A Memdoor workflow implementing the GVS5H multi-agent method (Lee, 2026) achieves a mean score of 9.8/10 over five passes on the first ten LiveCodeBench Hard problems using five Qwen3.8-27B models, surpassing the original authors' harness (9.6) and Fable 5 alone (8.0). The workflow employs five distinct roles—planner, two parallel ideators (requesting the most direct provable algorithm and an alternative technique family), solver, and manager—plus a repair worker, all operating with fresh contexts and communicating exclusively via files on disk. A shell script validates success against hidden tests, not model judgment. The configuration uses a 128K token reply cap, temperatures of 0.3 (plan), 0.4 (ideation), and 0.2 (manager/worker), top_p 0.95, top_k 20, and up to 10 rounds. Across 50 runs (49 solved), the median model time was 40 minutes at a cost of $0.39 per run ($35.79 total for 15.3M output tokens on OpenRouter via Modal, CoreWeave, Parasail, and Wafer FP8 hosts). Two failure modes persisted from the original loop: a manager inventing an incorrect counterexample causing a regression on abc399_e, and a manager failing to terminate after a correct solution, overwriting it on arc196_c/d. The workflow is executable via Memdoor CLI or TUI, requiring only a task.md and paired test files, with steps defined in YAML DAGs compatible with the mario engine.