Loreley Ran 1,008 Repo-Scale Evolution Jobs and Found Quality-Diversity Search No Better Than Sequential Editing

arXiv 2608.19703·medium signalXBlueskyLinkedInCopy link

arXiv 2608.19703 keeps whole repository states in a Quality-Diversity archive, producing candidates as Git commits in isolated worktrees judged by a project-supplied evaluator, then samples archived states as parents or context for later edits. In a matched Zstandard experiment of seven paired blocks and 48 candidate jobs per policy per block (1,008 jobs total), QD landed 0.135% below sequential champion editing (95% BCa interval -0.556% to +0.161%) and 0.320% above independent root proposals. Sequential had the highest observed 48-job mean and median, and while archive retention and stepping-stone sampling demonstrably occurred, the controlled experiment did not establish any endpoint benefit.