VERA-SE: Verification-Guided Autonomous Software Repair with Risk-Aware Patch Selection

VERA-SE presents a verification-guided autonomous software repair framework that combines large language model (LLM)-based patch generation with structural analysis, behavioral preservation checks, change-risk estimation, and risk-aware patch selection. The framework separates candidate generation from verification, evaluating multiple proposed repairs rather than accepting patches solely because they pass visible tests. The approach is evaluated using two experimental settings. A controlled benchmark comprising six repair tasks and 240 candidate instances compares visible-test-only validation, visible-plus-hidden-test validation, and the full VERA-SE verifier. In this setting, VERA-SE reduced the observed false-acceptance rate to 0%, compared with 74.7% for visible-test-only validation and 27.6% for visible-plus-hidden-test validation, while incorrectly rejecting one oracle-correct candidate. A separate real-world pilot evaluated ten localized defects from the Black Python formatter in BugsInPy. Across 40 requested generations, 28 valid candidate instances were produced, of which only one passed the held-out correctness oracle. In this setting, VERA-SE incorrectly accepted one patch and rejected the sole oracle-correct candidate. The findings demonstrate the potential of task-specific behavioral verification while identifying limitations in transferring generic structural and change-risk heuristics to real-world software defects. The research motivates stronger automated semantic verification methods for reliable LLM-based software repair.