LLM-Style Evidence Cards for Scientific Search Interfaces: A UI/UX Design Framework for Retrieval Transparency, Ranking Trust, and Visual Evidence Hierarchy

Abstract

Scientific search systems increasingly provide ranked documents, passages, and automated summaries, yet users must still determine whether retrieved evidence supports a claim, whether the presented information is sufficient for inspection, and why a result is highly ranked. This paper proposes an evidence-card UI/UX framework for scientific search that transforms retrieved articles into structured evidence cards containing a claim anchor, extractive evidence summary, support/refute/insufficient badge, confidence cue, citation cue, ranking rationale, and expandable source text. The framework is designed as a visual communication layer for retrieval transparency and evidence inspection rather than as a new claim-verification model or a live LLM system. Evaluation was conducted using the BEIR SciFact retrieval benchmark, the original SciFact train/dev datasets, and a SciFact-Open candidate-pool stress test. On the 300-query BEIR SciFact test set, the BM25-dominant hybrid baseline achieved nDCG@10 = 0.6667 and Recall@10 = 0.7858, while the proposed evidence-card pipeline achieved nDCG@10 = 0.6621 and Recall@10 = 0.7763. On the SciFact dev set, gold evidence appeared within the top three evidence-card candidates for 84.6% of evidence-bearing claims, and selected rationale sentences matched gold rationale annotations for 44.0% of gold evidence-document pairs. Interface-level analysis showed that the proposed card design increased the evidence visibility index from 0.2643 to 0.5920 and reduced the estimated first-pass scan-burden proxy from 115.50 to 84.08 seconds. These results suggest that evidence cards improve transparency by making relevance, uncertainty, confidence, and ranking rationale visible while preserving access to source evidence.