{"data":{"kind":"file","path":"README.md","version_id":"ionwz4b0tqpehn2z8esz7st6","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2295,"modified_at":"2026-07-16T13:47:42.965000","content_hash":"d654e73836a47d40388cccc483e1637d2dc0e057588e68410f6331df93eddb46"},"entries":[],"content":"# beads-combined-cycles-v1\n\nA self-contained Verifiers v1 repository-editing environment based on Beads\nimmediately before the combined scheduling-cycle fix. The model edits\n`/testbed`; hidden Go tests are installed only after the rollout. The image\ncontains a single synthetic Git base commit and no upstream refs/history, so\nthe answer commit cannot be discovered through local Git metadata.\n\nNo model provider, paid API, or external database is used by the reward. The\nonly remote call during a rollout is the coding harness calling the model being\nevaluated. All scoring uses Go tests and an in-process embedded Dolt database.\n\n## Task and reward\n\nVersion 0.5.0 exposes one complete cross-layer task. The prompt explicitly\nrequires the model to finish the shared graph rule, storage paths, domain policy,\nbulk insertion, and graph apply rather than treating any one package as a\nstandalone curriculum exercise.\n\nThe scalar reward is the sum of four deterministic binary components worth\n0.25 each:\n\n- `domain_policy`: hierarchy protection, mixed cycles, and error precedence\n  through the domain/SQLite path.\n- `embedded_storage`: end-to-end combined-graph behavior through embedded Dolt.\n- `bulk_add`: parent-first insertion, final cycle validation, and rollback.\n- `graph_apply`: final merged checking, implicit parents, and atomic rollback.\n\nThese are partial outcomes within the same task, not independent tasks. A model\ncan receive useful credit for a correct layer, but only a score of 1.0 means the\ncomplete repository behavior is fixed. No LLM judge is used.\n\n## Local verification\n\n```bash\ndocker build -t beads-combined-cycles:local .\nuv venv\nuv pip install -e .\n\nuv run validate beads-combined-cycles-v1 \\\n  --runtime.type docker \\\n  --taskset.image beads-combined-cycles:local \\\n  --rich false\n```\n\nThe validation command is model-free: it applies the private reference patch\nand requires all four reward components to pass on the one task.\n\n## Publish privately\n\n```bash\nprime images push beads-combined-cycles:v0.2.0 --private\nprime env push --visibility PRIVATE\n```\n\nAfter publishing, `configs/eval-prime.toml` runs one 150-turn Qwen pilot. Run\nthree pilots before committing to ten, because the target 10-30% pass rate must\nbe measured rather than inferred from the PR size.\n","encoding":"utf-8","truncated":false,"total_bytes":2295},"status":null}