Contradiction Detection

Contradiction detection targets two primary failure modes:

• Internal Contradiction: The model generates statements within its own output that conflict with each other (e.g., 'The event is scheduled for Monday... It will take place on Tuesday.').
• External Contradiction: The model's output conflicts with information provided in the source context or with established, verifiable facts. This is the primary defense against hallucination when using Retrieval-Augmented Generation (RAG) architectures. The prompt must specify which type of check to perform, as the verification logic differs.

Effective prompts are not implicit; they use clear, imperative language. A standard structure includes:

1. Task Directive: 'Identify any contradictory statements.'
2. Scope Definition: 'Check between your answer and the provided source.'
3. Resolution Mandate: 'If a contradiction is found, revise your answer to be consistent.'
4. Output Format: 'Present the original contradiction and your corrected statement.' This moves the model from generative mode into a self-verification and error correction loop, a form of Recursive Error Correction applied within a single inference call.

Contradiction detection is rarely a standalone instruction. It is a key component within larger fact-checking loop architectures. A common pattern is:

• Step 1: Generate an initial response.
• Step 2: Instruct the model to extract all factual claims from its response.
• Step 3: Perform contradiction detection between these claims and the source.
• Step 4: Revise or flag the output. This stepwise verification process decomposes the complex task of ensuring factual fidelity into manageable, instructed sub-tasks, making the model's reasoning more transparent and reliable.

The technique's efficacy is fundamentally tied to source attribution. To detect an external contradiction, the model must know what to check against. Prompts must therefore:

• Anchor the response to specific provided documents (contextual anchoring).
• Mandate the use of citations or references.
• Enforce a no fabrication rule, stating all information must derive from sources. Without clear attribution, the model lacks the ground truth needed to identify a contradiction, rendering the instruction ineffective. This makes it a cornerstone of source-based generation methodologies.

To be useful for observability and trust, the output of contradiction detection must be structured. Prompts should enforce a deterministic output format, such as:

codeCopy
- Claim: [The original contradictory claim]
- Source Conflict: [The conflicting information from the source]
- Resolution: [The corrected, consistent statement]

This structured output serves as an audit trail, allowing developers and AI Governance teams to verify the model's self-correction process. It transforms the prompt from a black-box request into a transparent algorithmic explainability tool.

The technique has inherent limitations:

• Failure to Detect: Models may lack the reasoning depth to identify subtle logical contradictions.
• Over-Correction: It may incorrectly 'correct' valid statements due to misinterpretation.
• Confidence Misalignment: A model might be highly confident in its initial, contradictory statement. Therefore, contradiction detection prompts often work best when combined with calibration prompts that adjust the model's confidence and instructions for uncertainty acknowledgment. It is a powerful guardrail, not an infallible solution, and must be part of a broader hallucination mitigation strategy.

This prompt instructs the model to review its own generated text for logical contradictions. It is a self-verification step applied before finalizing an output.

Example Prompt: "First, draft a response to the user's query. Then, in a second step labeled 'Consistency Review,' systematically check your draft for any statements that contradict each other. List each contradiction found. Finally, produce a revised response that resolves all identified issues."

Key Mechanism: Forces the model to adopt a dual role—generator and critic—activating different reasoning pathways to catch errors its initial pass may have missed.

This pattern directs the model to compare its proposed answer against provided source documents, flagging any points of disagreement. It enforces source-based generation.

Example Prompt: "You are given the following source text and a question. Provide an answer based solely on the source. After your answer, include a section titled 'Alignment Check.' In this section, quote the exact sentences from the source that support each of your answer's key claims. If any claim lacks direct support, note 'Unsupported' and revise your answer to remove it."

Use Case: Critical for Retrieval-Augmented Generation (RAG) systems to ensure the model does not ignore or contradict retrieved evidence.

Used when multiple source documents are provided, this prompt guides the model to synthesize information while explicitly identifying and reconciling discrepancies between sources. It is a form of multi-source synthesis with a contradiction focus.

Example Prompt: "You have been provided with three documents on the same topic. Your task is to write a unified summary. First, create a table with two columns: 'Key Fact' and 'Source Agreement.' For each fact, note if all sources agree, or list the conflicting statements. Then, write a summary that addresses these conflicts, using phrasing such as 'Source A states X, while Source B reports Y.'"

Outcome: Produces a transparent output that acknowledges uncertainty or conflicting evidence rather than presenting a single, potentially fabricated consensus.

This architecture decomposes contradiction detection into a strict, structured verification pipeline. The model must output its reasoning in a predefined format that makes the check auditable.

Example Prompt: "Follow these steps precisely:

1. Generate Claims: List the 5 main factual claims in your answer.
2. Provide Evidence: For each claim, cite the exact sentence from the provided context that supports it. Use quotation marks.
3. Flag Contradictions: Review the evidence. If any two claims are supported by evidence that logically contradicts, state 'CONTRADICTION: [Claim A] vs [Claim B]'.
4. Output Final Answer: Provide a corrected answer that resolves any contradictions from step 3."

Benefit: The enforced structure minimizes the chance the model will skip or gloss over the verification step.

This prompt specifically targets contradictions involving dates, sequences, or statistics. It uses temporal bounding and explicit rules to catch impossible timelines or conflicting numbers.

Example Prompt: "When describing the project timeline, ensure all dates and durations are logically consistent. After your description, include a 'Sanity Check': verify that start dates are before end dates, that phase durations add up correctly, and that no event is said to occur in two different years. If you find an inconsistency, correct it before submitting your final answer."

Application: Highly effective for generating reports, historical summaries, or financial projections where numerical consistency is paramount.

This prompt establishes an absolute rule (a no fabrication rule) that triggers a specific failure state if a contradiction is detected, rather than asking for a revision. It is useful for high-stakes, automated pipelines.

Example Prompt: "Analyze the following two statements. Determine if they are logically contradictory. Your output must be ONLY a valid JSON object: {"contradiction_detected": boolean, "reasoning": "string"}. If contradiction_detected is true, you must NOT proceed to answer the user's question. Only if it is false should you provide the final answer."

Engineering Value: Turns contradiction detection into a deterministic guardrail that can programmatically halt a pipeline, preventing the delivery of flawed information.

A grounding prompt is an instruction that explicitly requires a language model to base its response on provided source material, verifiable facts, or a specific knowledge base to prevent fabrication. It acts as the foundational constraint for factual output.

• Mechanism: Directs the model to treat the provided context as the sole authoritative source.
• Example Instruction: "Answer the question using only the information provided in the following document. Do not use any prior knowledge."
• Relationship to Contradiction Detection: Grounding provides the source material against which contradictions are later identified. It establishes the 'ground truth' for the verification process.

A factual consistency check is a prompt instruction that directs a model to verify that all statements in its output are internally consistent and align with established facts or provided context. It is a broader evaluation that includes, but is not limited to, contradiction detection.

• Scope: Can check for consistency against external knowledge, temporal logic, and commonsense reasoning, not just internal contradictions.
• Example Instruction: "Review your generated summary. List any factual claims that cannot be directly supported by the source text or that conflict with known historical dates."
• Key Difference: While contradiction detection specifically flags conflicts, a factual consistency check may also flag unsupported or anachronistic statements that aren't directly contradictory.

A self-verification prompt is an instruction that guides a model to act as its own critic, systematically checking its initial response for errors, inconsistencies, or unsupported claims. This creates an internal feedback loop for quality control.

• Process: Typically involves a multi-step prompt where the model first generates a draft, then switches to a 'critic' role to evaluate it.
• Example Instruction: "You are now a fact-checker. Review the previous answer. Identify any statements that are speculative, lack citation, or contradict the source. Provide a revised version."
• Implementation: Contradiction detection is often a core subroutine within a self-verification prompt, used to find conflicts during the critique phase.

A cross-reference instruction is a prompt directive that requires a model to compare information across multiple provided sources to establish consensus and identify discrepancies before responding. It is a proactive method for contradiction detection during the information synthesis phase.

• Use Case: Essential for tasks like summarizing multiple documents, legal discovery, or research synthesis.
• Example Instruction: "You have been provided with three reports on the same event. Before answering, compare the key facts (dates, figures, names) across all sources. Note any points of disagreement."
• Output: The model's response should explicitly acknowledge and resolve conflicts, or present them neutrally if resolution is impossible.

Structured verification is a prompt pattern that forces a model to output its fact-checking process in a predefined format, such as a table of claims and supporting evidence. This makes the verification logic explicit, auditable, and easier to parse programmatically.

• Format Enforcement: Often uses markdown tables, JSON, or XML to structure the output.
• Example Instruction: "Generate your answer. Then, produce a verification table with two columns: 'Claim' and 'Supporting Evidence from Source'. If a claim has no evidence or contradicts evidence, mark it as 'UNVERIFIED'."
• Advantage: This pattern operationalizes contradiction detection, turning a qualitative check into a structured, machine-readable output that can be integrated into automated pipelines.

The no fabrication rule is an absolute prompt prohibition that explicitly instructs the model not to invent details, quotes, data, or citations that are not present in the provided context. It is the most direct guardrail against hallucination.

• Nature: A non-negotiable constraint, often stated in strong, imperative language.
• Example Instruction: "Do not add any details, examples, or numerical data that are not explicitly mentioned in the provided text. If the information is not present, state 'Not specified in the source.'"
• Foundation for Detection: This rule establishes the baseline expectation. Contradiction detection then becomes the mechanism for enforcing this rule by identifying when the model's output has strayed from (and thus contradicted) the source-limited context.