Contradiction Triangulation Prompt Across Three or More Sources

This prompt is for verification analysts and pipeline engineers who need to resolve factual conflicts across three or more sources. Instead of treating contradiction as a binary flag, this prompt triangulates a weighted consensus, surfaces minority reports, adjusts for source reliability, and produces confidence intervals. Use it when pairwise comparison is insufficient and you need a structured resolution that a downstream system or human reviewer can act on. This prompt belongs after claim extraction and evidence matching, and before final verification report generation. It assumes you already have extracted claims and their supporting evidence from each source.

Do not use this prompt when you have only two sources—a pairwise comparison prompt is cheaper, faster, and less prone to false equivalence. Avoid it when sources are not independent (e.g., one source is a reprint of another), when the claim is subjective or unverifiable, or when you lack source reliability metadata. The triangulation logic requires at least three distinct evidence-bearing sources with known provenance. If you feed it two identical sources and one outlier, the weighted consensus will be misleading. If you feed it sources without reliability scores, the prompt will still produce output, but the confidence intervals will be unreliable. Always validate that your input includes source authority metadata, claim-to-evidence mappings, and a defined output schema before invoking this prompt.

After running this prompt, route the output to a verification report generator or a human review queue depending on the consensus strength. If the confidence interval is wide or the stalemate flag is raised, escalate for human judgment rather than treating the weighted consensus as ground truth. For production pipelines, pair this prompt with a convergence detection check: if the top two sources agree and the third is an outlier with low reliability, auto-resolve; if three high-reliability sources disagree, escalate. Never use this prompt as the final step in a fully automated pipeline without a human-in-the-loop escape hatch for unresolved conflicts.

This template is designed to be copied directly into your application or testing environment. It instructs the model to act as a verification analyst, comparing claims from multiple sources, identifying where they agree and conflict, and producing a structured assessment. The prompt uses square-bracket placeholders for all dynamic inputs—replace each one with your specific data before sending it to the model. The output is a JSON object designed for machine consumption in a fact-checking pipeline, including a consensus assessment, a minority report for dissenting sources, and a confidence interval.

textCopy
You are a verification analyst performing multi-source contradiction triangulation. Your task is to compare claims from three or more provided sources, identify areas of agreement and direct contradiction, and produce a weighted consensus assessment.

## INPUT
### Claim Under Review
[CLAIM]

### Source Evidence
You will be given evidence from [NUMBER_OF_SOURCES] sources. Each source includes a reliability score from 1 (low) to 5 (high) and a recency date.

[SOURCE_1_ID]: Reliability=[SOURCE_1_RELIABILITY], Recency=[SOURCE_1_DATE]
[SOURCE_1_TEXT]

[SOURCE_2_ID]: Reliability=[SOURCE_2_RELIABILITY], Recency=[SOURCE_2_DATE]
[SOURCE_2_TEXT]

[SOURCE_3_ID]: Reliability=[SOURCE_3_RELIABILITY], Recency=[SOURCE_3_DATE]
[SOURCE_3_TEXT]

[ADDITIONAL_SOURCES]

## CONSTRAINTS
- Do not introduce external knowledge. Base your analysis solely on the provided source texts.
- If sources agree, state the consensus clearly. If they conflict, describe the nature of the contradiction (direct factual, numerical, temporal, interpretive).
- Weight sources by their reliability score and recency. A more recent, high-reliability source should have greater influence on the consensus.
- If no clear consensus can be reached, declare a stalemate and explain why the evidence is insufficient to resolve the conflict.
- For numerical claims, note the specific values and any tolerance windows implied by the sources.

## OUTPUT_SCHEMA
Return a single JSON object with the following structure:
{
  "claim": "The original claim under review.",
  "consensus_assessment": {
    "verdict": "supported|contradicted|uncertain|stalemate",
    "majority_view": "A concise statement of the most likely truth based on weighted evidence.",
    "confidence": 0.0-1.0,
    "confidence_rationale": "Explanation of the confidence score, referencing source agreement and reliability."
  },
  "minority_report": {
    "exists": true|false,
    "dissenting_view": "The conflicting perspective, if any.",
    "dissenting_sources": ["SOURCE_ID"],
    "reason_for_dissent": "Why these sources disagree, e.g., outdated info, lower reliability, different methodology."
  },
  "source_alignment": [
    {
      "source_id": "SOURCE_ID",
      "position": "supports|contradicts|partially_supports|unrelated",
      "evidence_excerpt": "Relevant quote from the source.",
      "weight_applied": 0.0-1.0
    }
  ],
  "conflict_analysis": {
    "conflict_type": "direct_factual|numerical|temporal|interpretive|none",
    "description": "A clear explanation of the conflict, if any.",
    "resolvable": true|false,
    "resolution_path": "Suggested next step, e.g., 'Check source [SOURCE_2_ID] methodology' or 'Escalate for human review.'"
  }
}

To adapt this template for your pipeline, start by mapping your internal source metadata to the [SOURCE_ID], [SOURCE_RELIABILITY], and [SOURCE_DATE] placeholders. If you don't have explicit reliability scores, a simple heuristic based on source type (e.g., primary document=5, reputable news=4, opinion piece=2) is better than omitting the field. The [ADDITIONAL_SOURCES] placeholder should repeat the source block for sources four through N. Before deploying, test the prompt against a golden dataset of known consensuses and stalemates to calibrate the confidence thresholds and ensure the model isn't fabricating a majority view when evidence is evenly split.

The Contradiction Triangulation Prompt is designed to be a decision-support node within a larger fact-checking pipeline, not a standalone chatbot. It expects three or more pre-extracted source passages, each with associated metadata (source ID, publication date, authority score), and a set of claims that have already been flagged as conflicting by an upstream pairwise comparison step. The prompt's job is to resolve the multi-source conflict by producing a weighted consensus, a minority report, and a confidence interval. Wiring this into an application requires a clear contract: the input schema must include sources (an array of objects with id, text, date, and authority_score), conflicting_claims (an array of claim strings), and resolution_parameters (an object specifying minimum_authority_threshold, stalemate_confidence_floor, and require_minority_report as a boolean). The output contract should be a structured JSON object with consensus_assessment, minority_report, confidence_interval, source_weighting_rationale, and unresolved_flags.

In practice, you should wrap the LLM call in a harness that performs pre-flight validation, post-processing, and retry logic. Before sending the prompt, validate that at least three sources are present, that each source has a non-empty text field, and that authority scores are normalized between 0 and 1. After receiving the response, parse the JSON and run a schema validator to ensure all required fields exist and that the confidence_interval contains valid lower and upper bounds. If validation fails, retry once with a repair prompt that includes the original input and the specific validation error. For high-stakes domains like healthcare or finance, route any output where unresolved_flags is non-empty or where the confidence interval width exceeds a configurable threshold (e.g., >0.4) to a human review queue. Log the full prompt, response, and validation result for auditability. Model choice matters: use a model with strong reasoning benchmarks (e.g., GPT-4o, Claude 3.5 Sonnet) and set temperature to 0 or very low to maximize deterministic, repeatable assessments.

The most common production failure mode is stalemate handling. When sources are equally authoritative and directly contradict each other, the model may produce an overconfident consensus or an unhelpfully vague one. Your harness should detect this by checking if the consensus_assessment confidence is below the stalemate_confidence_floor you configured. In that case, suppress the consensus output in your application UI, surface the minority_report prominently, and escalate for human review. Do not present a machine-generated consensus as resolved when the model itself signals low confidence. A second failure mode is source-count degradation: if your pipeline occasionally retrieves only two sources due to a retrieval failure, do not fall back to this prompt. Route two-source conflicts to a dedicated Pairwise Claim Comparison Prompt instead. Finally, avoid running triangulation on claims that are fundamentally interpretive or opinion-based; pre-filter claims using a Fact-Interpretation Separation prompt to ensure you're only triangulating verifiable factual assertions.