Contradiction-Aware Summary Generation Prompt

This prompt is designed for summarization pipelines where suppressing source conflicts is a critical failure mode, not a minor flaw. The primary job-to-be-done is generating a concise summary from multiple documents while explicitly surfacing contradictions, representing conflicting viewpoints faithfully, and marking uncertainty where evidence diverges. The ideal user is an engineering lead or AI builder integrating summarization into a fact-checking, legal review, financial audit, or multi-source research product. You need this prompt when your downstream consumers—analysts, reviewers, or automated systems—make decisions based on the summary and must be warned about unresolved factual disagreements, not given a falsely coherent narrative.

Do not use this prompt when you need a single, authoritative narrative that resolves conflicts by selecting a 'best' source. It is also inappropriate for low-risk content aggregation where contradiction flagging adds noise without actionable value. The prompt requires at least two source documents with potentially conflicting claims as input. If your pipeline only has a single source, use a standard summarization prompt with evidence grounding instead. The operational context assumes you have already retrieved or collected source material; this prompt does not perform retrieval. It operates on provided context only and must not invent external facts to resolve disputes.

Before implementing this prompt, ensure your evaluation harness includes specific checks for conflict suppression bias—the tendency of models to harmonize disagreements into a bland consensus. You must also define a clear output contract that separates factual summaries from contradiction flags, preventing the model from burying conflicts in footnotes. Wire this prompt into a pipeline where the output feeds a verification UI or a downstream triage system, not a public-facing content surface without human review. If the contradiction severity is high-stakes (e.g., clinical data, legal filings), always route the flagged output for human approval before any action is taken.

This template is designed for summarization pipelines where hiding disagreement is worse than reporting it. Use it when your input consists of multiple source documents that may conflict on facts, interpretations, or timelines. The prompt instructs the model to produce a summary that includes explicit contradiction flags, conflicting viewpoint representation, and calibrated uncertainty markers. Do not use this prompt when you need a single coherent narrative that resolves all tension—this prompt is for surfacing conflict, not resolving it.

textCopy
You are a contradiction-aware summarization system. Your task is to produce a summary that faithfully represents the content of the provided source documents, with explicit attention to conflicts, disagreements, and inconsistencies between them.

## INPUT
Source documents are provided below. Each document is labeled with a source ID in square brackets.

[SOURCES]

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "summary": "A concise overall summary of the topic covered by the sources.",
  "points_of_agreement": [
    {
      "statement": "A factual statement that all sources agree on.",
      "supporting_sources": ["source_id_1", "source_id_2"]
    }
  ],
  "contradictions": [
    {
      "id": "contradiction-1",
      "claim_a": "The claim as stated in the first source.",
      "source_a": "source_id",
      "claim_b": "The conflicting claim from another source.",
      "source_b": "source_id",
      "conflict_type": "direct_contradiction | partial_mismatch | temporal_inconsistency | numerical_disagreement | logical_conflict",
      "severity": "high | medium | low",
      "explanation": "A brief explanation of why these claims conflict."
    }
  ],
  "unresolved_issues": [
    {
      "issue": "A topic where sources disagree and no clear resolution is possible from the provided material.",
      "positions": ["Summary of position from source A", "Summary of position from source B"],
      "recommendation": "What a reader should do with this unresolved conflict (e.g., seek additional sources, treat both claims with caution)."
    }
  ],
  "uncertainty_markers": [
    {
      "statement": "A statement where the model has low confidence due to conflicting or insufficient evidence.",
      "reason": "Why this statement is uncertain."
    }
  ]
}

## CONSTRAINTS
- Do not fabricate agreement where sources disagree.
- Do not omit contradictions to produce a cleaner summary.
- If only one source addresses a claim, mark it as uncertain rather than treating it as agreed-upon fact.
- For each contradiction, cite the exact source IDs. Do not invent source references.
- If no contradictions exist, return an empty contradictions array. Do not invent conflicts.
- Preserve the original framing of each source's claims. Do not rephrase to make them seem more compatible than they are.

## EXAMPLES
[EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]

To adapt this template, replace the square-bracket placeholders with your specific inputs. The [SOURCES] placeholder should contain your documents with clear source IDs. The [EXAMPLES] placeholder is critical for production use—provide at least one few-shot example showing a correct output with both agreements and contradictions to calibrate the model's behavior. The [RISK_LEVEL] placeholder should be set to high for domains like healthcare, finance, or legal where contradiction suppression could cause harm; this triggers stricter adherence to the conflict-surfacing instructions. After generating the output, validate the JSON schema and verify that every cited source ID actually exists in the input. For high-risk applications, route outputs with high severity contradictions to human review before publication.

This prompt is not a standalone chat instruction; it is a component in a summarization pipeline that must surface, rather than suppress, source conflicts. The primary integration point is after evidence retrieval and before the final summary is presented to a user or downstream system. The application layer is responsible for assembling the [INPUT_DOCUMENTS], which should be pre-segmented and deduplicated source passages. The [CONTRADICTION_RULES] placeholder should be populated from a configuration store, not hardcoded, allowing operations teams to tune sensitivity without touching the prompt text. The model call should be wrapped in a function that enforces the [OUTPUT_SCHEMA], typically a JSON object with summary_text, contradiction_flags (an array of objects with claim_a, claim_b, source_a, source_b, conflict_type), and uncertainty_markers.

Validation is the critical harness component here. After receiving the model response, you must run a structural validation against the expected schema, rejecting any response where contradiction_flags is not a valid array or where required fields are missing. A more important semantic validation must check for 'conflict suppression bias'—the model's tendency to harmonize away genuine disagreements. Implement a secondary LLM call or a heuristic check that compares the number of flagged contradictions against a known baseline from a pairwise contradiction detection prompt run on the same inputs. If the summary's contradiction count is statistically lower than the pairwise check suggests, log a warning, increment a suppression_risk metric, and either retry with a lower temperature or escalate for human review. All model responses, including validation failures, must be logged with the prompt version, input document hashes, and a timestamp for auditability.

For model choice, use a model with strong instruction-following and long-context handling, as the combined input documents can be large. GPT-4o or Claude 3.5 Sonnet are suitable starting points. Set the temperature low (0.1–0.2) to reduce variability in the structured output. Implement a retry strategy with exponential backoff for schema validation failures, but cap retries at three attempts to control latency. If the prompt is used in a high-stakes domain like medical or financial summarization, the final output must be routed to a human review queue, and the uncertainty_markers field should be visually highlighted in the review interface. Avoid wiring this prompt directly into a customer-facing chat without these validation and review guardrails; a raw summary that silently drops a critical contradiction creates more risk than no summary at all.