Candidate: leaf rust
Observed signs match a backend-validated model within its supported crops. Shown as a ranked candidate for agronomist review — not a confirmed diagnosis.
R&D stage · honesty-gated · fail-closed
FieldPatho
Field-first plant-disease intelligence for phones. Its defining feature is honesty-gating: every result declares how it was validated — on-device, backend, candidate, reference, or unknown — so a suspected match is never shown as a confirmed diagnosis.
What it is
Lab- and booth-trained disease classifiers collapse on real in-the-wild field captures — and confidently mislabel, which is dangerous when a grower acts on a wrong call. FieldPatho is built the other way around: field-first, and structurally unwilling to present a guess, a textbook symptom, or a synthetic image as something the camera confirmed.
It ships as two scoped pieces: a buildable, signed iOS app (FieldPatho Scout) with guided capture, a manual "teach a crop or disease" workflow, and integrity-checked history; and a Python ML core with type-enforced honesty contracts, a crop-aware taxonomy, and a leakage-safe data spine. Disease inference is deliberately held fail-closed — it abstains until field-tuned promotion evidence exists.
Why it's different
The market is full of free "plant doctor" apps (Plantix, xarvio, Agrio, Cropwise) that surface a single confident label — and the independent literature already warns they hit 80–95% only in optimal conditions and degrade badly in real fields. None of them declares its evidence level or abstains. FieldPatho makes evidence-level and abstention a first-class, type-enforced part of every result, and stays vendor-neutral on inputs — built for agronomists and advisors who cannot act on an unverified call.
What we don't claim. There is no field-validated automated diagnosis here yet — every headline metric is promotion-blocked or fail-closed, and the healthy-vs-diseased gate is a release no-go. We sell the trust architecture and the capture workflow, not diagnostic accuracy. "Any disease on any crop" is a research direction, not a delivered feature.
Prepared demo
Every result wears its evidence level.
Illustrative result cards. The point isn't a confident label — it's that the app tells you exactly how much to trust what it shows, and abstains when it should.
Out of coverage
This crop/condition is outside the model's validated support. The app abstains and routes to manual capture or expert review rather than guessing.
Textbook symptom
A reference symptom image from the taxonomy — clearly labeled as a reference, never as something the camera observed.
Illustrative of the honesty-gating states enforced in the FieldPatho core. Live diagnosis is fail-closed pending field validation.
Start here
Advisors: help shape trustworthy field diagnostics.
If you run field scouting or advisory work and want a triage tool that abstains instead of bluffing, let's talk about a design-partner pilot.