Answers are grounded in this paper and cite their sources.
Context: Open-source vulnerability auditing increasingly combines static analyzers, rule-based scanners, and learning-based detectors, but triage workflows often treat raw detector scores as if they were reliable probabilities. Objective: This paper studies how security teams can turn heterogeneous detector outputs into auditable decisions that account for uncertainty, review cost, and unsafe-ignore risk. Methods: We propose a calibrated and cost-sensitive governance framework that separates detector scoring from final action selection and explicitly models three actions: Fix, Review, and Ignore. The main experiment evaluates 2740 registered OWASP Benchmark Java prediction rows with independent train, calibration, and test splits, using outputs from CodeBERT, CodeQL, Semgrep, and SpotBugs/FindSecBugs. Results: The registered CodeBERT executor collapses under the fixed 0.5 threshold because its scores are concentrated around 0.997. A calibration-split threshold produces TP/FP/TN/FN = 139/124/61/69 on the test split, with F1 = 0.5902, AUROC = 0.5256, and AUPRC = 0.5455. Under the registered baseline protocol, Semgrep outperforms the rectified CodeBERT output. Calibration reduces overconfidence, hybrid review-oriented policies eliminate unsafe auto-ignore on the test split by routing nearly all samples to review, and a review-budget simulation shows that increasing review capacity from 10% to 100% reduces false-positive fixes from 167 to 0 under the selected rule-union policy. A Vul4J real-project study further executes three Java project cases, producing Semgrep JSON, SpotBugs XML, compiled CodeQL SARIF artifacts, and official target-test evidence for the selected vulnerable and fixed revisions. Conclusion: The results indicate that calibrated and cost-sensitive governance provides a practical decision layer for vulnerability auditing by improving auditability, reducing unsafe automatic dismissal under review-aware policies, and making review workload and safety-cost trade-offs explicit.
No takes yet. Share an insight, caveat, or question.
Feng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: