# SparseWake Artifact Audit Report

Audit date: 2026-05-06

Scope: anonymous SparseWake dataset release. The manuscript source checked for consistency was `main_v11.tex`; it was read but not compiled or edited.

## Overall Status

| Requirement | Status | Notes |
|---|---|---|
| Anonymous release contents | PASS WITH ACTION | Release tree text scan found no author names, private emails, private repository URLs, local paths, usernames, machine names, or institution names. The parent repository `.git` history is not anonymous and must not be included in the review ZIP. |
| Manuscript-to-artifact coverage | PASS WITH NOTES | Manuscript datasets, tables, result summaries, and referenced figure PDFs are present. Main and supplement figure PDFs are included under manuscript-compatible paths. |
| HDF5 benchmark data | PASS | Three full processed HDF5 files and one sample HDF5 are present. HDF5 root/dataset attributes contain no identity metadata. |
| Reproducible benchmark code | PASS | Sample verification, table reproduction, figure reproduction, and quick training smoke test run from the release tree. |
| Data integrity | PASS | Checksums are listed in `data/checksums.sha256`; `scripts/verify_dataset.py` can verify sample structure and checksum entries. |
| Documentation | PASS | README, dataset card, evaluation card, provenance, field documentation, limitations, license, CITATION.cff, Croissant metadata, audit report, manifest, and release checklist are present. |
| Double-blind packaging | PASS WITH ACTION | Package the release-tree contents only. Do not package the parent repository `.git` directory or development logs outside the release tree. |

## Manuscript-to-Artifact Manifest

| Manuscript item | Expected artifact file | Status | Notes |
|---|---|---|---|
| Benchmark overview figure, `figures/main/Fig1.pdf` | `figures/main/Fig1.pdf` | PASS | PDF included; embedded local source reference was anonymized. |
| Main localization figure, `figures/main/Fig2.pdf` | `figures/main/Fig2.pdf`; reproducible summary figure `figures/fig02_main_results.*` | PASS | Main PDF included; script-generated PDF/SVG also present. |
| Self-signal figure, `figures/main/Fig3.pdf` | `figures/main/Fig3.pdf`; reproducible summary figure `figures/fig03_self_signal_control.*` | PASS | Main PDF included; script-generated PDF/SVG also present. |
| Architecture screen figure, `figures/supp/Fig4.pdf` | `figures/supp/Fig4.pdf`; `figures/figS4_architecture_screen.*` | PASS | CSV source: `data/results/model_screen.csv`. |
| Noise/SNR diagnostics, `figures/supp/Fig5.pdf` | `figures/supp/Fig5.pdf`; CSV source `data/results/noise_snr_diagnostics.csv` | PASS | Figure PDF included. |
| Orientation diagnostics, `figures/supp/Fig6.pdf` | `figures/supp/Fig6.pdf`; CSV source `data/results/orientation_metrics.csv` | PASS | Figure PDF included. |
| Sample-size convergence, `figures/supp/Fig7.pdf` | `figures/supp/Fig7.pdf`; `figures/figS5_sample_size_convergence.*` | PASS | CSV source: `data/results/sample_size_convergence.csv`. |
| Dataset validation, `figures/supp/Fig8.pdf` | `figures/supp/Fig8.pdf` | PASS | Figure PDF included. |
| Spatial error heatmap, `figures/supp/Fig9.pdf` | `figures/supp/Fig9.pdf` | PASS | Figure PDF included. |
| Benchmark specification table | `README.md`, `docs/benchmark_protocol.md`, `docs/data_fields.md` | PASS | Protocol, fields, sensors, labels, metrics, and split are documented. |
| Dataset roles table | `data/processed/README.md`, `data/processed/download_manifest.json` | PASS | Role names map to release HDF5 files. |
| Architecture-screen pilot | `data/results/model_screen.csv`, `configs/model_screen.yaml` | PASS | Result summary included; pilot HDF5 is not released as a separate full dataset because it is auxiliary. |
| External component benchmark | `data/processed/abm_sensing_dataset_v04_close_orient30_potential_160k.h5` | PASS | Contains wake-only, potential-only, wake-plus-potential, labels, poses, regions, and sensor world positions. |
| Paired self-signal controls, aligned | `data/processed/abm_sensing_dataset_v05_self_external_paired_160k_compressed.h5` | PASS | Contains external-only, self-only, and total-like arrays. |
| Paired self-signal controls, randomized orientation | `data/processed/abm_sensing_dataset_v05_self_external_orient30_160k.h5` | PASS | Contains external/self/total arrays for randomized follower orientation. |
| History sweep H=1,4,8,16,24 | `data/results/main_numbers.csv`, `data/results/v04_h_sweep_summary_table.csv` | PASS | Values match manuscript prose within rounding. |
| Full-state `location_theta` | `data/results/orientation_metrics.csv`, `configs/main_v04.yaml` | PASS | Close-wake position RMSE `0.040` and heading MAE `0.53 deg` trace to CSV. |
| Sensor ablation 6,4,3,2 sensors | `data/results/sensor_ablation.csv`, `data/results/v04_sensor_subset_summary_table.csv` | PASS | All counts referenced in the manuscript are present. |
| Common raw-noise stress test | `data/results/component_noise.csv`, `data/results/v04_component_common_noise_summary_table.csv` | PASS | Potential-only, wake-only, and wake-plus-potential rows are present. |
| Self-signal controls | `data/results/aligned_self_signal_control.csv`, `data/results/self_signal_randomized.csv`, `data/results/v05_self_control_table.csv` | PASS | Aligned and randomized-orientation controls are present. |
| Total-like noise robustness | `data/results/v05_total_noise_table.csv`, `data/results/main_numbers.csv` | PASS | Noise multipliers through 0.05 are present. |
| Sample-size convergence | `data/results/sample_size_convergence.csv`, `configs/sample_size.yaml` | PASS | 500, 1000, 2500, and 4000 training-pose conditions are present. |
| Compute report | `data/results/compute_resources.csv`, this report | PASS WITH UNCERTAINTY | Exact CPU/GPU/RAM were not logged in the frozen run and are marked unresolved below. |

## Data and Code Checks

HDF5 files checked:

| File | Status | Notes |
|---|---|---|
| `data/processed/abm_sensing_dataset_v04_close_orient30_potential_160k.h5` | PASS | `160000` samples; fields include `X_raw`, `X_wake_raw`, `X_potential_raw`, `y`, `groups`, `region_id`, `follower_pose`, `leader_pose`, and `sensor_world_positions`. |
| `data/processed/abm_sensing_dataset_v05_self_external_orient30_160k.h5` | PASS | `160000` samples; includes external, self, and total-like component arrays. |
| `data/processed/abm_sensing_dataset_v05_self_external_paired_160k_compressed.h5` | PASS | `160000` samples; includes aligned paired self-signal component arrays. |
| `data/sample/sparsewake_sample.h5` | PASS | `2048` samples and explicit `pose_id`; used for smoke tests. |

Implementation checks:

| Check | Status | Notes |
|---|---|---|
| Global follower coordinates excluded from model inputs | PASS | Training code loads only selected `X_*` component arrays as `input`; `follower_pose` and `sensor_world_positions` remain metadata. |
| Held-out-pose test logic | PASS | `src/sparsewake/splits.py` matches the experiment driver: `default_rng(seed).choice` selects held-out test poses, then sklearn `train_test_split(..., test_size=0.15, random_state=seed)` splits the remaining samples into train/validation. |
| Train-set-only standardization | PASS | `src/sparsewake/train.py` fits input mean/std only on `train_idx`. |
| `raw_norm` feature dimension | PASS | `src/sparsewake/features.py` produces `37` features per time step for six sensors and `888` dimensions for H=24. |
| Main metric implementation | PASS | Position RMSE and circular `theta_rel` MAE are implemented in `src/sparsewake/metrics.py`. |

## Anonymity Scan

Text files scanned: Markdown, Python, YAML, JSON, CSV, CFF, SVG, checksum, and text files in the release tree.

Binary/metadata checked: HDF5 root and dataset attributes; PDF/SVG metadata and raw byte search for known local identifiers and path patterns.

Findings:

- Release tree text scan: PASS.
- Croissant metadata: PASS. It uses anonymous creator/citation fields and contains no creator email, publisher institution, private URL, or local path.
- HDF5 attributes: PASS. Full HDF5 files have no root or dataset attributes; sample HDF5 attributes are generic.
- PDF/SVG metadata: PASS after sanitizing one embedded local source-file reference in `figures/main/Fig1.pdf`. Remaining creator/producer metadata names software only, such as Matplotlib or Adobe PDF library.
- Git history: FAIL FOR PARENT REPOSITORY. The parent repository history contains author names, emails, and a private remote path. Do not include `.git` in the anonymous ZIP.

## Commands Tested

These commands were run from the release-tree root:

```bash
python scripts/verify_dataset.py --data data/sample/sparsewake_sample.h5
..\.venv\Scripts\python.exe scripts/train_temporal_mlp.py --config configs/main_v04.yaml --data data/sample/sparsewake_sample.h5 --quick
python scripts/reproduce_tables.py --results data/results --out tables
python scripts/make_main_figures.py --results data/results --out figures
python scripts/make_supp_figures.py --results data/results --out figures
```

Additional structural checks:

```bash
..\.venv\Scripts\python.exe -c "import sys; sys.path.insert(0,'src'); from sparsewake.data import load_h5; from sparsewake.features import raw_norm_features, build_design_matrix; d=load_h5('data/sample/sparsewake_sample.h5'); print(raw_norm_features(d['input']).shape); print(build_design_matrix(d, history=24)[0].shape)"
git log --all --format="%H %an %ae %s" -n 40
```

The default system `python` environment did not include `torch`; the repository virtual environment did. A fresh reviewer environment should install `requirements.txt` or `environment.yml`.

## Files Changed During Audit

| File | Change |
|---|---|
| `src/sparsewake/features.py` | Corrected `raw_norm` to match `abm_sensing_dataset/experiments/sensing_v2/feature_builders.py`: raw channels, normalized raw channels, and one denominator per time step. |
| `src/sparsewake/splits.py` | Corrected split helper to match `run_sensing_v2_experiments.py::make_pose_holdout`: held-out test poses via `default_rng(seed).choice`, then sklearn train/validation split within non-test samples. |
| `requirements.txt`, `environment.yml` | Added `scikit-learn` because the actual experiment split uses `sklearn.model_selection.train_test_split`. |
| `scripts/verify_dataset.py` | Added checksum verification option. |
| `README.md` | Added checksum verification to the sample verification command. |
| `dataset_card.md` | Corrected train/validation/test protocol wording. |
| `docs/benchmark_protocol.md` | Corrected split protocol wording. |
| `evaluation_card.md` | Corrected split protocol wording. |
| `figures/main/Fig1.pdf`, `figures/main/Fig2.pdf`, `figures/main/Fig3.pdf` | Added manuscript-compatible main figure PDFs; sanitized embedded source path in Fig1. |
| `figures/supp/Fig4.pdf` through `figures/supp/Fig9.pdf` | Added manuscript-compatible supplement figure PDFs. |
| `tables/quick_train_metrics.json` | Updated by smoke-test training. |
| `AUDIT_REPORT.md` | Added this report. |
| `RELEASE_CHECKLIST.md` | Added final upload checklist. |
| `MANIFEST.json` | Added generated file manifest. |
| `data/checksums.sha256` | Updated final checksums. |

## Unresolved Issues Before Upload

1. `data/processed/download_manifest.json` should contain relative paths or anonymous hosted paths for the full HDF5 files.
2. The parent repository `.git` history is not anonymous. The final review upload must contain only release-tree contents and must not include `.git`.
3. Exact CPU/GPU model, RAM, MATLAB version, and dataset-generation wall time were not logged in the frozen run. The artifact reports conservative wall-clock times for representative three-seed runs and marks hardware details as unknown.
4. The architecture-screen pilot is represented by summary CSV/config rather than a separate pilot HDF5 release. This is consistent with the manuscript role because the pilot is auxiliary; reviewers can reproduce benchmark evaluation from the released full processed HDF5 files.

## Manuscript Statements Needing Revision

No numeric scientific result needed revision based on this audit.

Recommended non-scientific wording if the artifact is hosted outside the review system: state that the full processed HDF5 files are included directly in the anonymous dataset repository.