{"data":{"kind":"file","path":"README.md","version_id":"sastgkxxxh875pm3se8vyi6h","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2629,"modified_at":"2026-08-02T22:26:59.207000","content_hash":"8a3ab9ac8b66f6a3a2c92bae7d0435150d8b658ad80d86fea415ac1c118f07f7"},"entries":[],"content":"# Beer Distribution Game\n\nA public Verifiers environment for evaluating tool-using LLMs on delayed\nsupply-chain control. One model controls one role through a strict\n`place_order(quantity)` tool while deterministic scripted policies control the\nother roles. The simulator is deterministic and replayable: the same scenario\nseed and action trace always produce the same outcome.\n\nThe environment provides five tiers: steady demand, persistent stochastic demand,\na hidden regime shift, partial pipeline observability, and strategic scarcity in a\ntwo-retailer Y network. Episodes are seeded, exactly replayable, and graded from\nthe simulator trace. The primary result is controlled-role cost normalized against\na same-seed adaptive base-stock reference; service, bullwhip, and system externality\nremain separate metrics.\n\n## Human baseline Gradio app\n\nPlay Tier 5 Y as the **Wholesaler** (identical Hub seeds, FOW observations,\nhorizon, and grading). See [`README_HUMAN.md`](README_HUMAN.md) and\n[`docs/HUMAN_TO_MODEL.md`](../../docs/HUMAN_TO_MODEL.md).\n\n```bash\npython3 -m pip install -r requirements-space.txt\nPYTHONPATH=. python3 human_app/app.py\n```\n\n## Local validation\n\nFrom this directory with `uv` installed:\n\n```bash\nuv sync\nuv run validate beer-distribution-game --runtime.type subprocess\n```\n\nExample dry-run configuration:\n\n```bash\nuv run eval @ eval.toml --dry-run True\n```\n\nThe package exports one `BeerTaskset` and one bundled `BeerHarness`, following the\nnative Verifiers 0.2 contract. It requires no API keys for simulation or grading.\nThe checked-in evaluation config disables result upload by default.\n\n## Current development evidence\n\nThe steady retailer task is a protocol/control screen. DeepSeek V4 Flash matched\nthe corrected base-stock reference exactly (cost 69, reward 0.5). The complementary\nTier 5 Y-wholesaler task is harder: across three development seeds the model cost\nwas 1,111.8 ± 213.2 versus 850.7 ± 326.1 for paired base-stock, with reward\n0.423 ± 0.060. All 108 actions were protocol-clean. This is a development finding,\nnot a held-out or multi-model benchmark result.\n\nEvaluation configurations are named by role and scope:\n\n- `eval_akash_retailer_smoke.toml` and `eval_akash_retailer_dev.toml`;\n- `eval_akash_wholesaler_y_smoke.toml` and\n  `eval_akash_wholesaler_y_dev.toml`.\n\nStart with a one-seed smoke. All Akash configurations run sequentially, keep\nuploads disabled, and read the API key from `AKASH_API_KEY`; credentials are\nnever stored in configs or compact result artifacts.\n\nSee the repository-level specifications for the complete interface, reward, and\ndifficulty contracts.\n","encoding":"utf-8","truncated":false,"total_bytes":2629},"status":null}