Purpose
Human review is a designed service, not an emergency inbox. It must have ownership, queues, response targets, authority boundaries and measurable quality.
Review roles
| Role | Authority |
|---|---|
| Field officer | Validate field context, request evidence, provide routine local follow-up. |
| Agronomist | Resolve uncertain diagnoses, approve complex scientific conclusions. |
| Regulatory specialist | Interpret product registration, dose, interval and legal restrictions. |
| Duty supervisor | Triage critical cases, reassign work, invoke operational controls. |
| Knowledge steward | Correct source metadata, withdraw guidance and initiate re-indexing. |
| Safety lead | Review serious failures, authorize restrictions and restoration. |
Queue design
Cases are routed by:
- risk and urgency;
- district and crop;
- required expertise;
- evidence completeness;
- farmer accessibility;
- language or dialect;
- service-level deadline;
- and current reviewer capacity.
A case may have parallel scientific and regulatory reviews, but one accountable owner must control the final disposition.
Suggested service levels
| Queue | Target |
|---|---|
| Critical exposure or quarantine concern | Immediate emergency workflow; not a normal asynchronous queue. |
| High-risk treatment approval | Same operating day during staffed hours. |
| Uncertain diagnosis with active crop damage | Within one working day. |
| Routine follow-up | Within two working days. |
| Knowledge correction | Prioritized by affected-case count and severity. |
Final service levels require departmental approval and staffing analysis.
Reviewer workspace
The interface should show:
- farmer narrative and normalized transcript;
- original and annotated media;
- environmental and field context;
- ranked hypotheses and alternatives;
- local evidence and current policy results;
- external research with source quality;
- model uncertainty and missing evidence;
- previous reviewer actions;
- a structured decision form.
Reviewers must be able to approve, amend, request more evidence, escalate, reject, or mark the case outside platform scope.
Disagreement handling
- Preserve each independent opinion.
- Ask whether disagreement is factual, interpretive, regulatory or caused by missing evidence.
- Obtain additional evidence where it can resolve the issue.
- Route unresolved scientific disagreement to an agronomist and regulatory disagreement to the designated authority.
- Record the final rationale and dissent.
- Convert recurrent disagreements into evaluation cases or knowledge-policy revisions.
Capacity model
Estimate demand using:
daily completed cases × escalation rate × average review minutes ÷ productive reviewer minutes
Separate rates by risk class. Add seasonal peaks, absenteeism, rework, travel or farmer callback time. Do not assume reviewer capacity is unlimited.
Feedback integrity
Reviewer decisions are not automatically ground truth. Quality controls should include:
- reviewer calibration exercises;
- second review for sampled or high-risk cases;
- agreement measurement;
- rationale completeness;
- conflict-of-interest controls;
- and adjudicated gold-standard sets.
Only reviewed, de-identified and appropriately authorized cases may enter model or retrieval evaluation datasets.