An AI answer may include a citation and still be wrong. The cited passage might discuss the same topic without supporting the specific claim. Evidence verification adds a separate step that checks whether the retrieved text truly justifies the generated statement.
Relevance is not the same as support
A search result can be highly relevant because it contains similar terms. Support is stricter: the evidence must agree with the claim, apply to the correct context, and avoid contradiction. This distinction is especially important in technical and medical domains.
Verification as a classification task
A verifier can receive the claim, the case context, and the retrieved evidence, then predict support or non-support. Strong negative examples are essential. They can include wrong states, related but insufficient passages, contradictory evidence, and text from a different case.
Use the score to control behavior
A low verification score can trigger another retrieval attempt, request more context, show uncertainty, or prevent the answer from being presented as fact. This creates a practical safety layer between retrieval and final generation.
Test difficult cases
Random negatives are often too easy. Evaluation should include evidence that looks plausible but fails in a subtle way. Swapped evidence and closely related non-support examples reveal whether the verifier learned genuine support or only topic similarity.
Reliability requires multiple checks
No single metric or component guarantees correctness. Useful systems combine retrieval evaluation, evidence verification, answer grounding, uncertainty handling, and human review for high-impact decisions.