AERIS-Chart: Evidence-Sufficient And Uncertainty-Calibrated Retrieval For Chart Claim Verification

Chart-grounded fact verification determines whether textual claims are supported by quantitative evidence in visualizations. Building on FactViz3M, we address the limitation that relevant retrieved charts may still lack sufficient evidence for binary verification. We formulate verification as a three-way task: Supported, Refuted, and Not Enough Information (NEI). We propose AERIS-Chart, integrating claim-conditioned multi-candidate retrieval, structured chart-evidence graphs, numerical–relational verification, and evidence-sufficiency estimation. Experiments show that AERIS-Chart achieves the strongest performance, reaching 81.3% accuracy and 79.8 Macro-F1, with 83.5, 81.7, and 74.2 F1 for Supported, Refuted, and NEI, respectively. Compared with multi-chart retrieval, it improves accuracy by 3.5 points and Macro-F1 by 3.7 points, while reducing ECE from 0.083 to 0.061. These results demonstrate that explicitly modeling the evidence chain improves verification accuracy and calibration, particularly when visual evidence is insufficient. The findings highlight calibrated abstention as important for realistic deployment.