An LLM-based tool can now point developers to the specific lines most likely to cause production outages — before the code ships.
Researchers built a Diff Risk Score model that reads a code change and assigns risk. Instead of stopping there, they tapped the model's internal attention weights to surface which lines, hunks, or files the model fixated on. The top two flagged hunks covered the expert-labeled outage-causing lines 53.85% of the time, while asking developers to look at only about 26% of the changed lines on average. Because the explanations come from inference the model already runs, there is no meaningful latency penalty.
Code review is where bugs are cheapest to catch, yet most automated tools either flag style violations or hand back an opaque pass/fail score. This approach threads a real needle: it narrows reviewer attention without hiding the reasoning behind a black box. That auditability matters in organizations where developers are skeptical of ML-driven tooling they cannot interrogate.
The 54% hit rate is a genuine result, not a marketing number — but it also means nearly half of outage-causing changes would not appear in the top highlights. Treat it as a useful signal, not a safety net.