# SHWRP-2026-033 v1.0 — optical crop-disease screening prototype

This directory is the public version 1.0 research package and executable,
abstention-aware optical crop-disease screening baseline for DJI Mavic 3M-compatible
RGB plus G/R/RE/NIR inputs.

- DOI: <https://doi.org/10.5281/zenodo.21787704>
- Zenodo record: <https://zenodo.org/records/21787704>
- Article: <https://bento24235111.com/research-journal/m3m-optical-disease-screening>
- Published and revised: 2026-08-04
- Status: public working paper; not externally peer reviewed

The software does **not** diagnose a pathogen, control a UAV, determine food
safety, or recommend treatment. Its published output vocabulary is restricted to:

- `REJECTED_QC`
- `REJECTED_OOD`
- `NO_FLAG`
- `REVIEW_REQUIRED`

The bundled model was trained only on deterministic synthetic fixtures. Its
`promotionStatus` is therefore forced to `RESEARCH_ONLY`. Synthetic metrics verify
software behaviour and must never be reported as field plant-disease accuracy.

## Reproduce the engineering evidence

```bash
python -m pip install -r requirements.txt
python -m pytest -q -p no:cacheprovider
python -m m3m_disease demo --output outputs/demo --seed 20260804
python verify_reproducibility.py
python generate_figures.py
python build_pdf.py
python render_pdf.py
python validate_study.py
python package_release.py
python validate_package.py
python update_zenodo_request.py
```

The core CLI also exposes `fixture`, `validate`, `train`, `evaluate`, and
`predict`. Run `python -m m3m_disease --help` for the exact contracts.

## Research artifacts

- `paper.md` and `paper.pdf`: full searchable public working paper.
- `m3m_disease/`: auditable NumPy/Pillow implementation.
- `schemas/`: capture, label, dataset, split, model, model-card, inference, and
  run-manifest contracts.
- `tests/`: 33 tests covering features, QC/OOD, leakage prevention, model hash,
  CLI, end-to-end execution, and promotion fail-closed behaviour.
- `outputs/demo/`: deterministic synthetic fixture, model, model card, split
  receipt, evaluation, and inference receipt.
- `study-protocol.md`: prospective Taiwanese field-validation protocol.
- `mavic3m-capture-and-qc.md`: raw-DN, calibration, registration, and fail-closed
  capture requirements.
- `field-label-form.csv`: empty prospective observation template; it contains no
  fabricated field observations.
- `dataset-audit.md` and `outputs/external-blb-sample-audit.json`: public-source
  audit and limited format compatibility check.
- `source-registry.csv`: primary-source registry with retrieval dates.
- `release-manifest.json`: per-file release integrity record embedded in the ZIP.
- `package-validation-report.json`: independent post-package integrity check.
- `zenodo-request.json`: exactly two upload artifacts with bytes and SHA-256.

## Data and evidence boundary

No third-party raw image, notebook, Google Drive mirror, farm coordinate, or
personal data is part of the release allowlist. The BLB UAV Dataset record at
`10.6084/m9.figshare.26955862.v3` is CC BY 4.0, but users should retrieve it from
its original DOI. A single image-mask pair, class CSV, and notebook were inspected
in an excluded temporary environment; only hashes and derived format-check results
are retained here.

The Zenodo record `10.5281/zenodo.17008818` is a useful Mavic 3M compatibility
candidate, but its License field was blank when checked on 2026-08-04. It is not
redistributed, used for public training, or included in model weights.

## Rights, AI use, conflicts, and corrections

Version 1.0 is licensed under CC BY 4.0 only for author-owned original analysis,
text, research software, and original figures. Third-party datasets, names,
trademarks, URLs, images, excerpts, manuals, and software retain their original
rights and terms. See `LICENSE.txt` and `RIGHTS-NOTICE.txt`.

Generative AI (OpenAI Codex) assisted source-lead organization, research code,
tests, drafting, figures, and layout. AI is not an author. CHOU, BING-HAN / 周秉漢
is responsible for source verification, methods, interpretation, release, and
corrections.

The author received no commission or funding from DJI, dataset authors, platforms,
pesticide vendors, or UAV service firms. The author may later provide website, AI,
or agricultural-technology services and could indirectly benefit from this
engineering-capability demonstration.

Corrections, missing attribution, and rights concerns may be sent with specific
evidence to `handson0102@hotmail.com`. Verified issues will be handled through a
versioned correction, restriction, replacement, or removal process.