Abstract
<title>Abstract</title> <p> <bold>Context:</bold> Spectrum-based fault localization ranks executable lines but does not tell a developer where to stop inspecting. <bold>Objective:</bold> We formulate stopping as prediction of the smallest score-tie boundary containing a mapped faulty line and seek bug-specific cutoffs for a user-selected marginal miss risk. <bold>Method:</bold> ConformalFL applies one-sided conformalized quantile regression to pre-fault spectrum features. We evaluate 248 candidate-present bugs out of a pinned universe of 395 historical Defects4J bugs, using 30 project-stratified splits, nominal coverage levels of 80%, 90%, and 95%, fixed inspection policies, and project, chronological, and cross-language stress tests. <bold>Results:</bold> At nominal 90%, ConformalFL attained 91.8% mean coverage while inspecting 30.2% of executable candidates. A constant conformal cutoff attained the same mean coverage at 44.2% inspection. However, preregistered Top-10% inspection was empirically close (90.1% coverage at 28.3% inspection), and 18.9% of repeated cases required full inspection after tie expansion. Cross-language transfer retained high empirical coverage but often lost useful compression. <bold>Conclusions:</bold> ConformalFL does not empirically dominate a well-chosen fixed cutoff on this benchmark. Its contribution is an auditable mapping from a requested miss risk to a tie-complete, bug-adaptive inspection set under calibration–deployment exchangeability; it does not guarantee conditional or shifted-project coverage, improve ranking quality, or measure human debugging time. </p>