A ledger-based, zero-shot self-orchestration method coordinates multiple model instances to decompose problems and achieve frontier-like coding performance at a fraction of the cost, tested across eleven open and closed-weight models on LiveCodeBench problems.

The codebase (v2-current) includes clear environment setup, keys, and uv-based example commands, with results written to runs/results.json and workspaces stored in runs/ws/.

Results are not universal: some models show no improvement or degradation, indicating the method depends on specific model capabilities.

Artifacts and materials include figures such as gvs5h-mark-humans.png and benchmark_livecodebench.png; paper and run data are under CC BY 4.0, with code under MIT License and details in NOTICE.md.

Licensing: Code under MIT; paper and run data under CC BY 4.0, with additional licensing notes in NOTICE.md.

Key results show Orchestrated Qwen3.8 Flash Next achieving 93.0% pass@1 versus 90.4% for Fable 5 at about 9% of the cost; locally served Qwen3.8-27B improves from 66.8% to 92.4% pass@1.

Across eleven models, gains reach up to 25.6 percentage points on pinned backends, with two configurations hitting frontier-level accuracy.

Implications for teams: self-hosted coding assistants could largely bridge the gap with proprietary APIs if base models have sufficient reasoning to benefit from problem decomposition.

The work is released under CC BY 4.0, and cost analysis shows the multi-agent orchestration achieves frontier-like accuracy at significantly lower cost on a log-cost scale.

Experimental scope spans eleven open and closed-weight models on the latest LiveCodeBench problems, with results framed as passing benchmarks and tied to pinned backends and cost.

Five instances of Qwen3.8-27B can match or exceed Claude Fable 5 on LiveCodeBench Hard without fine-tuning or RL, using ledger-based zero-shot self-orchestration.

Code is available on GitHub under MIT license, with running requirements including uv, Python 3.12, and an API key for the target model.