Okuafo MaizeGuard Edge v1.5

00 / Overview

Calibrated two-model maize symptom screening system

A public development candidate combining categorical and specialist FP16 TensorFlow Lite models with locked fusion calibration and explicit abstention for offline, human-reviewed maize-leaf symptom screening.

01 / Status

Public access and evidence basis.

Public repository visibility identifies access, not production approval or an open-source license grant. Review the current repository terms before copying, deploying, redistributing or modifying the artifacts.

02 / Intended capabilities

Published workflow and intended capabilities.

• Screen maize-leaf symptom images across five documented pattern categories
• Combine a categorical model and a specialist model through a locked fusion calibrator
• Abstain when calibrated confidence is below the published acceptance threshold
• Run locally with FP16 TensorFlow Lite models on compatible edge hardware
• Preserve model scores, confidence and uncertainty for qualified human review

03 / Research objective

Research objective

Improve confidence handling in focused maize symptom screening by combining two model perspectives, calibrating the final decision and retaining abstention as a required outcome.

04 / Model files

Model files

• Categorical model: 24.5 MiB FP16 TensorFlow Lite artifact
• Specialist model: 24.5 MiB FP16 TensorFlow Lite artifact
• Locked dependency-free fusion resolver and calibration contract
• Source Keras checkpoints, manifest, checksums and development evidence

05 / Inputs

Inputs

• 384 x 384 RGB maize-leaf symptom image
• Float32 tensor derived from the image in the raw 0-255 range

06 / Outputs

Outputs

• Bacterial leaf streak, common rust, gray leaf spot, northern leaf blight or other disease
• Calibrated fusion probabilities, confidence, margin and acceptance decision
• Explicit abstention when the locked confidence threshold is not met

07 / Architecture overview

Architecture overview

A categorical classifier and a specialist confirmation model produce separate five-class scores. A locked degree-two one-vs-rest logistic calibrator fuses those scores and returns either the selected symptom category or an abstention for review.

08 / Development evidence

Development evidence

The Hugging Face package publishes five-fold out-of-fold development evidence, per-class measurements, a full development-set saved-model-to-TFLite decision-parity audit and the exact calibration threshold used by the runtime. These are internal development measurements, not field-validation claims.

09 / Known limitation

Known limitation

Gray leaf spot recall remains below the strict internal release gate. The candidate must retain abstention and qualified human review and must not be represented as independently validated, agronomist validated or production approved.

10 / Edge verification

Edge verification

The published FP16 TFLite artifacts use built-in operations only, produced deterministic repeated-batch outputs in the recorded audit and matched the saved-model fusion decisions across the documented development set.

11 / Data and provenance

Data and provenance

Training images are not distributed with the model. The public package includes artifact hashes, development evidence and source contracts; access to the model does not grant rights to third-party source data.

12 / Safety considerations

Safety considerations

Treat every accepted output as screening support, not a confirmed diagnosis. Abstentions and uncertain cases require human review. Treatment, pesticide, insurance, credit and other consequential decisions require a qualified agronomist, extension officer or appropriate specialist.

13 / Intended use

Intended use

Offline, human-reviewed maize-leaf symptom screening; controlled user validation; field-scouting research; and evaluation of calibrated edge-classification workflows.

14 / Out-of-scope use

Out-of-scope use

Autonomous diagnosis, treatment selection, pesticide or dosage advice, replacement of professional agronomy review, or consequential agricultural and financial decisions.

15 / Access and licensing

Access and licensing

The public Hugging Face package licenses model weights under CC BY 4.0 and included code under Apache 2.0. Third-party data and other materials remain governed by their own terms.

16 / Citation

Citation

Sinuosity Physical Intelligence Lab and OkuafoAI. Okuafo MaizeGuard Edge v1.5. Public development candidate, 2026.

17 / Version history

Version history

Version 1.5 was published on Hugging Face on 25 July 2026 with two FP16 TensorFlow Lite models, source checkpoints, a locked fusion runtime, checksums and development evidence. Version 1.4 remains available as the preceding two-stage screening record.

18 / Responsible disclosure

Responsible disclosure

Potential safety, provenance, licensing or security concerns may be reported to hello@sinuosity.ai.