One photograph. A plant in a garden bed, early evening, handheld phone. Not a benchmark image — the kind of picture a grower actually takes. Nothing below depends on which plant it is: it is what happens to any photograph.
9 models looked at this plant and named 7 different diseases. The most confident was 100% sure of bacterial spot.
Every one of these is a real, trained, shipped model, and every one of them answered. None of them said “I am not sure”, because none of them can.
Why every one of them answered
A classifier trained on seven cherry diseases cannot represent “this is not a cherry tree”. Its probabilities must sum to one, so it always answers, confidently.
The obvious reply — then just run the right model — assumes you know the crop. On this photograph our own crop models do not agree:
What it showed instead
“No coverage yet.”
A refusal that names what the app lacks, then asks for the one thing that would fix it: tell us what disease you see.
One of those 9 answers appears further down, at 62%, under two Experimental pills. The other eight — including bacterial spot at 100% — are never shown.
It is not hiding what it knows. It is deciding what may carry the weight of a diagnosis, and demoting the rest.
Why we think this is the thing to build
Our position, not a measurement. The numbers inside it are measured.
The evidence behind all of that
All of it read from artifacts the phone wrote at capture time. Nothing was re-run to make this page.
Every answer, for every photograph — the full grid
20 scans, 22 trained heads, 320 committed answers. Fewer than 20×22, because the two visits were three weeks and one app version apart and did not carry the same models.
healthynot a crop I knowbacterial spotblack rotblight or necrosis likecommon rustearly blightgray leaf spotlate blightleaf moldleaf scorchleaf spot likemosaic or virus likemosaic virusnorthern leaf blighta crop in the catalogpowdery mildewpowdery mildew likerustscabseptoria leaf spota crop in the cataloga crop in the catalogyellow leaf curl virus
Read the rows. One disease model calls this plant powdery mildew 20 times out of 20 — a single colour straight across. Another calls it bacterial spot 20 out of 20. Neither has ever seen this plant in training, and neither can say so — and neither would behave any differently on a photograph you took this morning.
What this page does not claim
- It is not an accuracy measurement. These captures carry no disease ground-truth labels. Nobody has adjudicated what is actually wrong with these plants, so which model is right is unknown — the disagreement is a fact about the models, not about the plants.
- The models here were asked the wrong question. A cherry model run on a plant that is not a cherry tree is out of its domain by construction. That is the point being demonstrated, and it is not a measurement of how good any of these models is at its own job.
- Refusing is not the same as being right. It is correct behaviour given the app’s coverage, not evidence that its coverage is adequate. This crop is not covered yet.
- One garden, one phone, one operator. Four plants across two visits is a demonstration, not a sample.
- Development preview. No model in this app is validated on device, and nothing here should be used to treat a plant.
Rebuild with python scripts/render_app_demo.py. The
machine-readable summary is beside this file as app_demo.json; the
method and caveats are in reports/track_app_demos.md.