Contradiction Explanation Generation Prompt

Use the Contradiction Explanation Generation Prompt when your application must surface evidential disagreements to end users, compliance reviewers, or analysts who need to assess source reliability themselves. The core job-to-be-done is transparency: you have already detected that two or more evidence sources directly contradict each other on a specific claim, and now you need a human-readable explanation that grounds every reason for conflict in the provided source content. This prompt is designed for explainability features in research assistants, decision-support systems, and RAG applications where users need to understand why sources disagree, not just that they disagree. The ideal user is a product engineer or AI builder integrating this into a pipeline that has already completed conflict detection—this prompt does not find conflicts; it explains them.

The prompt requires specific, pre-validated inputs to function correctly. You must provide the exact claim under dispute, the full text of the conflicting source passages, and any relevant metadata such as publication dates, author credentials, or document types. Without this context, the model will speculate about reasons for conflict rather than grounding its explanation in evidence. The prompt forces the model to cite specific spans from each source when describing what each source claims, preventing hallucinated or vague attributions. This is not a general-purpose summarization prompt—it assumes you have already identified a genuine contradiction through a separate detection step, such as the Multi-Source Conflict Detection Prompt Template or a structured comparison pipeline.

Do not use this prompt when the disagreement is merely a difference in scope, methodology, or terminology rather than a direct factual contradiction. If sources discuss different populations, time periods, or definitions of a term, a contradiction explanation will mislead users by framing compatible statements as conflicting. Similarly, avoid this prompt when you need to resolve or reconcile the conflict—this prompt explains the disagreement without taking a position on which source is correct. For resolution, use the Conflict Resolution Decision Prompt for AI Assistants or the Evidence Reconciliation Prompt for Analyst Reports. Finally, never use this prompt without human review in regulated or high-risk domains such as healthcare, legal, or financial compliance, where a poorly framed explanation of source conflict could imply false equivalence between authoritative and unreliable sources. Always pair this prompt with source reliability metadata and a human-in-the-loop review step when the explanation will be shown to end users making consequential decisions.

This prompt template generates a structured explanation of a contradiction between two source passages. It is designed for transparency and explainability features in RAG applications, research assistants, and decision-support systems. The model is instructed to ground every part of its explanation in the provided source content, avoiding speculation about author intent, methodology, or external factors not explicitly mentioned in the passages. Use this template when you need to show users why sources disagree, not just that they disagree.

textCopy
You are an evidence analyst. Your task is to explain a contradiction between two source passages. You must ground every statement in the provided text. Do not speculate about author intent, methodology, or external factors unless they are explicitly stated in the passages.

## SOURCE A
[SOURCE_A_TEXT]

## SOURCE B
[SOURCE_B_TEXT]

## CONTRADICTED CLAIM
[CONTRADICTED_CLAIM]

## INSTRUCTIONS
1. Identify what Source A claims about the contradicted claim.
2. Identify what Source B claims about the contradicted claim.
3. Explain the nature of the disagreement. Classify it as one of: factual contradiction, scope mismatch, temporal change, definitional difference, or methodological difference.
4. Identify any textual evidence within the passages that might explain the disagreement (e.g., different date ranges, populations, definitions).
5. State what remains unresolved and what a reader should verify independently.

## CONSTRAINTS
- Do not invent information not present in the passages.
- If the passages do not provide enough context to classify the disagreement confidently, state that explicitly.
- Use direct quotes from the passages to support each point.
- If the passages agree more than they disagree, acknowledge areas of agreement before explaining the conflict.

## OUTPUT FORMAT
Return a JSON object with the following schema:
{
  "source_a_position": "string (Source A's claim, with supporting quote)",
  "source_b_position": "string (Source B's claim, with supporting quote)",
  "disagreement_type": "factual_contradiction | scope_mismatch | temporal_change | definitional_difference | methodological_difference | uncertain",
  "possible_explanations": ["string (explanation grounded in passage text)"],
  "areas_of_agreement": "string or null",
  "unresolved_questions": ["string (what a reader should verify)"],
  "confidence": "high | medium | low (how confidently the disagreement can be characterized from the passages alone)"
}

Adaptation guidance: Replace [SOURCE_A_TEXT] and [SOURCE_B_TEXT] with the full text of the conflicting passages. Replace [CONTRADICTED_CLAIM] with a concise statement of the specific claim in dispute. For multi-source scenarios, extend the template with additional source blocks. In high-stakes domains such as healthcare or legal review, add a [RISK_LEVEL] parameter and a human-review gate before the explanation reaches end users. Validate the output JSON against the schema before surfacing it in your application. If the model returns "disagreement_type": "uncertain" or "confidence": "low", route the output for human review rather than presenting it as a definitive explanation.

The Contradiction Explanation Generation Prompt is not a standalone widget; it is a component in a larger evidence pipeline. Before this prompt fires, your system must have already detected a conflict between at least two source passages and extracted the specific contradictory claims. The prompt's job is to produce a human-readable explanation grounded in those source texts, not to discover the conflict itself. Wire it downstream of a conflict detection step (such as the Multi-Source Conflict Detection Prompt Template or Conflicting Passage Pair Identification Prompt) so that the explanation generator receives clean, verified input: the claim in dispute, the two conflicting source passages, and any available source metadata (publication date, author, methodology notes).

Build a validation wrapper around the model call that enforces grounding before the explanation reaches any user. The validator should check that every factual assertion in the generated explanation can be traced back to the provided source passages. A practical approach: extract all declarative sentences from the explanation, embed them, and run a cosine similarity check against sentence embeddings of the source passages. Flag any sentence below a similarity threshold for human review. Additionally, scan for speculative language markers ('might have,' 'could be,' 'perhaps') and cross-reference them against the source metadata—if the prompt speculates about a source's methodology without that information being present in the input, block the output. For high-stakes domains (legal, medical, financial), route all explanations through a human review queue before publication, with the validator's grounding scores attached to help reviewers prioritize.

Model choice matters for this task. The prompt requires careful reasoning about source content without introducing hallucinated context. Use a model with strong instruction-following and low hallucination rates on grounded summarization tasks—frontier models from major providers are appropriate here. Set temperature low (0.0–0.2) to reduce creative drift. Implement a retry strategy with escalating interventions: on first validation failure, resubmit with the validator's specific grounding failures appended as additional constraints. On second failure, escalate to a human reviewer with both the original sources and the failed explanation attempts. Log every generation attempt, validation result, and human decision for audit and prompt improvement. Never silently fall back to an unvalidated explanation.

The output schema should be strict. Require the model to return structured JSON with clearly separated fields: the contradictory claim, source A's position with citation, source B's position with citation, the explanation of factors that might account for the disagreement, and a confidence indicator. Parse this JSON in your application layer before rendering. If the model returns malformed JSON, use a repair prompt (see Output Repair and Validation Prompts pillar) rather than accepting a degraded text response. The structured output lets you render explanations differently for different surfaces—a concise inline tooltip versus an expanded audit view—while keeping the underlying data consistent and testable.

Test this prompt against a curated set of known contradictions with ground-truth explanations written by domain experts. Measure explanation faithfulness (does the explanation only use information present in the sources?), completeness (does it address all salient differences?), and neutrality (does it favor one source without evidence-based justification?). Run regression tests whenever you update the prompt, the model version, or the upstream conflict detection logic. A silent degradation in explanation quality can erode user trust faster than a refused answer.