker's communicative intent. Literal accuracy is necessary but insufficient. A quote can be word-perfect yet still distort meaning through selective extraction, tone erasure, or context removal. This prompt produces an intent-alignment score with evidence from surrounding source context, enabling automated triage before human review.

Use this when your verification workflow must distinguish between 'the words are correct' and 'the meaning is preserved.' It is designed to sit downstream from a Quote-to-Source Comparison Prompt, which handles string-level fidelity. Once a quote passes a literal match, this prompt evaluates whether the use of that quote is faithful to the speaker's position. The output is a structured score and rationale that can be used to route borderline cases to human reviewers, auto-approve high-confidence matches, or flag content for revision. The harness should include a threshold for automatic escalation: for example, any intent-alignment score below 0.8 triggers a human review task with the full evidence package.

Do not use this prompt for simple string-matching or binary misquotation detection. That work belongs to the Quote-to-Source Comparison Prompt Template or Misquotation Detection Prompt in this pillar. This prompt is also not a replacement for a full editorial review; it is a triage and acceleration tool. In high-stakes domains like legal or medical reporting, a human must always review the final determination, regardless of the model's confidence score.

This prompt template is the core instruction set for the Speaker Intent Preservation Evaluation workflow. It is designed to be copied directly into your system prompt or a dedicated user message in your verification pipeline. The template uses square-bracket placeholders for all dynamic inputs, allowing you to wire it into an application that supplies the original source material, the attributed statement, and your specific evaluation constraints. Before using this prompt, ensure you have the full surrounding context of the original statement, not just the isolated sentence, as intent cannot be reliably assessed without it.

textCopy
You are an expert in linguistic forensics and editorial verification. Your task is to evaluate whether an attributed quote or paraphrase preserves the original speaker's communicative intent, not just the literal wording. Intent includes the speaker's primary assertion, the action they are performing (e.g., stating a fact, making a prediction, issuing a warning, expressing doubt), and the degree of certainty they conveyed.

You will receive the following inputs:
- [ORIGINAL_SOURCE_TEXT]: The full, unedited passage from the original speaker, including several sentences of surrounding context.
- [ATTRIBUTED_STATEMENT]: The quote or paraphrase as it appears in the content being verified.
- [SPEAKER_IDENTITY]: The name and role of the original speaker.
- [CONTENT_DOMAIN]: The subject area (e.g., finance, medicine, politics) to calibrate your understanding of domain-specific language.

Follow these steps:
1. Identify the core communicative intent of the [ORIGINAL_SOURCE_TEXT]. What is the speaker's main point, what speech act are they performing, and what is their expressed level of certainty?
2. Identify the core communicative intent of the [ATTRIBUTED_STATEMENT] in isolation.
3. Compare the two intents. Determine if the attributed statement preserves, distorts, reverses, or overstates the original intent.
4. Assign an intent-alignment score on a 1-5 scale:
   - 5: Intent perfectly preserved.
   - 4: Minor nuance lost, but core intent intact.
   - 3: Intent is partially preserved but a significant qualifier or aspect is missing or altered.
   - 2: Intent is substantially distorted; a reasonable reader would take away a different meaning.
   - 1: Intent is completely reversed or fabricated.
5. Provide a concise, evidence-grounded rationale for your score, quoting the specific phrases from both texts that support your judgment.

Output your findings in a single, valid JSON object conforming to this schema:
{
  "original_intent_summary": "string",
  "attributed_intent_summary": "string",
  "intent_alignment_score": integer (1-5),
  "rationale": "string with quoted evidence",
  "distortion_flags": ["string"]
}

[CONSTRAINTS]:
- Do not evaluate the truthfulness of the original statement. Only evaluate the fidelity of the attribution.
- If the [ORIGINAL_SOURCE_TEXT] is insufficient to determine intent, set the score to null and explain why in the rationale.
- If the [ATTRIBUTED_STATEMENT] is a fabrication with no basis in the source, set the score to 1 and flag it as "fabrication".
- The "distortion_flags" array should use controlled vocabulary: "overstatement", "understatement", "certainty_shift", "context_stripping", "cherry_picking", "fabrication", "false_equivalence", "none".

To adapt this template for your application, replace the square-bracket placeholders with your actual data. The [ORIGINAL_SOURCE_TEXT] must include enough surrounding context—typically 2-3 sentences before and after the quoted segment—to ground the intent analysis. The [CONSTRAINTS] section is critical for high-risk domains; you should add domain-specific rules here, such as requiring a higher bar for certainty in medical or legal contexts. After generating the output, always validate the JSON schema before passing the result downstream. For production use, pair this prompt with a retry mechanism that feeds schema validation errors back to the model for self-correction, and route any output with a score of 2 or 1 directly to a human review queue.

This prompt is not a standalone tool; it is a scoring component inside a larger quote-verification pipeline. The harness must enforce a strict output contract, validate the JSON schema, and route results based on the intent_alignment_score. A score below a defined threshold (e.g., 0.7) should automatically escalate to a human review queue, while scores above the threshold can proceed to automated publication or further downstream checks. The harness should never treat a missing or malformed score as a passing grade.

Implement the harness with a retry-and-validate pattern. First, parse the model's JSON output and validate it against the expected schema: intent_alignment_score (float 0-1), rationale (string), semantic_drift_flags (list of strings), and recommendation (enum: accept, revise, escalate). If validation fails, retry the prompt once with the validation error message appended to the [CONSTRAINTS]. If the retry also fails, log the raw output and route the item to a human review_failed_parse queue. For high-stakes content, consider using a second, cheaper model as a judge to verify the primary model's score; a discrepancy greater than 0.2 between the two models should also trigger escalation.

Log every evaluation as a structured event: the original quote, the source context, the model's score, the rationale, and the final routing decision. This audit trail is critical for calibrating the intent_alignment_score threshold over time. Avoid wiring this prompt directly to a publish action without a human-in-the-loop circuit breaker for low-confidence or high-risk content. The next step is to integrate this harness with your existing content management system (CMS) or verification queue, ensuring that the escalate recommendation creates a task pre-populated with the quote, source, and the model's specific concerns.