Linux kernel review study finds 99% semantic stability across patch revisions

A team led by Lucas Ciziks (Universidade Federal de São Paulo) with Paulo Meirelles and Marco Aurélio Gerosa (Northern Arizona University) has published a first-of-its-kind measurement of how much code review actually alters patches in the Linux kernel. Tracking 10,117 function trajectories through numbered revisions of the Linux IIO subsystem, they compared each function's semantic similarity across versions against a baseline of unrelated function pairs. The headline finding: despite the common belief that review substantially reshapes code, 75.3% of tracked functions are never textually modified at all, inflating naive similarity scores to near-perfect levels.

When the team restricted the analysis to functions that actually received edits, semantic purpose remained largely intact—mean similarity of 0.990 versus a 0.909 baseline. Drift in meaning concentrates in the first review round, but that effect largely disappears once you account for the growing share of untouched functions in later rounds. A small, statistically detectable residual effect remains, raising an open question: does near-ceiling similarity mean reviewers successfully preserve intent, or that existing similarity tools are blind to small but significant localized edits? The paper, presented at VEM 2026, positions itself as a first look at semantic stability across review cycles, with next steps to refine measurement granularity and validate against real bug-introducing changes.