When Valid Citations Are Not Enough: Auditing Scientific Claim Support Across Evidence Versions

Abstract Scientific claims can remain unchanged even as their supporting evidence changes, making their continued validity a question of both source provenance and interpretation. To examine how evidence availability and reference judgments shape that assessment, we audited a source-cited verification workflow that records source changes, claim-support states, and update consequences separately. The study combined six constructed numerical pairs with eight unchanged claims from four published papers evaluated against original and deliberately altered result tables. Paired-version inference was compared with separately executed single- version assessments, and frozen reference judgments were checked against the evidence actually supplied to each assessor. We then conducted a separate, post-exposure diagnostic using expanded published context and admissible source roles, with new references fixed before inference. In the numerical demonstration, both approaches detected the single loss of support without additional loss alerts. In the original published-table study, all 24 responses passed citation and input validation, yet some reference judgments relied on context absent from the inference packets. Against the original references, paired and single-version assessment identified two and one of four intended losses, respectively, with state agreement of 10/16 and 11/16. Rescoring the unchanged paired predictions against diagnostic references changed agreement to 12/16. In the expanded diagnostic, both approaches detected all three resolved losses and preserved all three supported controls; agreement with packet-specific AI references was 16/16 and 15/16. Causal and test/validation qualifications remained unresolved. These results show how evidence availability and reference selection affect the interpretation of version-specific claim support, even when source citations and input records are valid.