Okuafo MaizeGuard Edge v1.5
Public development candidate — user testing required.
This model has not been independently or agronomist validated. It is not approved for production use or autonomous diagnosis.
MaizeGuard Edge v1.5 is a two-model FP16 TFLite screening pipeline for five maize leaf symptom classes:
- bacterial leaf streak
- common rust
- gray leaf spot
- northern leaf blight
- other disease
The categorical and specialist models are combined by the locked calibrator in
runtime/fusion_calibrator.json. The runtime returns __abstain__ when the
calibrated confidence is below 0.7360778873342575. An abstention must be sent
for human review rather than converted into a diagnosis.
Development evidence
All headline metrics below are internal five-fold out-of-fold development measurements, not field-validation or production claims.
| Metric | Result |
|---|---|
| Raw accuracy | 89.74% |
| Raw macro F1 | 87.41% |
| Selective accepted accuracy | 97.23% |
| Selective coverage | 80.09% |
| FP16 TFLite decision parity | 1,306 / 1,306 |
The exported TFLite models use built-in operations only and produced deterministic repeated-batch outputs in the recorded parity audit.
Known limitation
Gray leaf spot raw recall is 59.46%, below the strict 78% release gate.
The only failed development gate is oof_named_recall_gte_0_78. This candidate
must therefore retain abstention and human review and must not be described as
validated.
Files
models/: the categorical and specialist FP16 TFLite models plus their source Keras checkpoints.runtime/: the dependency-free fusion resolver, locked calibrator, class contract, and one-image test CLI.evidence/: the development report, 1,306-image TFLite parity report, quantization-threshold adjustment receipt, and development-review exemption.manifest.json: SHA-256 inventory and release-state flags for the exact candidate.
Test one image
python -m pip install -r requirements-test.txt
python runtime/run_fusion_tflite.py \
--bundle . \
--image /path/to/maize-leaf.jpg
Treat every output as screening support for a qualified human reviewer.
Release state
- Development review exemption: active
- Ready for user validation: yes
- Public development publication: authorized
- Public model-weight licence: CC BY 4.0
- Independently validated: no
- Agronomist validated: no
- Validated release eligible: no
- Production use approved: no
Model weights are licensed under CC BY 4.0 in LICENSE. Included code is
licensed under Apache-2.0 in APACHE-2.0.txt. The development evidence retains
its original release-state flags to preserve the audit record; public
availability does not convert those measurements into validation.
- Downloads last month
- 30