Curious what the actual tooling looked like here. A lot of "AI code review" in practice means dumping code into an LLM and asking for issues - useful but inconsistent.

The sched_ext case specifically is interesting because extensible scheduler code has to maintain invariants that span the BPF-kernel boundary. Those kinds of cross-boundary invariant violations are expensive to catch in human review because you need deep familiarity with both sides. If AI is reliably picking those up it's genuinely useful, not just finding style issues.

Would be interesting to see the breakdown of bug classes found. Race conditions vs. logic errors vs. API misuse would say a lot about where the signal actually is.