Dimensions
| Dimension | Question |
|---|---|
| Knowledge confidence | Does the model understand the general subject? |
| Context completeness | Are case-specific facts adequate? |
| Evidence support | Are material claims supported? |
| Local applicability | Does evidence fit crop, district, stage and practice? |
| Freshness | Could rules, weather or research have changed? |
| Alternative risk | Is another explanation close enough to matter? |
| Source agreement | Do observations and sources agree? |
| Consequence severity | What happens if the action is wrong? |
Modes
| Mode | Use |
|---|---|
| MEMORY_ONLY | Stable, low-risk explanation with enough context |
| ASK_FOLLOWUP | One farmer fact would materially improve the answer |
| LOCAL_RETRIEVAL | Departmental applicability or prior cases matter |
| CONTEXT_TOOLS | Weather, soil, terrain or history matter |
| FOCUSED_WEB | Recent, unusual, contested or missing evidence |
| DEEP_RESEARCH | High consequence, rare case or source conflict |
| HUMAN_ESCALATION | Evidence remains insufficient or authority is exceeded |
Mandatory verification
Regardless of subjective confidence, verify current product registration; dose, dilution, interval and pre-harvest period; current weather/outbreak status; local restrictions; emerging threats; and high-consequence claims.
Research budget
Define maximum rounds, sources, domains/source classes, elapsed time, token/cost ceiling and escalation conditions.
Calibration
Calibrate model confidence on held-out local cases using reliability diagrams, expected calibration error and action accuracy by band. Thresholds vary by consequence.
Stop rule
Stop when the next action has low expected information gain relative to cost and evidence supports the proposed action at the required risk level. Otherwise continue or escalate.