CLAIM-2026-10-03-04

Of 2,967 episodes with usable action records, 1,638 (55%) issued a full fix implementing all reference-solution actions, and of those only 59% (966 of 1,638) pass.

Why it matters

The agent finding and implementing the right repair is not the bottleneck. Roughly two in five complete repairs still fail, which moves the engineering problem from capability to closure.

Sources

• arxiv2026full