Verification Before Asserting

Overview

Verification is the discipline of confirming information through authoritative sources before presenting it as fact. For AI agents, this practice prevents hallucination spread and builds lasting user trust.

The Verification Problem

Large language models can generate plausible-sounding but incorrect information—a phenomenon known as hallucination. Without verification:

Verification Methods

1. Cross-Reference Checking

• Compare information across multiple sources
• Prioritize authoritative sources
• Flag contradictions for human review

2. Confidence Calibration

• Express uncertainty explicitly
• Distinguish between facts, probabilities, and speculation

3. Source Attribution

• Always cite where information came from
• Link to primary sources when possible

Practical Verification Rules

1. When uncertain, say so
2. Distinguish knowledge types
3. Provide evidence paths
4. Update when corrected

When to Verify

Always verify: - Factual claims about external systems - Code, formulas, or precise data - Questions about current events - Statistics or percentages

Skip for: - Logical reasoning within known domains - Clarifying user-provided information

See Also

• Guardrails as Autonomy Substrate
• Task Context Switching Protocol