Fact-Interpretation Separation System Prompt

This prompt is designed for verification pipeline builders who need a reusable, model-level instruction to classify every sentence in a document. Its job is to separate verifiable factual statements from analysis, opinion, prediction, and rhetorical framing before any fact-checking occurs. The ideal user is an engineer or technical operator integrating this classification step into a larger automated workflow—such as claim extraction, evidence matching, or human review routing. You should use this prompt when downstream tasks require clean, structured labels and confidence scores to avoid misclassifying an analyst's speculation as a checkable fact. It is not a one-shot fact-checking prompt; it is a classification contract that runs before verification begins.

Do not use this prompt when you need a final truth verdict on a claim, when the input is a single isolated sentence with no surrounding context, or when the operational latency budget cannot accommodate sentence-by-sentence classification. This prompt also underperforms on content where facts and interpretation are deliberately fused to mislead—adversarial blurring requires a separate detection step. The prompt assumes the input document contains discernible linguistic boundaries between fact and non-fact statements. If your source material is entirely speculative (e.g., a futurist essay) or entirely factual (e.g., a weather log), the classification output will be uniform and the computational cost may not be justified. In those cases, a simpler keyword-based or rule-based filter may suffice.

Before wiring this prompt into production, you must define the downstream consumer of its output. If the labeled sentences feed an evidence-matching system, ensure that system can handle the confidence and boundary_rationale fields without discarding them. If the output routes ambiguous cases to human reviewers, agree on a confidence threshold below which escalation occurs—typically, sentences scored below 0.7 confidence for the fact or interpretation labels are good candidates for review. The prompt includes eval checks for consistent classification across document types, but you should supplement these with your own golden dataset of 50–100 sentences labeled by domain experts. Measure inter-annotator agreement between the model and your experts before trusting the output in high-risk domains such as legal, financial, or medical content. In those domains, always route low-confidence outputs to a human and log the model's classification alongside the reviewer's final label for audit and continuous improvement.

This system prompt is designed to be set once in the model's system instruction field and remain stable across multiple user messages, each containing a document to classify. It enforces a strict output schema, requires confidence scores for every classification, and demands boundary rationales when a single sentence contains multiple statement types. The prompt is self-contained: it defines the taxonomy, output format, and behavioral rules without relying on external context or conversation history.

textCopy
You are a precise document classifier. Your only job is to classify every sentence in the provided document into exactly one of these five categories:

1. FACT: A verifiable statement that can be checked against external evidence. Facts are objective claims about observable reality, events, measurements, or documented statements. Examples: "The company reported $4.2B in revenue." "The bill was signed on March 12."

2. INTERPRETATION: An inference, analysis, or conclusion drawn from facts but not directly verifiable as a standalone statement. Interpretations apply reasoning to evidence. Examples: "This revenue decline suggests weakening demand." "The delay indicates resource constraints."

3. OPINION: A subjective judgment, preference, or value statement that cannot be resolved by evidence alone. Opinions express personal or organizational stance. Examples: "This was a poor decision." "The design is elegant."

4. PREDICTION: A forward-looking statement about future events, outcomes, or states. Predictions are not currently verifiable regardless of how confidently stated. Examples: "Revenue will grow 12% next quarter." "The regulation is likely to pass."

5. RHETORICAL_FRAMING: Language whose primary function is persuasion, emphasis, narrative construction, or emotional appeal rather than conveying checkable content. Examples: "We are at a crossroads." "This changes everything."

RULES:
- Classify every sentence independently. Do not let surrounding sentences influence a classification.
- If a sentence contains multiple statement types, classify it as the dominant type AND include a boundary_rationale explaining which parts belong to which category.
- Provide a confidence score from 0.0 to 1.0 for every classification. Use 0.0 for complete uncertainty and 1.0 for absolute certainty.
- If confidence is below 0.7, include a specific reason in the confidence_note field.
- Do not summarize, analyze, or comment on the document beyond the classification output.
- Do not refuse to classify. If a sentence is ambiguous, classify it to the best of your ability with appropriately low confidence.

OUTPUT FORMAT: Return a JSON object with this exact structure:
{
  "classifications": [
    {
      "sentence_index": 0,
      "text": "[EXACT_SENTENCE_TEXT]",
      "category": "[FACT|INTERPRETATION|OPINION|PREDICTION|RHETORICAL_FRAMING]",
      "confidence": [0.0-1.0],
      "confidence_note": "[REQUIRED_IF_CONFIDENCE_BELOW_0.7]",
      "boundary_rationale": "[REQUIRED_IF_MULTIPLE_TYPES_PRESENT]"
    }
  ]
}

[INPUT_DOCUMENT]

To adapt this prompt for your domain, modify the category definitions and examples to match your terminology. For legal documents, replace the generic examples with clause-level illustrations. For financial content, add domain-specific edge cases like distinguishing reported figures from adjusted measures. The output schema should remain stable; only the taxonomy descriptions and examples should change. Before deploying, run this prompt against a golden dataset of 50-100 sentences with known classifications and measure agreement against human labels. If agreement falls below 0.85 F1 on FACT vs. INTERPRETATION boundaries, add domain-specific few-shot examples to the category definitions rather than increasing instruction length.

This system prompt is designed as a reusable classification layer that sits between content ingestion and downstream verification. It expects raw text input and returns structured labels for every sentence. In production, you'll typically deploy this as a pre-processing step: raw documents flow in, the model classifies each sentence as fact, interpretation, opinion, prediction, or rhetorical framing, and only sentences tagged as 'fact' proceed to evidence matching and claim verification. Sentences tagged as interpretation, opinion, or prediction should be routed to separate analysis pipelines or flagged for human review rather than being silently treated as verifiable claims.

Wire the prompt into your application by constructing a request that combines this system instruction with the target document as user input. Use a model that supports structured output with a defined JSON schema matching the expected output shape: an array of sentence objects, each containing sentence_index, text, classification, confidence, and boundary_rationale. Implement a validation layer that checks for schema compliance, missing fields, confidence scores below your threshold (typically 0.7 for auto-routing), and sentences where the model flags boundary ambiguity. For high-stakes domains like legal or medical content, set a stricter confidence floor (0.85+) and route anything below it to human review. Log every classification decision with the model version, prompt hash, input document identifier, and timestamp for audit trails.

Build retry logic for malformed outputs: if the response fails schema validation, retry once with the same prompt and a stronger constraint instruction appended. If the second attempt fails, escalate the entire document to a human review queue rather than silently dropping sentences. For batch processing, implement parallel requests with a concurrency limit appropriate to your rate limits, and include a post-processing step that checks for cross-document classification consistency using the eval rubric from the companion Fact-Interpretation Separation Eval Rubric Prompt. Avoid wiring this prompt directly into user-facing features without a human-in-the-loop fallback—misclassifying an interpretation as a fact can cascade into false verification results downstream. Start with a shadow mode deployment where classifications are logged and compared against human labels before enabling automated routing.