# ORBench reproduction protocol v1

Seed: 25. Code version: 0.1.0. Solver: OR-Tools CP-SAT, 1 thread, 6–8 seconds.

1. Generate the catalog with `python scripts/build_artifacts.py`.
2. Each problem stores `schema_payload`, bilingual prompts, `gold_status`, `gold_objective`, and `gold_solution`.
3. Evaluated methods: `structured-solver`, `heuristic`, `json-extractor+solver`, `template-extractor+solver`, `naive-direct`.
4. Listed but not run: `Qwen/Qwen3-4B-Instruct-2507`, `Qwen/Qwen3.8-27B`, `moonshotai/Kimi-K3`, `zai-org/GLM-5.3`, `deepseek-ai/DeepSeek-V4.1-Flash`. Cells are `not_evaluated`.
5. Gap = max(0, (method_obj − gold_obj) / max(1, |gold_obj|)) only if gold status is optimal and the method plan is feasible.
6. Paid APIs stay off. Do not substitute a second paid provider when the first is missing.
7. Synthetic results in this dataset are not customer data.

Run `pytest tests/test_pipeline.py` after generation. A passing suite is the acceptance check for constraint validation, split hygiene, and honest unevaluated cells.