Memdoor’s team-solve example separates a planner, two ideators, a solver, a manager and repair steps, using fresh contexts and files for coordination. Its public experiment repeated the first ten LiveCodeBench Hard problems five times and passed 49 of 50 runs, with a shell grader rather than a model judging success. Related coverage highlights the score, but the comparison figures were not rerun under the same harness. The repository explicitly notes possible training contamination from older contest problems; median runtime was about 40 minutes and the longest run took four hours. Latency and reproducibility therefore matter alongside the score.