Self-Verification Prompt

This mechanism breaks the verification task into a mandatory, sequential checklist. The model is instructed to first generate a response, then execute a series of discrete verification steps before finalizing its output.

Key steps often include:

• Claim Extraction: Isolate all factual assertions from the draft response.
• Evidence Sourcing: For each claim, locate supporting text or data within the provided context.
• Consistency Check: Ensure no claims contradict each other or the source material.
• Plausibility Assessment: Flag statements that violate basic real-world logic or domain-specific principles.

This structured approach replaces a vague "check your work" instruction with a deterministic, auditable process.

To make the verification process transparent and machine-readable, the prompt mandates the model to output its critique in a predefined, structured format. This transforms an internal reasoning step into an external artifact.

Common enforced formats include:

• Verification Tables: Outputting a table with columns for 'Claim', 'Supporting Evidence', 'Confidence', and 'Flag'.
• Annotated Responses: Interleaving the final answer with inline citations or bracketed evidence tags.
• Separate Audit Log: Producing a distinct JSON or YAML object detailing the verification steps, findings, and any revisions made.

This mechanism ensures the model's self-critique is explicit, not implicit, allowing for downstream validation by other systems or human reviewers.

This is an iterative mechanism where the prompt creates a rule: every factual claim must be backed by a specific, retrievable piece of evidence from the provided context. The model is instructed to operate in a loop:

1. Generate a statement.
2. Immediately search the context for a direct match or logical support.
3. If evidence is found, the statement is approved.
4. If evidence is not found, the statement must be revised or removed.

This creates a no-fabrication guardrail. It directly implements source-based generation, preventing the model from relying on its parametric memory unless that memory is explicitly activated and cited by the provided context. It is the core of techniques like Retrieval-Augmented Generation (RAG) when enforced via prompt.

This mechanism instructs the model to actively seek out and resolve conflicts, both internal and external. It moves beyond checking for support to checking for logical coherence.

The prompt typically directs two types of checks:

• Internal Consistency: Are any two statements in the model's own output mutually exclusive? (e.g., "The event occurred in January" and "The event occurred in Q2").
• External Consistency: Do any statements contradict the explicitly provided source documents or known facts defined in the prompt's knowledge cutoff?

Upon detection, the model is instructed to resolve the contradiction by deferring to the higher-authority source (usually the provided context), marking the conflicting claim as uncertain, or seeking clarification. This is a key component of factual consistency checks and multi-source synthesis.

This mechanism addresses the model's tendency toward overconfidence. The prompt instructs the model to assign and act upon a confidence level for its information, tying assertion to certainty.

Common instructions include:

• Implement a Confidence Threshold: "Only state information if your confidence exceeds 90%. Otherwise, say 'I cannot verify this.'"
• Use Uncertainty Tokens: "Preface any claim with [HIGH_CONFIDENCE], [MEDIUM_CONFIDENCE], or [LOW_CONFIDENCE] based on the clarity of evidence."
• Mandate Uncertainty Acknowledgment: "If evidence is partial or ambiguous, explicitly state the limitations of your knowledge."

This moves the model from generating binary right/wrong answers to producing calibrated outputs that signal their own reliability, a critical feature for high-stakes applications.

This mechanism strictly limits the domain of the response to prevent off-topic fabrication or extrapolation. The prompt acts as a tether, binding the model's output to a defined space.

It is achieved through explicit instructions like:

• Temporal Bounding: "Your knowledge is limited to events before 2023. Do not reference later events."
• Source Limitation: "Base your answer only on the provided document excerpts. Do not use general knowledge."
• Domain Constraint: "Restrict your analysis to the financial metrics provided. Do not speculate on market sentiment."

This bounded generation technique reduces hallucination by shrinking the solution space the model is allowed to explore. It works in tandem with knowledge cutoff declarations to create a precise operational sandbox for the model.

A grounding prompt explicitly requires a language model to base its response solely on provided source material, verifiable facts, or a specific knowledge base. This technique directly prevents fabrication by tethering the model's output to an external reference.

• Core Mechanism: The instruction acts as a constraint, limiting the model's generative space to the content of the provided context.
• Example Instruction: "Answer the following question using only the information provided in the document below. Do not use any prior knowledge."
• Primary Use: Essential for Retrieval-Augmented Generation (RAG) systems, legal document analysis, and customer support bots using knowledge bases.

A fact-checking loop is a multi-step prompt architecture that instructs a model to iteratively generate a response, then critique and revise it for factual accuracy. It formalizes the self-verification process into a repeatable cycle.

• Standard Pattern: Often follows a Generate → Critique → Revise sequence.
• Implementation: The critique step uses a separate verification prompt (e.g., "Identify any unsupported claims in the above text").
• Advantage: Breaks down the complex task of simultaneous generation and verification, leading to more thorough error detection than a single-step prompt.

A source attribution instruction is a prompt directive that mandates the model to cite the specific documents, data points, or references supporting each factual claim. This creates an audit trail for the output.

• Enforces Transparency: Makes the model's evidence explicit, allowing for human verification.
• Common Formats: Instructions can require inline citations (e.g., [Doc1]), footnotes, or a bibliography.
• Critical for: Academic writing assistants, research synthesis tools, and any application where provenance and avoiding plagiarism are paramount.

Structured verification is a prompt pattern that forces a model to output its fact-checking process in a predefined, machine-readable format. This moves verification from an implicit process to an explicit, inspectable output.

• Typical Output Format: A table with columns for Claim, Supporting Evidence, Source Location, and Verification Status.
• Benefit: The structured output can be programmatically parsed and validated by downstream systems, enabling automated quality checks.
• Use Case: High-stakes content generation pipelines where every claim must be logged and verified.

A confidence threshold is a prompt parameter instructing a model to only state information if its internal certainty exceeds a specified level. A calibration prompt is designed to improve the model's own accuracy in estimating this certainty.

• Threshold Instruction: "If you are less than 90% confident in the answer, say 'I cannot answer with sufficient confidence.'"
• Calibration Goal: Aligns the model's expressed confidence with the actual likelihood of correctness, reducing overconfident hallucinations.
• Application: Critical in medical or financial Q&A systems, where expressing uncertainty is safer than a confident but wrong answer.

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 provided source material. It is a key sub-task of self-verification.

• Process: The model is asked to list all claims, compare them for logical consistency, and flag any pairs that contradict.
• Instruction Example: "Review the following summary. Identify any two sentences that contradict each other. Explain the contradiction."
• Importance: Catches a common failure mode where a model generates plausible-sounding but mutually exclusive statements.