This is the authors' abstract. We don't add key points for this paper.
Software defect prediction is commonly evaluated as classification, although quality-assurance use also requires a decision about which model outputs need an additional review escalation. We present Cost–Risk–Effort Software Triage for Defect Prediction (CReST-DP), a decision-layer evaluation framework rather than a new defect classifier or reject-option principle. The primary evaluation track (Track A) covers 18 static datasets, 5 seeds, and 5 outer folds (450 settings) using mutually disjoint model-fit, learner/threshold-validation, and policy-validation subsets; all settings completed with zero processing errors and zero selection–policy overlap. The no-routing backbone achieved mean area under the receiver operating characteristic curve (AUC) 0.834 and Matthews correlation coefficient (MCC) 0.388. Inspection left 70.13% of artifacts without an additional triage-triggered review, with accepted error (AE) 0.0842, accepted false-negative exposure (AFNE) 0.2444, and routed-effort fraction 0.3385 under the line-change workload proxy. Matched random routing at the identical coverage yielded AE 0.1359 and AFNE 0.3708; dataset-bootstrap mean differences (Inspection minus random) were -0.0518 for AE and -0.1263 for AFNE. Deterministic single-signal rules can attain lower AE or AFNE but route more of the line-change workload proxy, so Inspection is interpreted as a risk–workload compromise rather than a universally superior rejection rule. Risk–coverage, signal, effort-score, policy-constant, alternative-effort, and uncapped-Camel sensitivities support this bounded interpretation. The secondary evaluation track (Track B), an ApacheJIT semantic-evidence analysis, remains exploratory because its original semantic extraction stage cannot be prospectively reconstructed from the preserved artifacts.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Alsawalqah et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: