Not estimates.
Evidence.

All yield forecasts validated against 3-year historical actuals per crop type and microclimate zone before model deployment. Models must achieve R² ≥ 0.82 against holdout sets to enter production.

Every forecast output includes upper and lower confidence bounds derived from ensemble model disagreement. Bands widen automatically when input data quality degrades or sensor coverage drops below thresholds.

Weekly comparison of predicted vs. observed values at the field level. When prediction error exceeds ±8% of the rolling mean, the model is automatically flagged for recalibration and affected outputs are marked with reduced confidence.

Every data point displayed in FieldMind carries a provenance tag showing its source sensor, spatial resolution, temporal resolution, and time of last validation. No interpolated or estimated values are presented without explicit labeling.

Multi-source data is cross-referenced before ingestion. Satellite-derived soil moisture must correlate within ±12% of ground-truth sensor readings to be included in model training. Discrepancies trigger automatic exclusion and manual review flagging.