| # Adding a policy | |
| 1. Use a unique lowercase, hyphenated model ID under `policies/`. | |
| 2. Use the `embodichain-rl-run-v1` layout: `checkpoint.pt`, `run-manifest.json`, | |
| `configs/train.yaml`, `configs/env.yaml` and `evaluation.json`. Keep manifest | |
| and configuration references relative. State any applicable weight license explicitly. | |
| 3. Record the training source, checkpoint selection basis, tested software | |
| versions, limitations and checkpoint metadata in evaluation.json. Keep MP4 files, robot | |
| assets, local paths, credentials and training logs outside this repository. | |
| 4. Check checkpoint/configuration consistency and evaluate with the declared | |
| runtime and an empty asset cache before publication. | |
| 5. Update index.json and the model files in the same reviewed repository | |
| revision. Keep the revision history referenced by released EmbodiChain | |
| versions. New environment implementations or model formats need matching | |
| EmbodiChain/task-package support before they can be evaluated. | |
| Keep shared usage instructions in EmbodiChain documentation and link them from | |
| the repository README. Individual model README files are not required. | |