Dynamic Spatial Microclimate & User-Mesh Forecast Engine

National models smooth their answer across a wide cell and a coarse vertical layer. On a ridge, in a hollow, or beside a large body of water, that smoothing is the entire error. This engine treats each local site as its own correcting node rather than a sample of a regional average.

Three specific failures are named and addressed: water effects applied without checking which way the wind is actually blowing, nighttime lows treated as a flat offset from the day's high, and humidity reported without the canopy conditions that actually decide whether mould takes hold.

Every body of water within 100 miles is evaluated. A lake only counts when the wind is arriving from it, inside a 45-degree cone; the effect then scales with the water-to-air temperature difference, the cosine of how far off-centre the alignment is, the inverse of distance, and the log of surface area.

That is why a lake 70 miles away can move the number one day and be ignored the next. The page names the water it used, the bearing, and the degrees it moved the reading.

The overnight low is not the day's high minus a constant. The recent 72-hour range is compared against the 90-day range, and the difference is amplified by its own square so that a genuinely widening or collapsing spread is carried through rather than averaged away. Ocean state and cloud cover both bend the result.

The dew point is an absolute floor: air cannot cool past the point where condensation releases its heat back, so the projected low is never allowed below it.

Vapour pressure deficit and continuous leaf wetness are combined with a temperature response centred on the range spores prefer, then lifted where a cool, saturated night suggests an inversion holding fog in place. The output is a single pressure figure with a plain instruction attached.

Home weather stations connect over USB, serial, Bluetooth or the local network. The handset strips serials, device identifiers and precise position before anything is sent, readings are bucketed into roughly kilometre-wide cells, and a consensus filter drops sensors that have gone bad before the rest are merged.

The structure above is disclosed in full. The phase multipliers, stretch ratios and seeds that make it produce a specific number are not, and never leave the server — which is what allows the method to be sold as metered calls without giving it away.