We publish skill, not promises.
The platform is in closed beta with design partners. Published skill figures will appear on this page when the beta ends. Until then, this page describes how every module is validated, and where it is weak.
What every module is compared with.
Skill is how closely a module’s output matches independent measurements it has not seen. Each module is compared with the reference that fits what it measures.
Crop yield
Compared with official yield statistics for seasons the model has not seen.
Climate and water balance
Compared with national weather-station records. This also checks the climate inputs behind every module.
Groundwater
Compared with readings from national borehole networks.
Rivers
Compared with national river-gauge records, river by river.
Reservoirs
Compared with records of water level and storage, where such records exist.
Ground subsidence
Compared with independent ground-based reference measurements, where such points exist near the site.
The rules before release.
The same rules apply to every module.
- Data the model has not seen. Comparisons use seasons, years and locations the model was not trained on.
- Skill for each basin or district. Results are reported separately, not as a single average.
- Uncertainty is checked too. Each stated range is tested for calibration against measurements the model has not seen. Every published number states its uncertainty and its limits.
- A gate before release. A module stays in pilot until it meets its accuracy targets. An indicator is released only after its skill for that basin or district has been published; below the threshold it is flagged as low confidence or withheld.
- Repeated checks. Comparisons are repeated as new reference data become available.
Known limits.
Where a module is weak, this page says so.
- Modules in pilot. Every module is currently in pilot and available to design partners only. Accuracy targets and measured results will be published together.
- Small, fragmented objects. Parcels, river reaches and reservoirs smaller than the resolution of the sensors carry wider uncertainty. This is flagged for each location.
- Sparse reference networks. Where boreholes, gauges or stations are few, such as in mountain basins, a module is harder to verify. Results there carry a low-confidence label.
- Small samples. Where only a few reference sites exist, validation rests on a small sample and is described as such. Coverage widens only as more reference data become available.
Join the closed beta.
Orbispect is onboarding design partners right now and building a waitlist for early access. Leave your work email and we'll reach out once the beta opens for your region. No sales deck.