Frozen model
Stage 7D A0 FP32 ONNX, with the fixed Stage 7C operating thresholds.
Small AI / Browser-local observation
A coffee-leaf review dossier where the compact model proposes and the human signs the record.
Specimen intake
The image is processed on this device. RoyaCheck does not upload the raw image for inference, and this MVP does not save the photo with the record.
Image intake
Use a current landscape image with one complete target leaf visible. Images smaller than 224×224 are routed to not sure.
Model not loaded yet.
Model stamp
Human sign-off
Do not copy the AI automatically. Review the image and choose the formal observation yourself.
Saved locally
Text-only. Nothing is sent automatically.
Local record drawer
Saved observations are stored only in this browser's IndexedDB. They may be removed by browser storage eviction, private/incognito browsing, or clearing site data. This MVP provides no backup or export, and the list is text-only without raw photos.
Text-only. Nothing is sent automatically.
Fixed-string card
This MVP includes a fixed-string local-language scaffold connected to the current AI proposal and human disposition. Actual Lugisu/Lumasaba wording is pending fluent human validation and is not claimed in this build.
Localization limitation: fixed strings only; no chatbot, no free-form translation, and no claim of validated Lugisu/Lumasaba usability or field readiness.
Claim ceiling
Stage 7D A0 FP32 ONNX, with the fixed Stage 7C operating thresholds.
On the frozen BRACOL internal holdout, 136 target-class leaves were eligible for target-class routing. The model produced 106 confident target-class outputs and 104 of those 106 were correct. The visible-rust route captured 66 of 95 rust leaves, while 27 of 95 rust leaves were routed to not sure and 2 of 95 rust leaves were routed to no visible rust. Other-condition false-visible-rust routing occurred for 7 of 117 other-condition leaves.
Those numbers are internal-holdout evidence only. They are not field validation, not RoCoLe external-transfer evidence, and not challenge-set evidence.
RoyaCheck does not diagnose a plant, select treatment, recommend pesticide use, explain falling yields, or replace qualified human review. It does not verify that an image is a coffee leaf.
This technical MVP includes a local-language fixed-string scaffold. Actual Lugisu/Lumasaba wording is pending fluent human validation. Validated localized usability has not been established.