Public prediction catalogue
Each surface explains its intended decision, method, evidence status, uncertainty and non-claim in crawlable public HTML.
ESTIMATE
Harvora forecasts overnight frost risk for every block as a probability. In an independent held-out benchmark across 2,356 New Zealand nights, it caught 90.48% of frost events (76 of 84) with a 3.79% false-alarm rate at the stated operating point.
Read the evidence brief ↗ESTIMATE
Harvora finds orchard spray windows worth using: of 308 windows it selected across 13 regions over two held-out seasons, 92.2% stayed rain-free — against 82.6% of unselected candidates. A selective timing signal for operational planning; the product label and local conditions always govern the final call.
Read the evidence brief ↗ESTIMATE
Harvora tracks each block's soil-water balance and signals irrigation timing before stress sets in. Where a prior-day moisture observation exists, the moisture-state call reaches 96.54% classification accuracy across 3,211 held-out region-days; the weather-only FAO-56 path stays dependable at root-zone r = 0.799.
Read the evidence brief ↗ESTIMATE
Harvora projects each block's harvest timing, and it sharpens as the season runs: median timing error is 1.5 days by day 60, down from 4.5 days at day 30, with the uncertainty band covering the regional spread 95.3% of the time.
Read the evidence brief ↗SOURCED
Harvora turns published disease science into block-level infection-risk windows — twelve crop-disease models, each naming the method it runs, with the machine-learning blend held to ROC-AUC 0.90–0.99 against that science across 18,064 held-out rows.
Read the evidence brief ↗ESTIMATE
Harvora counts bee-flyable hours across each block's flowering window, so bee placement and bloom weather get planned together. The bee-flyable weather classification reaches 91.7% accuracy (86.4% recall, 86.9% precision) across 2,544 spring hours.
Read the evidence brief ↗ESTIMATE
Harvora watches for damaging wind before it arrives: the watch caught 90.9% of damaging-wind days (209 of 230) across 1,378 days in 13 regions, with a 12.7% false-alarm rate — a day-level protection and labour signal, tuned to favour catching events over staying quiet.
Read the evidence brief ↗ESTIMATE
Harvora screens hail-conducive conditions at each block's coordinates out to seven days, catching 90.2% of held-out convective events (55 of 61) in regional model analysis. Hail is the hardest surface in the catalogue to validate — so its boundary is the most explicit.
Read the evidence brief ↗ESTIMATE
Harvora times the windows that matter for three pests — codling moth, potato tuber moth and tomato-potato psyllid — as 7-day probabilistic stage windows. Each lifecycle stack clears its AUC, Brier and calibration gates, and is built to direct scouting, not to assert a pest is present.
Read the evidence brief ↗ESTIMATE
Harvora reads the season's dryness pressure from 41 years of regional rainfall history conditioned on the current ENSO state — climatology-informed planning context across 13 representative regions, published for exactly what it is.
Read the evidence brief ↗