When Does Diagnosis Help Select Software Repairs? A Synthetic Sensitivity Study

This preprint presents an exploratory synthetic sensitivity study investigating when diagnostic information helps select software repair candidates. Using a controlled simulator, the study compares six candidate-selection policies across three mappings between diagnosis labels and candidate utility: identity-aligned, one-step rotation, and episode-wise random permutation. The primary outcome is the paired difference in selected gain between an adaptive diagnosis-informed ranking policy and an observed-performance ranking baseline. Under the identity-aligned mapping, the mean difference is +0.00542, with a descriptive 95% seed-level confidence interval of [0.00484, 0.00600]. Under the one-step rotation and episode-wise random mappings, the differences are −0.00412 [−0.00476, −0.00347] and −0.00174 [−0.00245, −0.00103], respectively. These results illustrate that the usefulness of diagnosis-informed candidate selection depends on the relationship between diagnostic labels and candidate utility. Accurate root-cause diagnosis alone does not necessarily identify the most beneficial repair candidate when that relationship is misaligned or unknown. The evidence is entirely synthetic. The study does not evaluate real-world software repairs, repair acceptance, deployment effectiveness, or safety guarantees. The findings should be interpreted as exploratory sensitivity results rather than evidence of real-world repair performance. This manuscript is a preprint and has not been peer reviewed.