SCIENTIFIC EVIDENCE CHARACTERISTICS AND THEIR ASSOCIATION WITH SUPPORTIVE VERSUS REFUTING ATTRIBUTION IN AUTOMATED CLAIM VERIFICATION

This analysis examines whether scientific abstract characteristics distinguish supporting from refuting evidence, addressing the interpretation of document presentation in scientific verification. Using the 2020 SciFact benchmark, it analyses 773 annotated claim document pairs covering 693 claims and 437 abstracts through descriptive comparisons, document weighting and partition sensitivity checks. Refuting pairs contain abstracts averaging 32.46 words longer than supporting pairs. The refuting share is 14.1 percentage points higher among structured abstracts, while mean relative evidence position differs by only 0.007. Equal document weighting preserves these pooled directions, but length and structure differences narrow substantially in the development partition. Furthermore, 179 abstracts support some claims and refute others despite unchanged document characteristics. These findings distinguish document presentation from the semantic relationship between evidence and a specific claim. The contribution is a descriptive account of how pooled presentation patterns coexist with partition sensitivity and claim dependent stance. Evidence services and benchmark developers should preserve claim specific provenance and disclose document overlap and weighting choices when reporting aggregate comparisons.