How the variety-attribution model works.

What goes in

Five sensor channels per node, the local microclimate, the variety planted, and the harvest outcome at season end.

• Soil moisture, NPK, pH, soil temperature, ambient humidity. Sampled every 15 minutes during the growing season.
• Microclimate priors from regional weather stations and SRTM elevation tiles.
• Harvest data is the slow loop. One row per (field, season, variety) at harvest.
• No personal data leaves the farm by default. Variety and outcome rows are opt-in for the data marketplace.

What comes out

Per-field, per-season ranked variety candidates with a predicted yield delta, a confidence band, and a vendor link. Plus the existing irrigation and fertiliser triggers.

• Predicted Δ yield: a point estimate with a low / medium / high confidence label.
• Vendor: a marketplace link to a seller that has the variety in stock for your region.
• Confidence drops sharply when fewer than three pilot farms in your microclimate have grown the variety.

How it is validated

Predictions are calibrated against published variety trial data and audited each season against held-out pilot fields.

• Held-out pilot fields are not used to train the season they are validated against.
• Trial-registry calibration removes the structural bias from breeders publishing best-case figures.

Source: AHDB Recommended Lists ↗

What it does not do (yet)

We are explicit about the limits because investors and growers should know them up front.

• No causal yield guarantees. Δ yield is a statistical estimate, not a warranty.
• No insurance-grade certification. We do not yet meet the actuarial standards for a parametric crop-insurance trigger.
• No automatic field action without grower confirmation. Auto-irrigate triggers a partner integration; the grower keeps the override.
• No livestock telemetry, no grain marketing, no tractor ISOBUS. These are not in scope.