Safety objective
The platform supports agricultural decisions; it does not manufacture certainty. Safety is achieved by making evidence, uncertainty, authorization and escalation explicit at every case transition.
Risk classes
| Class | Typical case | System behaviour |
|---|---|---|
| Low | General education, capture guidance, prevention explanation | Answer directly with clear provenance. |
| Medium | Plausible diagnosis with incomplete evidence | Provide cautious analysis, ask targeted questions, or request better media. |
| High | Chemical treatment, severe crop loss, conflicting evidence, possible resistance | Verify local authority and exact controls; route for officer review where required. |
| Critical | Human exposure, suspected quarantine condition, widespread outbreak signal | Stop normal advice flow, display emergency/escalation instructions, alert authorized personnel. |
Action matrix
| Evidence state | Required action |
|---|---|
| Strong agreement, low consequence | Answer and cite evidence. |
| One material fact missing | Ask the highest-information-gain question. |
| Poor image or uncertain entity | Request recapture or confirmation. |
| Scientific evidence exists but no local authorization | Explain the distinction; do not issue an authorized treatment. |
| Exact product/dose/interval cannot be verified | Withhold the exact instruction. |
| Sources materially conflict | Deep research or expert review. |
| High-risk diagnosis cannot be excluded | Escalate even if another diagnosis is more likely. |
Safe language contract
Responses must distinguish:
- observation: what the farmer, image, sensor or record shows;
- hypothesis: what may explain the observation;
- evidence: which local or external source supports it;
- uncertainty: what remains unresolved;
- action: what is safe now;
- authorization: which exact recommendation is locally permitted;
- escalation: who will review the case and why.
Prohibited behaviours include false certainty, invented citations, unsupported dosage, silently overriding a local rule, and presenting a web finding as departmental authorization.
Safety gates
- Input gate: critical crop, chemical, quantity and location entities are confirmed.
- Media gate: image quality and crop relevance are sufficient.
- Evidence gate: material claims have attributable support.
- Policy gate: exact legal and treatment constraints are verified.
- Risk gate: consequence and alternatives are assessed.
- Response gate: output follows the structured response contract.
- Delivery gate: spoken output is generated only from the approved text.
Kill switches
Authorized administrators must be able to disable:
- a model or provider;
- a prompt or workflow version;
- a document, chapter or evidence source;
- a product rule;
- external web research;
- chemical recommendations;
- a crop, district or condition;
- media retention;
- or the complete advisory service.
A kill switch must be immediate, auditable, reversible and tested.
Incident severity
| Severity | Example | Initial response |
|---|---|---|
| SEV-1 | Harmful authorized instruction, privacy breach, critical outbreak failure | Disable affected path, preserve evidence, executive and regulatory escalation. |
| SEV-2 | Systematic diagnostic or retrieval defect | Constrain scope, rollback, review impacted cases. |
| SEV-3 | Isolated incorrect answer caught by safeguards | Correct case, classify error, add regression test. |
| SEV-4 | UX, latency or noncritical integration defect | Normal backlog and monitoring. |
Post-incident process
Preserve the complete decision record; identify the failing stage; determine affected cases; issue corrections where appropriate; create regression tests; update model, prompt, source, rule or workflow versions; document authorization to restore service.
Live escalation
For high-risk or unresolved cases, the workflow may request an officer to join the current room. The AI provides a pre-join summary and must not continue issuing disputed treatment instructions during a warm handoff. If no officer is available, provide a safe interim action and explicit follow-up route.