Teaching an offline plant doctor to spot rust

Prepared 31 October 2026. These notes separate three kinds of evidence. The model comparisons describe research experiments; they do not establish the model used in a released app version.

1. Recorded experiments

The household results come from saved benchmark reports, not new full-pipeline runs. Exact counts were reconstructed from reported denominators and rounded percentages. Every denominator is at most 28, and each reconstructed count is uniquely determined. The article includes the full count table. The reported rust progression (0/24, 7/24, 15/24) is a category-level comparison; it does not establish before-and-after predictions for the individual hero photo.

These reports use cached whole-photo embeddings with mocked native-channel similarities. They measure the first Disease-tab label, the rules path, and the first suggestion of guided diagnosis. They do not measure a full guided session, live reticle crops, device latency, or biologically confirmed diagnoses.

The PlantDoc six-class comparison in Figure 01 comes from a research experiment note dated 24 October 2026. It has not been retrained or independently reproduced for this publication. Reported top-1: 68.6% for trait similarities versus 80.1% for image embeddings; n=236. This is a six-group task, distinct from an earlier 27 crop-and-disease-label experiment.

Same encoder, six grouped classes, 236 test photos; recorded research experiment.

The historical training procedure has methodological limits: trait standardization uses the whole dataset, and preprocessing is not refitted independently inside every cross-validation fold. Treat the historical comparison as exploratory evidence. Future comparisons should fit preprocessing inside folds and leave the final test untouched.

The v1 research head was subsequently refitted on all PlantDoc images. The v2 trainer also uses PlantDoc, including its original test split. Neither fitted head may now be evaluated on PlantDoc and described as held out.

2. Fresh research head-only inference

The replay uses saved classifier weights and 286 previously computed, 512-dimensional image embeddings, aligned with the reviewed test manifest. It evaluates a linear condition head rather than a complete plant-diagnosis system.

The replay normalizes embeddings and stored class vectors, computes logit = cosine × norm + bias, applies softmax, and names the winning class only if it has an enabled threshold and meets that threshold. It uses the fitted head’s rounded weights. It does not train anything, run the image encoder, or execute the rule fallback.

Head-only hits include 15/27 leaf miners, 15/24 rust, 13/26 scale insects, 6/10 thrips and 0/19 aphids. The head abstains on 24/25 healthy controls, which is not proof that those plants are healthy. The recorded combined pipeline has different counts because it includes fallback rules.

The six photo examples are deliberately selected: two matching labels, three wrong labels, and an abstention. They are not a random sample or a calibration study. Softmax scores are model outputs, not calibrated probabilities of biological diagnosis.

The original full-pipeline evaluation requires the separate app source, prompts, replay cache, and corresponding model snapshot. It was not run while preparing this article. No claim of exact historical app-source revision parity is made.

3. Fresh metadata audit

The audit checks the research manifests:

• 445 original test entries, 286 visually reviewed entries, 2,970 training entries.
• Zero shared photo IDs and observation URLs between the original test manifest and training.
• 205 shared observer IDs across the original test manifest and training; 150 for the reviewed subset. Those 150 account for 214 reviewed photos.
• Zero shared (issue, observer) pairs in the reviewed subset versus training. Per-issue exclusion is weaker than global observer exclusion.
• Two reviewed downy-mildew entries are tagged Phytophthora ramorum. Thirty of 153 training entries in that bucket are Phytophthora taxa. The condition mapping needs correction before interpreting that row as a clean downy-mildew evaluation. UC IPM: ramorum blight and sudden oak death.
• Current iNaturalist training counts include 12 spider-mite photos, 2 grey-mould photos, and no anthracnose photos. These are counts from the inspected manifest; older notes mention a different spider-mite count.

This audit does not detect perceptual duplicates, confirm all taxa, certify healthy controls, or estimate population accuracy. Research-grade observation identification is not lab confirmation of disease; cultivated healthy controls include casual-grade observations. The set was reused during development and is not a final untouched evaluation.

Photo sources and attribution

Photo rights remain with their credited contributors. Each of the six examples in the article’s photo-case section includes its contributor, observation link, and original photo licence record. The photos were resized; the composite adds letterboxing, text, and model outputs. No disease features were added or removed. CC BY licence versions were not stored in the cache; refer to the linked original photo licence records.

Proposed tests

A complete guided session has not been shown to improve diagnosis in these experiments. A proposed sequence checks image quality, records visible findings, inspects additional leaves and undersides, asks about relevant care changes, then re-ranks candidates or leaves the result uncertain.

Design sketch; diagnostic improvement remains unmeasured.

Removing green pixels is not established leaf segmentation: brown pots survive a colour threshold, variegation can resemble damage, and pests can live on green tissue. Missing an insect in one frame does not establish its absence.

Visual patterns also have exceptions. Bacterial leaf spots can begin circular and become angular; geometry alone cannot reliably separate bacteria from fungi. UC IPM: bacterial leaf spots.

Future validation should fit preprocessing inside training folds, group observers across issues, review duplicates and labels, and reserve an untouched final set. Care-problem claims need adequate watering and light examples. No new training run or live-phone evaluation was performed for this article.