Embodied heuristic learning faces a trade-off: a robot must adapt its program to new situations without losing behaviors that already work. Local repairs often miss structural problems, while broader rewrites risk introducing regressions that are expensive to discover through trial and error. The new SCOPE framework addresses this by treating failed and successful executions as complementary evidence for deciding what to change.

SCOPE builds a shared diagnostic graph annotated with execution evidence, using frontier-guided search to identify which checks remain unresolved. It then applies selective behavioral certification: stateful API replay restores recorded responses and mutations, letting the system verify that a proposed edit preserves selected historical successes before it is accepted. This diagnosis-and-certification loop, supplemented by sparse environment feedback, supports consecutive repairs across different configurations.

In experiments, SCOPE outperformed the strongest baseline in both simulation and real-world tasks. On LIBERO-PRO and Robosuite, it achieved mean success rates of 80.66% and 84.48%, respectively, corresponding to gains of 6.51 and 3.48 percentage points. The authors have released the code for the project.