Memdoor developer reports 49 successes in 50 coding runs
An October 31 test of a Qwen3.8-27B multi-agent workflow used ten LiveCodeBench Hard problems across five passes; the developer published run records but cautions that training contamination cannot be ruled out. The Fable 5 comparison uses the original study team's published results.