Imagine five cooks in separate kitchens, none of whom can see or talk to each other, each reading and writing to a shared recipe binder on the counter. One writes the menu, two independently propose dishes, a third picks and cooks, and a manager tastes the result and sends back notes — all coordinated through paper slips, never a shouted conversation. That is the core mechanism here: replacing a single powerful model's long context window with multiple cheap models that coordinate exclusively through files on disk. The committed claim, qualifiers stripped: five instances of Qwen3.8-27B, a mid-tier open model, can match or slightly exceed Fable 5 (a frontier model) on LiveCodeBench Hard problems when orchestrated through a structured multi-agent workflow. The Memdoor reimplementation scores 9.8/10 mean over five passes on ten problems, against 9.6 for the original GVS5H harness and 8.0 for Fable 5 alone. The workflow was not hand-written — the coder itself generated the files from a method description. The architecture is dead simple and that is the point. Five roles — planner, two parallel ideators, solver, manager — plus a repair worker, each starting from a blank context and communicating only through markdown files and a shared notes ledger. The key structural choice is that no agent sees another agent's raw conversation; they see only the artifacts. A shell script running actual test cases decides correctness, not a model's self-assessment. This is the file-mediated multi-agent pattern from GVS5H (Lee, 2026), where the insight is that fresh-context agents reading structured files outperform a single agent with a long, degrading context window. The integrity picture is unusually transparent for an open-source demo. All 50 runs (10 problems × 5 passes) are published with every request, reply, program, grader output, cost, and hosting provider. The authors are honest about contamination risk — the contest problems date from March–April 2025 and the model may have seen them in training. They are honest about the sample being only 10 problems, not the full 100 from the original paper. They publish the two failure modes: a manager inventing a wrong counterexample, and a correct program being overwritten by unnecessary continued iteration. The grading is done by the actual test harness against hidden tests, not model self-evaluation. The ladder comparison is direct and named. Against the GVS5H authors' own rebuilt harness on the same model (9.6/10), Memdoor's workflow slightly leads (9.8/10). Against Fable 5 alone (8.0/10), the gap is substantial. But the sample is small — ten problems, not a hundred — and the authors' columns come from their own regraded files, not an independent rerun. The cost efficiency is striking: 50 runs cost $35.79 total, median run time 40 minutes, median cost $0.39 per problem. All inference ran on OpenRouter across Modal, CoreWeave, Wafer, and Parasail endpoints in mixed precision. The milestone question is whether this file-mediated multi-agent pattern generalizes beyond competitive programming. LiveCodeBench Hard is a well-defined arena with clear test cases and unambiguous correctness. The real unlock would be demonstrating the same coordination gain on tasks where correctness is harder to verify — system design, refactoring, specification-ambiguous engineering problems. The 10-problem sample also needs scaling to the full 100 to be conclusive. The obvious next experiment the authors did not run: the full 100-problem LiveCodeBench Hard suite. The honest read is budget — at $0.39 median per run and five passes, the full suite would cost roughly $195, not prohibitive but not trivial for a demo. The more interesting unrun experiment is ablation: what happens with three agents instead of five? With one ideator instead of two? With no repair round? The parallel-ideation and file-ledger discipline are claimed as the key mechanisms, but no ablation isolates their individual contributions.
