AI Pattern · Reliability
Verification Pattern
Check the claims before responding.
Verification adds a fact-checking step before the answer is final. The system extracts the claims an answer makes and verifies each one, against retrieved sources, a recalculation, or an explicit check, then corrects or flags anything it cannot support before responding.
How it works
- 1Draft an answer
- 2Extract its factual claims
- 3Verify each against sources, a calculation, or a check
- 4Correct or drop unsupported claims, then respond
Strengths
- Cuts hallucinations on factual answers
- Produces citations and grounded confidence
- Catches arithmetic and logic slips
Watch-outs
- Extra cost and latency per claim
- Only as good as the sources it checks against
- Can still pass a subtly wrong claim
When to use it
High-stakes or factual answers, research, anything cited, numbers that matter. Overkill for casual conversation where being wrong is cheap.
Example prompt
For each factual claim in the draft, mark it Supported or Unsupported using only the provided sources. Then rewrite the answer to drop or correct every unsupported claim. Sources: <docs> Draft: <text>