Public technical brief
Evidence ledger
Prediction surface
94%
production companion catch rate
Method
XGBoost forecast-bias correction + physical weather ensemble
Decision value
Frost protection starts with a production operating point, then carries the independent reproduction beside it for a current weather-skill read.
How the number is built
Bias-corrected forecast ensembles turn regional minimum temperatures into a prioritised frost-protection operating point.
Evidence signal
Regional frost-risk indicator
The evidence contract
Crop-cycle evidence
Prediction surface
7-day
probabilistic crop-cycle outlook
Method
Published disease formulas + gated ML classifiers
Decision value
Potato and open-field tomato carry crop-cycle-aware late-blight conditions through the same national prediction and decision paths.
How the number is built
Published disease formulas and gated classifiers create named potato and tomato blight windows across the crop cycle.
Evidence signal
Formula-based disease pressure
Summer-hazard evidence
Sunburn and splitting rain event results remain visible as research-only monitoring evidence unless they exceed 90% recall while staying below 5% false alarms on the untouched chronological test. Research numbers are not live alert claims until that gate passes.
Prediction surface
90.2%
regional proxy recall · not observed hail
Method
Environmental ML over a convective-condition proxy
Decision value
The current seven-day output gives growers a regional convective-event warning surface, with held-out evidence retained beside the observed-hail programme.
How the number is built
Environmental ML scores regional convective conditions on a roughly 10 km grid and carries the highest forecast probability across seven days.
Evidence signal
Regional convective-risk indicator
Evidence pathways
Harvora delivers the screening, planning and prioritisation layer for each published surface today. Every signal carries its method, evidence tier and uncertainty; orchard-specific observations can add calibration depth where greater local specificity is valuable.
A weather-derived operating signal can be a proxy: an indirect signal used to estimate a harder-to-observe outcome. Harvora uses it as a practical screening and prioritisation layer; direct observations can add orchard-specific calibration where greater local specificity is valuable.
Growers can log what happened in the block, while partner records arrive through the same audited evidence path. Verified field outcomes are reviewed for calibration, then can inform future model versions, helping Harvora build a stronger local track record over time. A live model changes only after the new version passes its evaluation gate and receives explicit release approval.
Regional minimum-temperature conditions become an operating risk indicator for prioritisation.
Evidence expansion · The regional frost indicator is ready for prioritisation; independent block observations can add direct frost calibration depth.
A physical fruit-surface-temperature proxy sharpens the regional heat-environment read.
Evidence expansion · The regional heat screen is ready for canopy, shade and harvest planning; an instrumented fruit-temperature layer can add orchard-specific calibration.
Environmental ML over regional convective conditions provides a warning surface for protection planning.
Evidence expansion · The regional convective-risk screen is ready for protection planning; an independent observed hail holdout can add direct-event validation depth.
Thermal-time accumulation provides uncertainty-aware timing for harvest planning.
Evidence expansion · The heat-unit timing surface is ready for harvest planning; measured maturity, pick-date and packout records can add orchard-specific calibration.
Rule-based weather suitability and forecast uncertainty identify selective operating windows.
Evidence expansion · Product-specific efficacy and label-compliance registries can add a deeper decision layer alongside weather suitability.
Temperature, wind, rain and daylight become an operational bloom-weather window.
Evidence expansion · The bloom-weather window is ready for labour planning; fruit-set and pollinator-activity observations can add local calibration.
Crop thresholds and forecast uncertainty create a day-level protection and labour signal.
Evidence expansion · The wind watch is ready for day-level protection and labour planning; observed damage can add local calibration depth.
FAO-56 water-balance modelling and calibration turn weather into a block planning signal.
Evidence expansion · Connected physical probes can add block-level calibration depth to the weather-derived water-balance signal.
Published disease methods and guarded classifiers turn crop conditions into probabilistic risk windows.
Evidence expansion · The disease pressure surface is ready for probabilistic scouting; independent infection observations can add local calibration depth.
ENSO-conditioned history adds context before the operational forecast window opens.
Evidence expansion · Historical dryness context is ready for seasonal planning; calibrated forecast evidence can add another layer of decision depth.
Rain memory and documented drainage heuristics support practical access planning.
Evidence expansion · The access screen is ready for work-window planning; measured soil trafficability can add site-specific calibration.
Physical field evidence path: the Soil Scouter™ physical sensor is designed to bring observed soil conditions into block-level calibration and decision context.
Evidence expansion · Physical sensor evidence can add block-level calibration depth to the platform's decision context.
The Harvora Microclimate Station™ is designed as a site weather reference, paired with Soil Scouter™ readings distributed through the blocks.
Evidence expansion · Connected, quality-checked station observations can add local weather evidence for future block-level calibration.
Rain exposure during ripening becomes a timely planning signal.
Evidence expansion · The rain-exposure surface is ready for ripe-fruit planning; splitting and packout outcomes can add site-specific calibration.
Regression fixtures guard code changes only; they are not independent orchard outcome evidence.
