Evidence-Grounded Vision–Language Reporting for Underground Cable Inspection via Claim Verification and Preference Adaptation

Underground cable inspection requires reliable reports that describe visible defects while avoiding unsupported engineering conclusions and omissions. Existing detector-based workflows provide localized observations but lack integrated verification of heterogeneous claims and explicit control of report completeness. We hypothesize that claim-type-aware evidence verification, dependency reasoning, and coverage-aware preference adaptation can jointly improve reporting accuracy and completeness. Accordingly, we propose an evidence-grounded vision–language framework that verifies visual and engineering claims using detector evidence and source-traceable technical knowledge, propagates upstream uncertainty, and constructs reliability–coverage-guided preferences for parameter-efficient adaptation. On a simulated dataset containing 528 images and 1025 annotated defect instances, the proposed method improves Claim Macro-F1 from 0.548 (ordinary DPO) to 0.743, reduces the unsupported claim rate from 0.196 to 0.153, and decreases detector-referenced Defect Omission from 0.311 to 0.116. The Claim Macro-F1 difference from deterministic detector-and-rule-based reporting is not statistically significant. These findings support evidence-guided preliminary inspection reporting within the evaluated setting, while external validation in operating cable tunnels remains necessary.