Contradiction Resolution Prompt for Competing Evidence

This prompt is designed for RAG and multi-source research pipelines where retrieved evidence contains conflicting facts, dates, claims, or interpretations. The core job-to-be-done is to produce a structured conflict report that identifies specific contradictions, traces each claim to its source provenance, and recommends a resolution strategy—rather than letting the model silently choose one source or hallucinate a blended answer. The ideal user is an AI engineer or product developer building a system where downstream decisions depend on accurate evidence synthesis, such as legal research, clinical literature review, financial due diligence, or competitive intelligence. You need this prompt when your retrieval step returns documents that disagree and you cannot afford the model to paper over those disagreements with fluent but misleading prose.

The prompt requires several inputs to function correctly: a set of evidence passages with source identifiers, a specific question or claim being evaluated, and a defined output schema for the conflict report. You must also supply a risk level that controls whether the model recommends a resolution or simply flags the contradiction for human review. In high-stakes domains—healthcare, legal, compliance, safety—the resolution strategy field should be gated behind human approval. The prompt is not appropriate when all sources agree, when you need a single synthesized answer without provenance tracking, or when the contradiction is trivial and does not affect downstream decisions. Do not use this prompt for real-time chat where latency constraints prevent structured conflict analysis, or for tasks where the model lacks sufficient context to compare claims across sources.

Before deploying this prompt, ensure your retrieval pipeline preserves source metadata (document ID, date, author, authority tier) and that your application layer can enforce the human-review gate for high-risk resolutions. The prompt template includes placeholders for evidence, query, output schema, and risk level—adapt these to your domain's taxonomy and review workflow. After reading this section, proceed to the prompt template to copy the base structure, then review the implementation harness for guidance on validation, retries, and production wiring.

The following prompt template is designed to be copied directly into your prompt management system, IDE, or orchestration layer. It accepts a set of competing evidence passages and produces a structured conflict report that identifies contradictions, traces them to their sources, and recommends a resolution strategy. Every placeholder is enclosed in square brackets and must be replaced with real values before inference. The template is intentionally strict about output structure because downstream systems—review queues, logging pipelines, and audit tools—depend on consistent JSON.

textCopy
You are a contradiction resolution analyst. Your task is to examine a set of evidence passages that may contain conflicting claims about the same topic. You must identify every contradiction, trace each conflicting claim to its source, assess the relative reliability of each source, and recommend a resolution strategy.

## INPUT EVIDENCE
[EVIDENCE_PASSAGES]

## TASK INSTRUCTIONS
1. Extract every factual claim from the evidence passages.
2. Group claims that address the same fact, date, figure, event, or assertion.
3. For each group where claims disagree, produce a contradiction entry.
4. For each contradiction, assess source reliability based on [RELIABILITY_CRITERIA].
5. Recommend a resolution strategy from the allowed list: [RESOLUTION_STRATEGIES].
6. Flag any contradiction that requires human review based on [RISK_LEVEL].

## OUTPUT SCHEMA
Return a single JSON object with this exact structure:
{
  "contradictions": [
    {
      "contradiction_id": "string",
      "claim_group": "string describing the disputed fact",
      "conflicting_claims": [
        {
          "claim_text": "string",
          "source_id": "string matching source identifier from evidence",
          "source_excerpt": "verbatim quote from source",
          "reliability_assessment": "string"
        }
      ],
      "resolution_strategy": "string from allowed list",
      "resolution_rationale": "string explaining why this strategy was chosen",
      "requires_human_review": boolean,
      "human_review_reason": "string or null"
    }
  ],
  "uncontested_claims": [
    {
      "claim_text": "string",
      "supporting_sources": ["source_id"]
    }
  ],
  "resolution_summary": "string summarizing overall findings and recommended actions"
}

## CONSTRAINTS
- Do not invent claims absent from the evidence.
- Do not resolve contradictions by preferring the most fluent or confident-sounding source.
- Every conflicting claim must include a verbatim source excerpt.
- If reliability cannot be assessed, mark the contradiction for human review.
- If [RISK_LEVEL] is "high," all contradictions require human review regardless of confidence.
- Return only valid JSON. No markdown fences, no commentary outside the JSON object.

To adapt this template, replace [EVIDENCE_PASSAGES] with your retrieved or provided source material, formatted with clear source identifiers that the model can reference in its output. Define [RELIABILITY_CRITERIA] based on your domain—common criteria include source recency, author expertise, publication venue, corroboration count, and primary vs. secondary status. Set [RESOLUTION_STRATEGIES] to a closed list such as ["prefer_most_recent", "prefer_primary_source", "prefer_majority_corroborated", "flag_for_human_review", "report_disagreement_without_resolution"]. Set [RISK_LEVEL] to "low", "medium", or "high" based on your domain's tolerance for automated resolution. For high-stakes domains like healthcare, finance, or legal, always set risk level to "high" and ensure a human reviews every contradiction before any downstream action is taken.

The Contradiction Resolution Prompt is not a standalone chat interaction; it is a pipeline stage that sits between evidence retrieval and final answer generation. In a typical RAG architecture, you retrieve multiple passages, detect that they conflict on key facts, and then invoke this prompt to produce a structured conflict report. The report becomes input to a downstream decision: either an automated resolution strategy (prefer the most recent source, prefer the highest-authority source) or a human review queue. Wire this prompt as a synchronous call within an async processing pipeline, with a strict timeout and a well-defined output schema that your application can parse and act on.

Start by defining the input contract. The prompt expects a [QUERY] describing the question or claim under investigation, a [SOURCES] array where each source object includes an id, text, date, and authority field, and a [CONFLICT_DESCRIPTION] summarizing the nature of the disagreement. On the application side, validate that [SOURCES] contains at least two entries with non-empty text fields before calling the model. If the retrieval step returns only one source or sources that agree, skip this prompt entirely and route directly to answer generation. For model choice, prefer a model with strong reasoning and structured output capabilities (e.g., Claude 3.5 Sonnet, GPT-4o) because the task requires careful comparison and nuanced judgment. Set temperature to 0 or near-zero to maximize consistency across runs. Implement a JSON schema validator on the response: the output must contain a conflicts array where each conflict has claim, source_positions (mapping source IDs to their stances), resolution_strategy, and recommended_resolution fields. If the model returns malformed JSON, retry once with the same prompt plus a brief repair instruction appended to the system message. If the second attempt fails, log the raw output and escalate to a human operator rather than silently proceeding with a broken conflict report.

Human review gating is the most critical harness decision. For high-stakes domains—legal, healthcare, finance, safety—configure the pipeline so that any conflict report with a resolution_strategy of manual_review_required or a confidence score below a configurable threshold (e.g., 0.8) automatically routes to a review queue. Do not auto-resolve conflicts in these domains. For lower-stakes applications, you can implement automated resolution rules: if the conflict is purely about recency and one source is significantly newer, auto-select the newer source; if the conflict is about authority and one source is a primary document while another is a secondary summary, auto-select the primary. Log every automated resolution decision with the full conflict report, the rule that fired, and the selected resolution for auditability. Instrument the harness with metrics: conflict detection rate, resolution strategy distribution, human review escalation rate, and end-to-end latency from retrieval to resolved answer. These metrics will tell you whether the prompt is catching real contradictions or generating noise, and whether your resolution rules are aligned with actual source quality.