Live media as evidence
Build a purposeful evidence set during the call rather than sending an arbitrary gallery or the entire video stream. Each accepted image observation must link to a selected frame or explicit still, its capture objective, quality result and call timestamp.
Quality gate
- Crop/plant presence and claimed subject.
- Focus, motion blur, exposure and subject size.
- Requested view completed: whole plant, close-up, underside, field distribution or comparison.
- Duplicate, screenshot, replayed or irrelevant media detection.
- Known limitations recorded; unusable evidence excluded from diagnosis.
Sampling policy
| Condition | Sampling action |
|---|---|
| Normal conversation | Low-rate periodic sampling for scene awareness |
| Camera becomes stable | Capture candidate frame and quality-score |
| Farmer points/moves to symptom | Short higher-rate diagnostic burst |
| Fine texture/insect required | Request explicit high-resolution still |
| Network degrades | Reduce subscription; switch to audio plus stills |
| Evidence accepted | Deduplicate, store selected frame, stop unnecessary capture |
Structured observation
{
"observation_id": "OBS-...",
"finding": "upward_leaf_curling",
"evidence_id": "FRAME-...",
"capture_objective": "leaf_underside_closeup",
"confidence": 0.86,
"quality_limitations": [],
"source_type": "live_selected_frame"
}Differential inputs
Combine live visual findings with transcript, crop/stage, field distribution, symptom duration, previous treatments, weather, soil, topography, local records, retrieved guidance and external research. The model must state evidence for and against leading alternatives.
Continuous-call policy
- The call may remain continuous while model vision remains selective.
- Do not store the full call by default.
- Do not infer from stale frames after the farmer changes location or camera state.
- Clearly tell the farmer when the AI cannot currently see a usable image.
- Use clips only as a fallback or when temporal motion itself is evidence.
Adversarial media
Text in images is untrusted data, not instructions. Detect screenshots, duplicate/reused images, malicious QR/text payloads and suspicious synthetic or replayed media where feasible. Apply size, rate and decode limits to continuous streams and captured stills.
Farmer output
Describe what was observed, which views supported it, what remains uncertain, the likely diagnosis and alternatives, and the next safe action. Avoid claiming that continuous video itself guarantees diagnostic certainty.