Start with the repository quick start, then choose a guide.
| Guide | Purpose |
|---|---|
| Model | Architecture, parameter layout, output semantics and checkpoint precision |
| RLCD | Probability-aware RL, constrained abstention and the proposed multi-step extension |
| Inference | Independent HF decisions, native Engine, qualification and CUDA Graph |
| API | Requests, distributions, confidence and errors |
| Results | Full public results, regular benchmarks, calibration and latency |
| Inference latency | Current HF/eager/CUDA Graph timings, raw samples and reproduction |
| Evaluation | Reproduce the retained public benchmarks |
| Official Decision Index | Evaluate the GS1 HTTP service using upstream scripts |
| Playground | Typed queries, probabilities and raw JSON |
The gs1 wheel contains only model inference (model/), optimized execution (inference/) and local serving (api/). Benchmarks and qualification generation live in benchmark/; the replica launcher is scripts/serve_replicas.py. These tools and the Node playground run from the source checkout. Weights and evaluation inputs live outside Git.