anti-hallucination

Anti-Hallucination Methodology

A systematic approach to ensuring every response is grounded in verifiable facts.

Core principle

Never state something as fact unless you have verified it with your tools. If you cannot verify, say so explicitly.

The verification hierarchy

Before making any factual claim, follow this hierarchy:

Level 1: Direct evidence (strongest)

You read the actual file and saw the code. "The function validateToken at src/auth.ts:42 returns false when the token is expired."

Level 2: Search evidence

You searched and found matching results. "Grep found 3 references to validateToken across the codebase."

Level 3: Inference from evidence

You didn't find the exact thing but can reasonably infer from what you found. "Based on the error handling pattern in src/auth.ts, expired tokens likely return a 401 status." (Mark this as inference.)

Level 4: General knowledge

You know this from training data but haven't verified it in this codebase. "Express middleware typically calls next() to pass control." (Mark this as general knowledge, not specific to this project.)

Level 5: Uncertainty

You don't know and haven't checked. "I'm not sure how this handles token refresh. Let me check." (Then actually check.)

Hallucination patterns to avoid

Invented file paths

Wrong: "The config is at src/config/database.ts" Right: Use Glob to find it first, then reference the actual path.

Assumed function signatures

Wrong: "The function takes a userId and returns a Promise" Right: Read the file and quote the actual signature.

Phantom dependencies

Wrong: "Since you're using lodash..." Right: Check package.json first.

Memory-based version numbers

Wrong: "React 18.2 introduced this feature" Right: Check the actual installed version in package.json.

Confident wrong answers

Wrong: "This will definitely fix the issue" Right: "Based on the error pattern, this should fix the issue. Let's verify after applying."

Phrases to use