Bias-Audited Resume Screening with Calibrated Matching Scores, Selective Review, and Evidence-Grounded LLM Explanation Cards

Abstract

Automated resume screening requires a valid target, calibrated uncertainty, a defensible review policy, and explanations anchored to candidate evidence. This study evaluates those requirements on the datasetmaster resume collection. The 4,817-record snapshot contains a heterogeneous block of 217 records and a standardized block of 4,600 records balanced across 46 technical roles. Because the collection has neither job descriptions nor hiring outcomes, the target is internal proxy role retrieval, not employability or candidate–vacancy fit. A leakage audit removed identity fields, summaries, job titles, education majors, and project prose. The standardized cohort was divided into 2,760 training, 920 calibration, and 920 test profiles. The final temperature-scaled linear support vector machine obtained 0.2489 accuracy, 0.2325 macro-F1, 0.3978 top-3 accuracy, and 0.0401 expected calibration error. A logistic ablation rose from 0.2348 accuracy on strict qualifications to 0.9076 with project descriptions and 1.0000 with direct role text, demonstrating severe leakage. At the 80%-coverage threshold, test coverage was 0.8130 and accepted accuracy was 0.2955. Accuracy fell to 0.1346 on 104 mapped heterogeneous profiles. A lexical support gate reviewed 99.08% of all 217 heterogeneous records while retaining 94.35% of core records. All 920 structured cards passed source checks; an LLM then verbalized one card per role, and all 46 summaries preserved the role, route, safety constraints, and at least one exact evidence span. The results support human-supervised role triage, not autonomous hiring.