A winter storm cloud over a mountain range is usually inefficient: the flow forced up the barrier condenses supercooled liquid water — droplets colder than freezing that stay liquid because they have nothing to freeze on — faster than natural ice processes can convert it to snow. The leftover water evaporates on the lee side and is lost. Glaciogenic seeding releases silver iodide, whose crystal lattice is nearly a twin of ice, into that supercooled cloud. Droplets freeze on the AgI nuclei, the new crystals grow by vapor deposition and riming, and they fall out as additional snow in the target watershed. Seeding does not create clouds. It raises the precipitation efficiency of clouds that already exist — which is both why it can work and why it can never make weather out of a dry sky.

The physics sets hard boundaries. AgI barely nucleates warmer than about −6 °C and works well from about −8 °C down to about −25 °C, with a sweet spot near −15 °C; colder than that, nature is already making ice without help. The other family — hygroscopic seeding, salt flares burned under warm convective cloud bases to jump-start droplet coalescence — is the method used in Texas, the UAE and Thailand. Its statistics are suggestive, but its causal chain has never been closed observationally the way the winter chain has. That distinction structures everything on this site.

For sixty years the field's embarrassment was that nobody had ever watched seeding work end to end. The SNOWIE campaign (January–March 2017, 24 intensive observation periods over Idaho Power's operational area) did it: research aircraft and mountaintop radars tracked AgI from release, to ice initiation in supercooled cloud within about 30 minutes, to growth, to snow on the ground (French et al. 2018, PNAS) MEASURED. The attribution device was geometry: the seeding aircraft's crosswind legs produced trains of parallel radar reflectivity lines — the zig-zag signature — that nature cannot draw.

The follow-up (Friedrich et al. 2020, PNAS) put numbers on three cases: about 100, 196 and 275 acre-feet of liquid-equivalent snow per seeding event, spread over roughly 2,000 km² MEASURED. Real water — and modest water: a few hundred events at that yield is what stands behind program-scale claims of hundreds of thousands of acre-feet, and that extrapolation is model-based, not measured. Only three of the 24 periods were quantifiable at all, because the signal is only separable when natural snowfall is nearly absent.

Read the ladder honestly and the picture is coherent: the mechanism is proven, the per-event effect is measured and small, the seasonal effect is real but bounded by how rarely conditions are seedable, and the operational claims sit at the optimistic end of what the physics permits. The National Academies' 2003 verdict — "no convincing scientific proof of efficacy" — is what SNOWIE and Wyoming were designed to answer, and they answered it: proof of mechanism, yes; proof of large water yields, still no. Anyone selling more certainty than that, in either direction, is selling.

The newest rungs are the independent ones. In 2026, Bryce Judy's seeding-verification study rebuilt Utah's evaluation from public data alone and showed that the storm-scale "seeding effect" collapses from +335% to +1.5% as selection is designed out, while 46 winters of gauge data put a ±6% noise floor under any seasonal claim. That is precisely the ladder's shape recovered by an outsider from the state's own records — worked in full as Forensic Case 002.

The tool behind nearly every modern seeding evaluation is WRF-WxMod: NCAR's silver-iodide parameterization inside the Thompson aerosol-aware microphysics of the WRF model (Xue et al. 2013). It simulates the release, dispersion and nucleation of AgI from real generator sites and flight tracks, run as paired control/seeded simulations whose difference is the seeding effect. Idaho runs it in real time at 1.8 km as go/no-go guidance; Wyoming, Montana and Australia's Snowy Mountains use it for assessment.

Worth knowing: WRF-WxMod is not in the public WRF release — UCAR holds a trademark on it and licenses it — but the governing equations are fully published, and the base model is public domain. An open re-implementation, exercised first in idealized two-dimensional mountain flows and then in real cases, is on this site's roadmap; results from it will always wear the MODELED chip.

"Best time to seed" has a precise operational meaning. Seasonally: roughly November through April in the western U.S., because the water has to land as snowpack to be stored, and programs suspend when the snowpack is already fat (Colorado's permits force it; Utah suspends above the 95th percentile of snow-water equivalent). Within a storm: the seedable water is concentrated in the shallow, liquid-rich pre-frontal orographic phase and the unstable post-frontal westerlies — not the deep storm core, which is already efficiently snowing on its own. A duty meteorologist watches radiometer liquid water, the 700 mb temperature and wind, and model guidance, and lights generators half an hour to two hours before the plume is needed over the crest. Wyoming's trigger was 700 mb at −8 °C or colder with radiometer-confirmed liquid; Nevada's is crest-level −5 °C with winds under about 60 knots RECORD. All of which is to say: the deploy decision is a forecast problem — and forecasting it publicly is what Seedability is for.