# Preprint finds targeted edge tests reduce faulty code repair releases

_Published Friday, November 6, 2026 at 8:52 AM EST · Science, AI · Latest · Tier 2 — Notable_

A preregistered preprint reports that adding one edge-case input derived from a release note reduced the share of wrong code repairs passing a release check from 0.410 to 0.065. The study covered 73 tasks spanning 60 documented runtime or API changes and executed candidates on real runtimes against withheld contracts.

The result reached 0.061 on 15 unseen clusters. Later stages added more candidate fragments and hard distractors that matched legacy output on typical inputs. Even then, 0.168 of hard wrong programs passed the edge-input check. No model-written code was executed, and disputes over contract labels remain open.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2611.08269)

---
Canonical: https://techandbusiness.org/newswire/cCqWA0bxg39TbyZj0W7g4w
Published: 2026-11-06T13:52:43.214Z
Story chronology: 2026-11-06T05:00:00.000Z
Retrieved: 2026-11-06T22:57:22.465Z
Publisher: Tech & Business (techandbusiness.org)