Memdoor reports 49 successful coding runs in small multi-agent test

Memdoor reports that its workflow using Qwen3.8-27B solved 49 of 50 programming runs across ten LiveCodeBench Hard problems tested five times each. The workflow averaged 9.8 correct solutions out of ten, compared with 8.0 for Fable 5 alone in results taken from the original researchers' files. Separate model roles plan, propose algorithms, write code and review test results, exchanging information through files. The 50 runs cost $35.79 and took a median of 40 minutes of model time per run. The test covered ten problems rather than the original researchers' 100. Training contamination cannot be ruled out, and Fable 5 was not rerun in Memdoor's harness.