GraphAdmit: Evidence-Conditioned Mutation Admission for Repository Agents
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
A validator that discovers a broken import after execution arrives one state transition too late: files, dependency state, or an external system may already have changed. GraphAdmit turns that transition into an evidence-conditioned admission decision. For each structurally risky action, it materializes a dry-run graph and complete diff, intersects the requested scope with a host-frozen authorization envelope, and evaluates a model-independent universe of repository requirements, invariants, and refutations. Resolved evidence activates bounded sandbox probes; weaker evidence contracts the checked predicate set while unresolved obligations remain in the support denominator. Rejections return entity-level causes for bounded replanning, making GraphAdmit a complete execution policy rather than a proposal-only filter. Across eight paired seeds on a fixed 300-task SWE-bench Verified manifest (2,400 runs), the policy raises resolution from 29.4% to 31.2% while reducing harmful-edit tasks from 13.6% to 11.3%, high-severity harm from 4.0% to 2.8%, and rollbacks from 1.6 to 1.1 per task relative to a risk-matched scope-only policy. The paired effects exclude zero: resolution points (95% CI ), harmful-edit points (), severe-harm points (), and rollbacks (). On strong-support tasks, resolution improves by 3.2 points as harmful edits fall by 2.6 points. Common-threshold sweeps, four repair backbones, a 300-task nine-language transport manifest, and blinded adjudication preserve the severe-harm reduction. These policy-level results establish evidence-conditioned admission with structured feedback as a practical control point that redirects repository agents before structural damage without sacrificing task completion.
- Primary area:
- foundation or frontier models, including LLMs
- Keywords:
- tool agents and software engineering, GraphAdmit