Sign InOpen Brain
arXivPaperNeeds Review

Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering

Claim-level quotes can look verifiable while failing to support the full claim. Builders of retrieval agents should score citation coverage, verbatim accuracy, and entailment separately.

arXiv · Sep 14, 2026
Open Source Open MarkdownOpen JSON
Source Summary

A harness evaluates **12 LLMs** on **222 synthetic clinical questions** drawn from four practice guidelines. It checks whether each factual claim has a citation, whether the cited text is verbatim, and whether that text fully supports the claim.

Practical Implication

Builders should add claim-level entailment checks instead of treating citation presence or exact copying as proof. Retrieval-agent evaluations need separate measures for coverage, quotation fidelity, and complete substantiation.

Agent-Ready Context
A harness evaluates **12 LLMs** on **222 synthetic clinical questions** drawn from four practice guidelines. It checks whether each factual claim has a citation, whether the cited text is verbatim, and whether that text fully supports the claim.

Builders should add claim-level entailment checks instead of treating citation presence or exact copying as proof. Retrieval-agent evaluations need separate measures for coverage, quotation fidelity, and complete substantiation.

Most models quote source text for over 90% of claims, yet support can remain weak. Claude Opus 5 reaches **98.0% verbatim coverage** but only **37.1% full substantiation**; the clinical, synthetic setup limits direct generalization to other domains.
Connected Context · Feed7 Judgment

This replaces citation presence as a sufficient retrieval-eval signal with a three-part acceptance test: claim coverage, exact quotation, and full support. It reinforces prior warnings that apparently strong evaluators can miss domain-specific failures, while adding a directly inspectable evidence-to-claim check. Its clinical synthetic data supports the evaluation decomposition, not a universal performance estimate.

Domain-Specific Hallucination Detection in Large Language ModelsBoth show that a strong surface evaluation signal can conceal weak biomedical grounding, supporting domain-matched checks rather than reliance on a generic hallucination or citation score.Verifiable Environments for AI in Biology — Kenny Workman, LatchBioThe claim-level support test supplies a deterministic component for scientific evaluation, while the biology evidence explains why expert review is still needed when valid interpretations exceed brittle graders.Beyond Scores: Understanding LLM-as-a-Judge Mechanisms in Summarization EvaluationClaim-level decomposition provides observable failure categories that can help distinguish missed evidence from faulty final integration when auditing an LLM judge.Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought ReasoningBoth caution against mistaking legible text for valid evidence: verbatim quotations may not substantiate claims, just as readable reasoning steps may not cause correct answers.
Context Map
benchmarkresearch#agent-evals#retrieval#benchmark-integrity
Uncertainty
Most models quote source text for over 90% of claims, yet support can remain weak. Claude Opus 5 reaches **98.0% verbatim coverage** but only **37.1% full substantiation**; the clinical, synthetic setup limits direct generalization to other domains.