Contradictory Evidence Surface Prompt

This prompt is designed for a single, high-stakes job: when your RAG system retrieves multiple documents that disagree on a material fact, date, obligation, or interpretation, you must expose the conflict to the user rather than silently resolving it. The ideal user is an AI engineer or product builder in legal, finance, or compliance domains where hiding a disagreement creates regulatory, contractual, or decision risk. The required context is a set of retrieved passages that contain a genuine contradiction—not just different phrasings of the same fact, but competing evidence positions that a reasonable reader would recognize as inconsistent.

Use this prompt when the cost of a wrong answer is higher than the cost of an unresolved answer. For example, if two contracts assign the same obligation to different parties, or two financial filings report different revenue figures for the same period, the system's job is to present both positions with their sources, not to guess which is correct. The prompt is structured to produce a neutral, evidence-anchored output with explicit conflict metadata: each competing position gets its own block with source attribution, a direct quote, and a plain-language summary of what is claimed. A dedicated conflict summary section then explains the nature of the disagreement in terms of materiality, temporality, or interpretation.

Do not use this prompt for general Q&A where sources are consistent or where users expect a single definitive answer. It is not a fact-verification prompt—it surfaces conflicts but does not adjudicate them. It is not a synthesis prompt—it deliberately avoids merging contradictory claims into a compromise. If your system needs to resolve conflicts by preferring more recent, more authoritative, or more complete sources, use a temporal conflict resolution or evidence ranking prompt instead. This prompt is a surfacing tool, not a resolution tool. Before deploying it, ensure your retrieval pipeline is returning genuinely conflicting passages; feeding it consistent passages will produce false conflict flags that erode user trust.

This template is designed for high-stakes RAG applications in legal, finance, and compliance where sources disagree and the system must expose rather than hide conflict. The prompt instructs the model to identify competing positions, attribute each to its source, and present the disagreement without taking sides. Replace every square-bracket placeholder with your application's specific values before deployment.

textCopy
You are a contradiction-aware evidence analyst. Your job is to answer questions using only the provided context, and to actively surface any contradictory evidence you find.

## CONTEXT
[CONTEXT]

## QUESTION
[INPUT]

## INSTRUCTIONS
1. Identify all distinct positions or claims present in the context that relate to the question.
2. If the context contains contradictory evidence, you MUST surface it. Do not ignore, reconcile, or choose a winner.
3. For each position, provide:
   - The claim or finding.
   - The source identifier from the context.
   - Any relevant date, jurisdiction, or authority metadata if present.
4. If the context is insufficient to answer the question, state that explicitly and describe what is missing.
5. Do not introduce knowledge outside the provided context.
6. Do not express a preference for one position over another.

## OUTPUT FORMAT
Return a JSON object with this exact schema:
{
  "answerable": boolean,
  "positions": [
    {
      "claim": "string",
      "source": "string",
      "metadata": {
        "date": "string or null",
        "jurisdiction": "string or null",
        "authority": "string or null"
      },
      "supporting_excerpt": "string"
    }
  ],
  "conflicts": [
    {
      "description": "string describing the nature of the conflict",
      "position_indices": [0, 1]
    }
  ],
  "gaps": ["string describing missing information"],
  "neutral_summary": "string summarizing the state of evidence without taking sides"
}

## CONSTRAINTS
- Never fabricate source identifiers.
- If only one position exists, the conflicts array must be empty.
- If the context contains no relevant evidence, set answerable to false and return empty positions.
- Keep supporting excerpts under 200 characters.

To adapt this template, start by replacing [CONTEXT] with your retrieved passages, ensuring each chunk has a stable identifier. Replace [INPUT] with the user's question. The output schema is strict JSON; validate it before surfacing to users. For regulated domains, add a human review step before the neutral summary is published. Test the prompt with known contradictory document pairs to verify that both positions appear and that the model does not fabricate a resolution.

The Contradictory Evidence Surface Prompt is not a standalone Q&A prompt—it is a conflict-detection and surfacing component that sits between evidence retrieval and final answer delivery. In a production architecture, this prompt should be invoked when the retrieval step returns documents with conflicting claims, timestamps, or authority levels. The application layer is responsible for detecting potential conflict before calling this prompt. A pre-check step can use lightweight heuristics—such as named entity overlap with opposing sentiment, date range collisions, or source authority mismatches—to decide whether to route the query through a standard faithful synthesis prompt or this conflict-surfacing variant. Do not call this prompt for every query; it is designed for the subset of cases where hiding disagreement would be misleading.

The implementation harness requires several integration points. First, the prompt expects a structured [EVIDENCE_BLOCKS] input containing pre-processed passages with metadata: source identifier, publication date, authority level, and the raw text. Build a formatter that converts your retrieval results into this schema before injection. Second, implement output validation that checks for the required conflict structure: each competing position must have at least one supporting source, the neutral framing must be present, and no position should be presented as definitively correct unless the evidence is unanimous. Use a JSON schema validator on the structured output to enforce these constraints. If validation fails, retry once with the validation errors appended to the prompt as feedback. If the second attempt also fails, escalate to a human review queue with the original evidence blocks and both failed outputs attached.

Model choice matters here. This prompt benefits from models with strong reasoning and instruction-following capabilities, particularly for maintaining neutral framing under conflicting evidence. GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro are suitable starting points. Avoid smaller or older models that may collapse conflicting evidence into a false consensus or fabricate a resolution. For high-stakes domains like legal and finance, log every invocation with the full prompt, retrieved evidence, model response, and validation results. These logs serve as audit evidence and are essential for evaluating conflict completeness over time. Implement a periodic human review sample—at least 10% of conflict-surfacing responses—to check that the model is not systematically favoring certain sources or introducing bias in its neutral framing. Finally, wire the validated output into your answer delivery layer with clear UI affordances that distinguish competing positions, source attributions, and the absence of a resolved answer.