The most interesting thing about this result is not the score. It is that the full evidence trail — every request, every reply, every generated program, every grader output, every dollar spent — was published alongside it, including the failures.
The Method
GVS5H describes ledger-based zero-shot self-orchestration: fresh instances of a single model decompose a problem and coordinate through a shared filesystem holding a plan, notes and the current solution. No training, no orchestrator model — the file ledger is the coordination layer.
Across eleven open and closed-weight models on the 100 latest hard LiveCodeBench problems, the method produced gains of up to 25.6 percentage points on pinned backends. Orchestrated Qwen3.8 Flash Next reached 93.0% pass@1 against Fable 5’s 90.4% at 9% of the cost. Locally served open-weight Qwen3.8-27B went from 66.8% to 92.4%. The authors are explicit that gains are not universal — some models got no better, or worse.
The Independent Rebuild
The Memdoor team then rebuilt the harness as a declarative workflow with five roles — a planner, two parallel ideators, a solver, a manager — plus a repair worker, each starting from a fresh context and communicating only through files on disk. A shell script, not a model, decides success.
On the first ten problems, five passes each: mean 9.8 of 10, against 9.6 for the authors’ own harness and 8.0 for Fable 5 in a single call. 49 of 50 problems solved. The 50 runs cost $35.79 (15.3M output tokens), or $38.77 including sixteen attempts cut by a clock, stopped for a fix or paused. Median run: 40 minutes of model time and $0.39.
The Honest Part
The caveats are printed next to the numbers. Ten problems, not a hundred. The contests date from March 29 to April 6, 2025, so contamination cannot be ruled out. Fable 5’s column is the authors’ single-call file, not a run in this harness.
And two runs failed in instructive ways: in one, a manager invented a counterexample with a wrong expected answer, the team rewrote a correct program, and the rewrite passed the visible samples while failing the hidden test. In two others, a program that already passed every hidden test was overwritten because the manager never said done — at $3.44 and $4.24, the two most expensive runs in the set. Both are fixable without changing the method. Neither fix is in those numbers.
The pattern is the story: a cheap model in a disciplined loop with a written ledger will keep working after the context is gone — and the manager, not the coder, is where the failures live.
Related: GVS5H · Memdoor team-solve workflow · HN thread