0_abstract · 完整稿件源码

Repository-repair agents produce long interaction traces, while their tests usually score only the completed patch. Counterfactual edit credit connects these objects through an executable intervention on the final artifact. The method selects one surviving edit block without consulting the verifier outcome, restores its touched paths to their pre-block contents, and tests the resulting patch in a fresh environment. The signed change in test utility becomes an additive advantage on the generated tokens of the turns that produced the block. This construction requires at most one additional counterfactual verifier trial per eligible rollout and no additional policy continuation. A broadcast control receives the same signed and absolute auxiliary coefficient mass on a common reference trajectory, enabling a controlled comparison of allocation rules. We also specify a terminal-only control and a terminal-advantage edit-focus control that uses no counterfactual score. The accompanying protocol fixes training endpoints and separates tasks used for reinforcement learning from its evaluation complement. On 320 development trajectories, 244 removal queries are usable and 92 yield a nonzero contrast, including 17 from failed originals. These measurements support signal availability; policy-effectiveness measurements remain pending.