Abstract
Retrieval-augmented verifiers can mistake repeated publication for independent corroboration. We introduce EvidenceGraph, a stagewise framework that measures source lineage, controls repeated evidence, and separates retrieval, aggregation, and calibration decisions. A controlled 40-cluster intervention provides both a diagnosis and a defense. Copying one unsupported source in a fixed 20-slot context raises mean false-claim confidence by 0.339 and 0.327 on two frozen answer backbones. Collapsing passages by recorded origin holds the change below 0.001 and prevents additional false acceptances. A completed two-rater lineage audit reports kappa 0.951 on pair labels and adjusted Rand 0.983 on partitions. Our frozen estimator reaches pair F1 0.967 and cluster ARI 1.000 on constructed evidence, obtains pair F1 0.804 on the audited AVeriTeC subset, and is the strongest tested natural pair classifier overall. EvidenceGraph active selection improves accuracy by 5.7 points and reduces false acceptance by 20.9 points when redundant decoys hide useful evidence. A component study further identifies a diagonal-ridge aggregator with 93.5 percent accuracy and an argmax-preserving calibrator that reduces expected calibration error by 0.058, 0.341, and 0.096 across three natural benchmarks without changing a verdict. These gains also define the operating boundary. Across all natural pools, lineage quality falls to pair F1 0.568 and cluster ARI 0.401, while independence alone does not outperform passive ranking for gold-evidence recall. Our results establish provenance as a measurable control signal and show how to deploy it conservatively. Systems can expose common-origin structure, use redundancy controls in validated regimes, and preserve the verifier's decision boundary when transfer evidence is limited.