sinuosity/okuafo-maizeguard-edge-v1.4

Okuafo MaizeGuard Edge v1.4

Okuafo MaizeGuard Edge v1.4 is a compact, two-stage TensorFlow Lite system for offline maize-leaf screening. It first screens whether an image is diseased, healthy, or not suitable for maize screening. When a disease signal passes the configured confidence gate, a second model screens the leaf region for five disease-pattern categories.

Evidence statement: Research-preview maize screening model evaluated on documented development datasets and limited field samples.

This is not a claim of independent field validation, agronomic diagnosis, production readiness, release approval, or license clearance. The existing repository is publicly visible, but visibility is not permission to redistribute, deploy, or create derivative works. Every output requires qualified human review before treatment, pesticide, financial, or other consequential decisions.

Model files

File	Role	Output classes	Size
models/okuafo_maize_stage1_v1_4_fp16.tflite	maize/relevance and health gate	diseased, healthy, not_maize_or_unclear	24.5 MiB
models/okuafo_maize_stage2_disease_roi_v1_4_fp16.tflite	disease-pattern screening after a positive gate	bacterial_leaf_streak, common_rust, gray_leaf_spot, northern_leaf_blight, other_disease	24.5 MiB

Both models use an EfficientNetV2B3 backbone, a 384 × 384 × 3 RGB input, FP16-compressed weights, and raw RGB values in the 0–255 range. The input and output tensors remain float32.

Intended use

• Research review and authorized evaluation of offline, human-reviewed maize leaf screening on compatible mobile or edge hardware.
• Triage research for field-scouting and agronomy workflows with a qualified human reviewer.
• Research on small, deployable agricultural vision systems.
• A baseline for future models trained and evaluated with geographically and seasonally representative field data.

Out-of-scope use

• Autonomous diagnosis or treatment selection.
• Pesticide, fertilizer, insurance, credit, or yield decisions without expert review and independently verified evidence.
• Diagnosis from images that do not clearly show a maize leaf.
• Use as a general-purpose crop, livestock, or environmental classifier.
• Redistribution, deployment, or derivative use while LICENSE provides no license grant.

Quick start

The following command is for authorized local review of this staging package:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python inference/predict.py path/to/maize_leaf.jpg

The command prints structured JSON containing the gate decision, class probabilities, whether Stage 2 ran, and the final screening result. See inference/README.md for the exact integration boundary.

Local package validation

Run the fail-closed, standard-library package validator and the split-tool unit tests before sharing an authorized review copy:

python scripts/validate_release.py
python -m unittest discover -s eval -p 'test_*.py' -v

These checks verify claims/status gates, audit and repair-plan invariants, artifact hashes, metadata posture, SVG safety text, and rendered-card structure. They do not substitute for corrected-split retraining, model evaluation, license clearance, or independent field validation.

Evaluation evidence and split warning

Research-preview maize screening model evaluated on documented development datasets and limited field samples.

A retrospective audit of the original v1.4 metadata.csv found 80,313 records, 68,786 source + group IDs, and 2,205 groups crossing the train/validation/test partitions. The recorded split is therefore not a leakage-safe independent evaluation, and its historical metrics must not be used as public performance or field-results claims. The machine-readable findings are in eval/development-data-audit.json, and the reproducible audit is in eval/audit_split_manifest.py.

The deterministic repair tool generated a group-disjoint assignment plan for all 80,313 records: 57,914 train, 11,107 validation, and 11,292 test, with zero cross-split groups and zero cross-split records. See eval/rebuild_grouped_manifest.py and eval/corrected-split-plan.json. This plan fixes the assignment design only: the packaged models have not been retrained or evaluated on it, and an independent Ghana field holdout is not complete.

For traceability only, the packaged metadata retains these legacy development metrics:

Stage	Legacy development metric	Historical value
Stage 1 triage gate	accuracy	0.9920846819877625
Stage 2 disease-pattern screen	accuracy	0.9934966564178467

These values are not release metrics or field-accuracy estimates. A fresh retraining run and corrected evaluation using the group-disjoint plan remain open release gates. Limited field samples were used only to exercise the end-to-end routing path; the recorded diseased sample was not independently labelled by an agronomist. See eval/SMOKE_TESTS.md.

Training-data record

No source images are distributed with this model repository. Local training records cite PlantVillage, PlantDoc, Ghana CCMT, seasonal maize-disease imagery, a multi-crop maize-disease archive, OSF northern-leaf-blight imagery, and MIT Indoor Scenes for open-set rejection.

The source trail, acknowledgements, and unresolved license questions are recorded in CITATIONS.md. That record does not itself grant permission to redistribute the weights.

Limitations and risks

• Disease symptoms can overlap with nutrient stress, chemical injury, weather damage, pests, senescence, and mixed infections.
• The model does not estimate causal disease severity, recommend treatments, or measure chlorophyll/SPAD.
• Image blur, glare, shadows, occlusion, unusual cultivars, early symptoms, and background clutter may cause rejection or incorrect classification.
• other_disease means the image appears abnormal but does not support a named class; it is not a diagnosis.
• Performance may vary substantially by camera, farm, region, season, cultivar, and crop stage; those variations have not been independently characterized.

Responsible evaluation

Present every output as a screening result with confidence and an explicit uncertainty state. Preserve the not_maize_or_unclear and other_disease pathways. Ask the user to retake poor images, route consequential cases to a qualified human, and log the model version and thresholds so errors can be audited.

Release and license status

This directory mirrors a publicly visible research-preview repository. Public visibility describes access only; it does not mean the model has public-release or production approval. The package includes two model artifacts, metadata, reference inference code, checksums, source citations, safety guidance, and editable announcement artwork.

License review is pending and LICENSE provides no license grant. This update does not change repository visibility. Do not redistribute or deploy the model artifacts until every item in RELEASE_CHECKLIST.md is resolved and an explicit compatible model license is approved. Any decision to restrict repository access requires separate release-owner and legal authorization.

Citation

@software{sinuosity_okuafo_maizeguard_edge_2026,
  author       = {{Sinuosity AI}},
  title        = {Okuafo MaizeGuard Edge v1.4},
  year         = {2026},
  version      = {1.4},
  note         = {Research-preview two-stage TensorFlow Lite maize screening model; independent field validation and license review pending}
}

Maintainer

Sinuosity AI — physical intelligence research for systems that work beyond the lab.