# Start here

Choose the shortest path that fits your goal.

| Goal | First action |
|---|---|
| See the mechanism | Download and open `web/YEORIN_X_Live_Lab.html` |
| Explore the mathematical controller | Run `examples/lease_example.py` after setup |
| Reproduce the reference | Build; test; recorded replay |
| Try a new failure case | `tools/research.py experiment` with a new seed and run directory |
| Implement a host or subsystem | Read `docs/IMPLEMENTATION_GUIDE.md` and `ai/ARCHITECTURE.md` |
| Work with an AI agent | Give it `AGENTS.md`, then choose a task from `ai/TASKS.md` |
| Upload this release | Run the outer package's `UPLOAD_TO_HUGGINGFACE.bat` |

## Set up the research environment

Run from the directory containing this file. Install Python 3.11+ for this pinned replay profile, CMake 3.16+, and a C++17 compiler. On Windows, use the x64 Visual Studio developer terminal with the Desktop C++ workload and SDK. The independent uploader supports Python 3.10+.

```bash
python -m venv .venv
# Activate: Linux/macOS
source .venv/bin/activate
# Activate: Windows PowerShell
# .\.venv\Scripts\Activate.ps1
# Activate: Windows cmd
# .venv\Scripts\activate.bat
python -m pip install -r requirements-replay.txt
python -m pip install -e . --no-deps
python tools/research.py doctor
python examples/lease_example.py
python tools/research.py build
python tools/research.py test --out runs/first-test
python tools/research.py replay --recorded --backend cpp --out runs/first-replay
```

Expected reference evidence: 253 Python tests without skips, a C++ selftest, and 144 replay rows. The `recorded_comparison.json` file reports whether the recorded choices match within numerical tolerance. The reference primary summary has 129 active corrections and approximately 2.295348-pixel corrected RMSE. Dependency or platform differences can affect replay inference; inspect the comparison instead of assuming it matches.

Run names must be new. Choose a different `--out` to repeat. Successful tests verify the supplied scope; they do not establish game or optical performance.

## Explore a new seed

```bash
python tools/research.py experiment --scenario camera_lock --seed 9101 --out runs/lock-9101
```

Open the generated `summary.json`, `summary.csv`, `replay.json`, and `run_manifest.json`. The generator stores truth in `generated_case/`; the observer sees only `input_only/`. Zoom produces activation diagnostics with null geometry metrics because that scorer does not support changing FOV.

## Next reading

- `docs/EXPERIMENT_GUIDE.md`: scenarios, metric interpretation, baselines and new-run workflow.
- `docs/IMPLEMENTATION_GUIDE.md`: subsystem interfaces, units, source ownership, completion and admission.
- `docs/DATASET_SCHEMA.md`: JSONL tables and binary arrays.
- `docs/TROUBLESHOOTING.md`: environment, library discovery and platform failures.
- `docs/READINESS_AND_LIMITS.md`: verified components and remaining integration gaps.

The browser is immediately usable; the research workflow requires dependencies. The Windows source needs target-platform validation. No universal game installer is supplied.