GVS5H is a training-free coding harness in which fresh-context model instances coordinate through shared plans, notes and solution files. The project reports that Qwen3.8-27B improved from 66.8% to 92.4% on 100 hard LiveCodeBench problems. Gains were not universal, and comparison conditions were not identical for every model. A related developer discussion questioned possible training-data overlap in older problems and whether the benchmark represents real agentic software work. The results motivate testing orchestration on one’s own tasks, not a conclusion of general development parity.