Factual Consistency Check

This is the foundational instruction that mandates the model to support every factual assertion with specific data, quotes, or references from the provided context. It transforms generation from an open-ended task into a source-based generation exercise.

• Mechanism: The prompt explicitly states that unsupported claims are prohibited.
• Example Instruction: "For every factual statement you make, you must cite the exact sentence from the provided document that supports it using [bracketed numbers]. Do not make any claims without a citation."

This mechanism instructs the model to perform self-verification by identifying and resolving conflicting statements. It operationalizes the fact-checking loop within a single prompt or a chained sequence.

• Mechanism: The model is told to compare its output against the source context line-by-line to flag inconsistencies.
• Process: Often structured as: 1. Generate a draft answer. 2. Extract all factual claims. 3. For each claim, verify against source. 4. Revise or flag unverifiable claims.
• Output Format: Frequently requires a structured verification table listing claims, evidence, and a consistency flag.

This technique strictly limits the model's response scope to the provided context, preventing extrapolation. It combines temporal bounding (e.g., "only use data from 2023") and domain constraints.

• Mechanism: The prompt explicitly defines the boundaries of permissible information, acting as a hallucination guardrail.
• Key Instructions: "Your knowledge is strictly limited to the document provided above." "Do not use any prior knowledge or information outside the provided text." "If the answer is not in the document, state 'Not provided in context.'"
• Effect: Forces deterministic output highly reproducible from the same input context.

This mechanism enforces a strict, machine-readable output format that separates claims from evidence, making the verification process explicit and auditable. It is a key enabler of algorithmic explainability.

• Common Formats: JSON, XML, or markdown tables with predefined keys like claim, supporting_quote, source_page, is_verified.
• Purpose: Structures the self-verification prompt into a reproducible pipeline. Allows downstream systems to parse and validate the model's own fact-check.
• Example: {"claim": "The treaty was signed in 1992.", "evidence": "As stated on page 4: 'The signing ceremony took place in Rio de Janeiro in June 1992.'", "consistent": true}

This mechanism programs the model to express calibrated uncertainty rather than guess. It uses a confidence threshold instruction to prevent low-probability fabrications.

• Mechanism: The prompt instructs the model to only state information if its internal certainty is high, otherwise to use specific hedging language or decline.
• Calibration Prompt Example: "If you are less than 90% confident about a fact based on the provided sources, state 'The evidence is unclear on this point' instead of providing a likely answer."
• Link to Sibling: Works in tandem with a knowledge cutoff instruction to manage temporal uncertainty.

For prompts with multiple context documents, this mechanism instructs the model to integrate information and resolve discrepancies, ensuring factual fidelity across a corpus.

• Mechanism: The model is directed to compare information across sources, identify consensus, and note conflicts before generating a final, consistent answer.
• Instruction Example: "Synthesize an answer from Documents A, B, and C. Where sources agree, state the fact. Where they conflict, note the discrepancy and cite the differing sources."
• Advanced Use: Forms the core of retrieval-augmented prompt architectures for complex, evidence-based Q&A.

A grounding prompt is an explicit instruction that requires a model to base its entire response on provided source material, such as a document, database, or specific knowledge base. This technique acts as a primary defense against fabrication by tethering the model's output to a concrete reference.

• Core Mechanism: The prompt explicitly states that answers must be derived only from the supplied context.
• Example Instruction: "Using only the information provided in the following report, answer the question. Do not use any prior knowledge."
• Key Benefit: It creates a clear boundary between the model's internal knowledge (which may be outdated or incorrect) and the authoritative source you provide.

A source attribution instruction mandates that a model cite the exact documents, data points, or passages that support each factual claim in its response. This transforms the output from an assertion into a verifiable argument.

• Enables Auditability: Each claim can be traced back to its origin, allowing for human verification.
• Common Formats: Instructions can specify formats like inline brackets (e.g., [Doc A, p.3]), footnotes, or APA citations.
• Secondary Effect: The requirement to find and cite evidence often naturally discourages the model from inventing unsupported information.

A self-verification prompt instructs the model to act as its own critic, systematically checking its initial draft for errors, inconsistencies, or unsupported claims before producing a final answer. This implements a basic form of internal review.

• Typical Structure: The prompt chains two steps: 1) Generate an answer, 2) Critique that answer for factual accuracy against the provided context.
• Architecture Pattern: This is a foundational element of more complex fact-checking loops and ReAct frameworks.
• Use Case: Highly effective for complex reasoning tasks where a single pass is prone to oversight.

Contradiction detection is a specific instruction that directs a model to identify and resolve conflicting statements either within its own output or between its output and the provided source material. It enforces logical coherence.

• Internal vs. External: Checks for self-contradiction within the generated text and for contradictions with the source context.
• Prompt Example: "Review the following summary. List any statements that contradict each other or contradict the source document. Then, produce a revised version that resolves these conflicts."
• Relation to RAG: A critical component in Retrieval-Augmented Generation Architectures to handle conflicting information from multiple retrieved documents.

Uncertainty acknowledgment is a prompt instruction that trains a model to explicitly state when it lacks sufficient information or is unsure about a fact, rather than guessing. This replaces fabrication with transparency.

• Mitigates 'Confident' Hallucinations: Directly counters the model's tendency to generate plausible-sounding but incorrect information with high confidence.
• Linked to Calibration: Works in tandem with calibration prompts to better align the model's stated certainty with actual accuracy.
• Enterprise Value: Essential for high-stakes domains like healthcare or finance, where knowing the limits of knowledge is as important as the answer itself.

The no fabrication rule is an absolute, non-negotiable prohibition within a prompt that explicitly instructs the model not to invent details, quotes, data, or citations absent from the provided context. It sets a zero-tolerance policy.

• Absolute Constraint: Often phrased as a command: "Do not make up any information. If the answer is not in the context, say so."
• Foundation for Trust: This rule is the bedrock of source-based generation and is a key hallucination guardrail.
• Implementation Note: It is most effective when combined with a grounding prompt that clearly defines the allowable source material.