SWE-Prometheus: governance-improvement agents break existing behavior in up to 23% of runs.

arXiv 2609.29465 gives agents a repo snapshot and an open-ended governance goal across 60 repos, scoring six dimensions with clean-environment probes and behavior gates. Ten models range from 0.0568 to 0.5760 mean Normalized Governance Improvement, with breakage from 0% to 23%. A template that ignores the repository entirely scores 0.272 by adding tests, CI and docs, and improves reproducible environments on zero repos. The benchmark separates agents that add governance files from agents that make the build more reliable, which is a distinction most CI-hygiene tooling doesn't make.