Proof, not promises
Every Harvora surface ships with a measured accuracy stat on real New Zealand data — benchmarked July 2026, provenance documented, caveats stated up front. This page is our technical brief, in the open.
The average NZ growing location faces 23.4 damaging frost nights a season, with exposure running to tens of thousands of dollars per hectare.
Catches 94% of frost nights at the economically-correct alarm setting. Production overnight-low forecasts run 15–18% sharper than the public forecast (1.38 °C RMSE at 24 h). Climatology-style tools miss 66% of frost nights.
Potato and open-field tomato, modelled at full platform depth
Crop-cycle context now flows through phenology, irrigation, disease, harvest readiness and economics. Four guarded blight models and three insect lifecycle stacks extend the same probabilistic serving contract across all 9 supported crop families and 13 NZ regions.
Potato and open-field tomato now carry crop-cycle-aware late-blight intelligence through the same national prediction, finance and UI paths as the established crop families.
A 1,000-sample surface blending 60% formula with a 40% guarded classifier when the packaged artifact schema and hash match. The engineering corpus contains 6,510 generated feature rows across the four annual blight models.
Seven more surfaces, covering the risks summer brings
Hail, sunburn, splitting rain, fire weather, crew heat safety, seasonal dryness and field access — the perils that arrive with heat, ripe fruit and harvest machinery. Each one beats the raw forecast or climatology on held-out data, and each card states exactly what was measured.
Hail can wipe out a season's fruit in minutes — and no forecast anywhere can call it block by block. Knowing when the atmosphere is primed is the edge that matters.
Machine-learned models read the convective environment (CAPE, lifted index) up to 7 days out, beating the raw forecast at every lead across a 58,526-row corpus spanning 13 regions. On the days it flags, real convective events occur ~25× more often than chance — with fully calibrated probabilities (Brier 0.0095, ECE 0.008).
Numbers you can trust exist only where a company says what it can't claim
Everything above is measured against real, independent data. A few things aren't provable yet — and rather than blur the line, we publish it. This discipline is the product.
- 01
Our disease classifiers (AUC 0.91–0.98) faithfully reproduce published agronomic risk formulas — that is reproduction fidelity, not field-validated infection prediction. Real validation lands as grower-confirmed outcomes accumulate.
- 02
Yield-from-satellite (NDVI) is not field-validated for any crop yet. Five of nine crop models are flagged as self-referential in their own metadata — in capitals. Tested honestly against 658 real Sentinel-2 observations, the trajectory lands at climatology (MAE 0.066 vs 0.068) — no added skill yet on two blocks.
- 03
A runoff model scoring a 92% improvement sits unshipped in our codebase, held at deployable:false — it learned to mimic physics labels, so the score isn't real-world skill. We'd rather hold it back than dress it up.
- 04
No named-competitor benchmark yet. When we publish one, it will be run the same way as everything on this page: real data, stated provenance.