# Reproducibility — WP-AI-0001

```bash
cd ia/wp-ai-0001
python3 -m pip install -e ".[dev]"            # or: uv pip install -e ".[dev]"
python3 -m pytest tests -q                     # unit + causal-sanity tests (CPU, < 30 s)
python3 scripts/audit_environment.py           # writes docs/ENVIRONMENT_AUDIT.json (add --cluster to probe nodes)
python3 scripts/run_pilot.py --quick --device cpu --seeds 0 1 --n-boot 200   # CPU smoke test, ≈ 25 s
python3 scripts/run_pilot.py --seeds 0 1 2 --steps 3000                        # Tier-0 pilot (MPS ≈ 5 min)
python3 scripts/train_models.py                                                 # 17-model matrix (≈ 13 min MPS)
python3 scripts/run_transplant.py                                               # exploratory transplant run (≈ 3 min)
python3 scripts/run_transplant.py --tag confirmatory --seed-base 90000          # confirmatory run (prereg v1.0)
python3 scripts/eval_persistent.py --run <graft run id> --seed-base 95000       # saved adapter in a fresh process
python3 scripts/confirmatory_analysis.py --run <confirmatory run id>            # pre-declared contrasts + Holm
python3 scripts/run_tier1.py [--quick] [--split dev|confirmatory] [--pairs a:b ...]   # open-weight models (MLX)
python3 scripts/build_figures.py --run <run id>; python3 scripts/build_tables.py --graft ... ; scripts/build_paper.sh
```

- Every run writes `experiments/manifests/<run_id>.json` (git commit, config hash, data hashes, seeds,
  libraries, hardware, status history), appends to `experiments/registry/runs.jsonl`, and stores results in
  `experiments/summaries/<run_id>/` (`results.json`, `SUMMARY.md`). Per-item predictions are in
  `artifacts/runs/<run_id>/` (git-ignored), models in `artifacts/models/<run_id>/`.
- Run IDs are deterministic functions of the experiment id and config hash; re-running with the same config
  reuses trained models from `artifacts/models/<run_id>/` (delete to retrain).
- Determinism: PyTorch on MPS was verified bitwise-repeatable for this model (`noop.deterministic_repeat`);
  localization sweeps repeated twice differ by 0.0. Training determinism across devices is *not* claimed
  (CPU vs MPS kernels differ); seeds reproduce training on the same device/library versions.
- Data are synthetic and regenerated from seeds; the split hashes recorded in manifests must match.
- Secrets: none required. No tokens are read; paths are relative to the project root.