Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard

Memdoor’s team-solve is a reusable workflow for tackling programming problems with multiple AI agents. A planner, two ideators, a solver, a manager and a repair worker operate in fresh contexts, exchanging plans, proposals, code and review instructions through files rather than chat. Commands run sample tests, but the final command - not a model’s judgment - determines whether the solution passes. Users can install the workflow, provide a task and paired input/output samples, and run it with a configured model. The workflow is intended to reproduce the method described by GVS5H, including its prompts and iterative process.

In a five-pass test on the first ten problems in GVS5H’s LiveCodeBench Hard set, the Qwen3.8-27B workflow solved 49 of 50 runs and averaged 9.8 out of 10, compared with 9.6 for the authors’ 27B harness and 8.0 for Fable 5 alone. These results cover only ten problems, not the authors’ 100, and Fable 5 was evaluated from the authors’ results rather than in the same harness. The runs used OpenRouter-hosted models, cost $35.79 in total, and took a median of 40 minutes each; training-data contamination cannot be ruled out. Failures included a manager overturning a correct solution based on a flawed counterexample and later rounds replacing solutions that already passed hidden tests. The write-up identifies safeguards for both problems but says they were not included in the reported scores.