No Fabrication Rule

The rule's primary instruction is an explicit, non-negotiable command. It is typically placed at the beginning of a system prompt to establish the highest-priority constraint.

• Key Phrasing: Uses imperative language like "DO NOT invent," "ONLY use provided context," and "If the information is not present, state 'I cannot find that information.'"
• Purpose: It sets a clear behavioral boundary, overriding the model's default tendency to generate plausible-sounding completions. This transforms the model from a creative writer into a precise information retrieval and synthesis engine constrained by its source material.

The rule is ineffective without a clearly defined source of truth. This card explains how the rule explicitly binds the model's output to provided materials.

• Explicit Referencing: The prompt must specify the bounded context, e.g., "based solely on the following document," "using only the provided dataset."
• Prevents Extrapolation: This prevents the model from using its parametric knowledge to fill gaps, ensuring every claim is traceable to the input. It is the technical implementation of source-based generation.

To enforce the rule, prompts often mandate a specific output format that separates claims from evidence, making fabrication easy to spot and verify.

• Evidence-Attached Formats: Requires responses in structures like:
  • Claim: [The statement]. Evidence: [Direct quote from context].
  • A table with columns for Fact, Source Paragraph, and Confidence.
• Enables Automated Checks: This structured output allows for post-hoc validation by other systems or humans, operationalizing the verifiable claim requirement. If a cell in the 'Evidence' column is empty, the rule has been violated.

Advanced implementations embed the rule within a self-verification prompt that instructs the model to critique its own draft output before finalizing it.

• Process: The prompt chains instructions: "1. Generate an answer. 2. Review each sentence. 3. For any sentence not directly supported by the context, flag it and rewrite it to be accurate or remove it."
• Iterative Refinement: This creates an internal fact-checking loop, significantly reducing subtle hallucinations where the model strays from the source. It is a key component of the ReAct (Reasoning and Acting) framework for reliable agents.

A critical part of the rule defines what the model should do when requested information is absent. Without this, models may default to fabrication.

• Explicit Uncertainty: The prompt must provide a clear alternative action, such as: "If the answer is not in the context, respond with 'The provided documents do not contain information on this topic.'"
• Promotes Honesty: This enforced uncertainty acknowledgment is more reliable than relying on a model's innate confidence calibration. It ensures the system fails safely and transparently.

The no fabrication rule is the behavioral cornerstone of Retrieval-Augmented Generation (RAG) architectures and secure tool calling.

• In RAG: The rule is applied to the retrieved context. The prompt instructs: "Answer using ONLY the retrieved documents below." This guarantees the generated answer is grounded in the enterprise knowledge base.
• In Tool Calling: When an agent uses an API (e.g., a database query), the rule constrains the model to report only the data returned by the tool, not to embellish it. This is essential for deterministic output in automated workflows.

A grounding prompt is an instruction that explicitly requires a language model to base its response solely on provided source material, verifiable facts, or a specific knowledge base. It acts as a primary defense against fabrication by tethering the model's output to an authoritative context.

• Mechanism: Instructs the model to use phrases like "Based on the provided document..." or "According to the source..."
• Example: "Answer the following question using only the information contained in the attached research paper. Do not introduce any external knowledge."
• Key Benefit: Creates a clear, auditable link between the source input and the model's output.

A source attribution instruction is a prompt directive that mandates the model to cite the specific documents, data points, or line numbers that support each factual claim in its response. This transforms assertions into verifiable statements.

• Enforces Transparency: Requires inline citations (e.g., [Doc A, Section 2.1]) or a reference list.
• Enables Audit: Allows a human or automated system to quickly cross-check the model's output against the source.
• Common Formats: Instructions can specify APA, MLA, or simple bracket-based citation styles.

An evidence requirement is a strict prompt directive that forbids the model from making any factual assertion without immediately supporting it with specific data, a direct quote, or a reference from the provided context. It operationalizes the "No Fabrication Rule."

• Absolute Constraint: The prompt explicitly states: "For every factual claim you make, you must provide the supporting evidence from the context."
• Structural Output: Often leads to responses formatted as claim-evidence pairs or tables.
• Impact: Drastically reduces unsupported extrapolation and inventive "filling in the blanks."

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 or fabricating a plausible-sounding answer.

• Critical for Safety: Prevents the model from presenting conjecture as fact in high-stakes domains like healthcare or law.
• Standard Phrasing: Instructions include: "If the context does not contain enough information to answer fully, state what is known and explicitly note the gap."
• Builds Trust: Creates more reliable and honest interactions by setting appropriate user expectations.

A fact-checking loop is a multi-step prompt architecture that instructs the model to generate a response, then critique and revise it for factual accuracy in one or more subsequent, instructed steps. It externalizes the verification process.

• Common Pattern: Step 1: Draft an answer. Step 2: Review the draft. List any claims that lack direct support in the context. Step 3: Produce a final answer, removing or correcting unsupported claims.
• Increases Latency: Adds computational steps but significantly improves output fidelity.
• Self-Correction: Leverages the model's ability to critique its own text when given explicit instructions to do so.

Contradiction detection is a prompt instruction that directs a model to identify and resolve conflicting statements within its own output or between its output and the provided source material before finalizing its response.

• Prevents Inconsistency: Catches errors where a model might make two opposing claims.
• Instruction Example: "Before providing your final answer, scan it for any internal contradictions or statements that contradict the provided sources. List and resolve them."
• Foundation for Logic: A basic form of logical consistency checking that is prompted rather than hard-coded.