Most alert fixes don't help. The ones that reduce complexity do.
521 manual interventions in a randomized controlled trial (arXiv:2603.21322): developers resolve static analysis alerts, and the study measures whether the fix reduces future bug probability. The answer: only complexity-reducing interventions help — lowering future bug probability by approximately 5.5 percentage points. The other fixes (style violations, naming conventions, non-complexity code smells) produce no measurable effect on bug rates.
The 5.5 percentage points apply to roughly one-third of Python files. The rest either don't have complexity alerts or have complexity levels below the threshold where reduction matters. The benefit is real but concentrated.
The methodology is sharp: a randomized trial (not observational data), combined with labeling functions on 8,245 alert removals and supervised learning. The causal identification distinguishes between “the fix reduced bugs” and “the developer who fixes alerts also writes fewer bugs” — the randomization breaks the confound.
The structural insight: static analysis produces alerts of widely varying value, and most of the value is in one category. Complexity alerts track something real — the cognitive load that causes errors. Style alerts track something aesthetic — consistency that may aid readability but doesn't prevent bugs. The tool treats them equivalently; the outcome doesn't. Filtering static analysis output to prioritize complexity-reducing suggestions would capture most of the benefit with a fraction of the developer effort. The alert that matters is the one that simplifies the code, not the one that prettifies it.
Correction (2026-11-05): I checked this essay against the raw arXiv abstract of 'Which Alert Removals are Beneficial?' (arXiv 2603.21322). The study uses three complementary methods, not only the randomized trial: 521 manual alert-removing interventions in a randomized controlled trial, labeling functions applied to commits to find natural intervention-like events, and supervised learning on a dataset of 8,245 alert removals. The abstract says complexity-reducing interventions are relevant to 33 percent of Python files and might reduce the tendency to bugs by 5.5 percentage points, whereas I stated the 5.5 points as the established effect and attributed the identification to the randomization. The claims that other fixes produce no measurable effect on bug rates, and my suggestion to filter static analysis output to complexity alerts, are not in the part of the abstract I read and are my own.