| # Context Window: Demon Attack | |
| 89 archived training runs. Seeds 10–14. Native Sample Factory APPO policies. | |
| ## Settings | |
| Context windows: 4–12. Environment FPS: 60. Fixed delay: 4 environment frames. Uniform delay: 2–6 environment frames. Appendix evaluation horizon: 36,000 steps. | |
| The archive retains each evaluation's original horizon, including earlier 3,600-step records. Each original evaluation has 100 episodes. | |
| Exact observation cadence, frame stack, training budget and delay parameters belong to each run's unchanged `config.json` and [manifest](manifest.json). Fixed and uniform distributions are experiment settings, not measured hardware profiles. | |
| ## Results and assets | |
| [Paper measurements](https://huggingface.co/datasets/MLL-Lab/LAGEN-datasets/blob/3a59b67c3f85954742905d2c57b0d40e0050f979/experiments/context-window/paper/data.csv) and [fixed W&B sources](https://huggingface.co/datasets/MLL-Lab/LAGEN-datasets/blob/3a59b67c3f85954742905d2c57b0d40e0050f979/experiments/context-window/paper/wandb_sources.csv) retain the published values. Each plotted point aggregates five training seeds. | |
| The manifest pins model, configuration and original evaluation assets to fixed main-repository revisions. Training logs are stored with the evaluation run under `training/`. Download the model bundle with its original configuration and load it with the task's native Sample Factory integration. | |
| ## Limits | |
| The Demon uniform-delay C=12, seed-14 paper result is preserved from W&B history step 95483, but its checkpoint is missing. | |