Looping Is Not Reliability: State-Bound Evidence and Typed Revision Contracts for Agentic Code Repair
2026-07-27 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors examine how coding agents repeatedly try to fix code but often fail to keep or verify the correct fixes. They analyze many repair attempts and find that while the chance of having a correct solution at some point increases, the chance of maintaining a currently correct fix decreases over revisions. They identify issues like outdated verification traces causing harm and propose a new system that tightly links verification evidence to specific code states to better track and preserve correct patches. Their implementation focuses on defining and enforcing this process rather than proving it improves actual repair success.
coding agentscode repairgenerate-test-revise loopverificationHumanEvalprogram repairtracerollbackloop contractsoftware correctness
Authors
Xueping Gao, Jianwei Yang, Qiang Yang
Abstract
Generate--test--revise loops are common in coding agents, but repetition alone provides no reliability guarantee. We study the gap between finding a correct patch and retaining, verifying, and submitting it. A sealed five-seed study over 30 HumanEval repairs produces 900 three-revision trajectories. Under forced revision, current correctness with current traces falls from 0.820 after one revision to 0.673 after two, although ever-correct rises to 0.847. Two common-state studies use 2,430 branches from identical frozen programs to remove post-treatment risk-set bias. In a prespecified 14B replication, stale traces harm 34/135 correct starts versus 4/135 with current traces, a 22.2-point increase (task-cluster 95\% CI $[8.9,37.0]$, exact Holm $p=0.0337$). A prospective 540-rollout policy eliminates observed correct-start harm but reduces wrong-start repair and fails its joint criterion. Repository experiments over 24 bugs and four coder stacks expose floor effects and component heterogeneity without Holm-significant effects. We therefore separate admission, preservation, grounded certification, competence, and liveness. We derive an evidence-bound typed loop contract and instantiate its mechanically enforceable subset in a reference implementation that binds verifier evidence to exact code states, preserves verified checkpoints, and emits auditable admission receipts. The implementation is an executable specification and conformance artifact, not evidence of improved repair competence or calibrated verifier dependence.