Calibrated Selective Fact-Checking via Evidence Chain Evaluation
06:00 · July 22, 2026 · arXiv cs.AI RSS
Large language models (LLMs) can achieve strong fact-checking accuracy, yet forced binary decisions conceal a critical reliability problem: systems may issue confident verdicts even when supporting evidence is weak, sparse, or internally inconsistent. We address this issue through Evidence Chain Evaluation (ECE), a selective fact-checking framework that permits abstention via an uncertain verdict instead of requiring a true/false decision for every claim. The evaluated system is a tool-using verification agent that gathers evidence through web search, scholarly search, and executable checks, and then returns a structured verdict with confidence and source-level metadata. On ECE-Bench, ECE achieves 91.6% standard accuracy, 93.7% coverage, and 97.8% selective accuracy on answered claims. Although ECE does not outperform the strongest retrieval baseline on aggregate calibration metrics such as Expected Calibration Error, Brier score, or AURC, it delivers a clear selective-prediction trade-off: the system maintains very high accuracy on answered claims while deferring 6 of 95 cases. These deferred cases are concentrated in lower-reliability evidence settings (5/6 at source level L4), supporting the view that abstention functions as a safety-oriented mechanism for handling epistemically weak evidence. Code is available at https://github.com/ cheshireyang/ECE.git
Summary
Large language models have grown capable at fact-checking when augmented with retrieval or tool use, yet they still issue confident true-or-false verdicts even when the supporting evidence is sparse, contradictory, or low-quality. Evidence Chain Evaluation (ECE) addresses this reliability gap by treating fact-checking as a selective-prediction task: the system may return an uncertain verdict and abstain rather than forcing a binary label. The evaluated implementation is a tool-routed verification agent that gathers evidence through web search, scholarly search, and executable code checks before emitting a structured output that includes a verdict, a scalar confidence score, and a source-reliability tag.
On the ECE-Bench benchmark the agent records 91.6 percent standard accuracy while answering 93.7 percent of claims, yielding 97.8 percent accuracy on the answered subset. Six of the ninety-five evaluated claims are deferred; five of those six originate from the lowest source-reliability tier. This distribution indicates that abstention concentrates on epistemically weak inputs, functioning as a safety mechanism that trades modest coverage for substantially higher accuracy on the claims the system does answer.
The paper supplies formal definitions for the relevant metrics. Standard accuracy counts an uncertain verdict as incorrect. Selective accuracy and coverage are computed only over the answered subset, separating the frequency of abstention from the correctness of committed answers. Expected calibration error, Brier score, and area under the risk-coverage curve are also reported, although the strongest retrieval baseline remains superior on these aggregate measures. The accompanying repository contains the evaluation scripts and the VerificationAgent implementation used to produce the reported figures.
Why it matters
This research is highly relevant for Dutch AI researchers and practitioners focusing on trustworthy and ethical AI, a key priority in the Netherlands and the EU. The abstention mechanism directly addresses LLM hallucination and reliability issues, offering actionable methodologies for building compliant, high-stakes verification pipelines under EU AI regulations.
