WeatherOrb

One weather brain for sky and sea: eleven forecast models on one machine, blended into a single field, and rendered as a map that can be scored against the store behind it.

The models are synced from Open-Meteo’s public archive. Six atmospheric members are regridded onto a common 0.0625° grid and published as one pseudo-model, weatherorb_blend — the layer the map opens on. The field is rebuilt every three hours; the last full round took 5,729 s over a 92 GB store, and the catalogue it serves is 136 layers across 15 models, the wave fleet and three static atlases included.

The blend is more than an average with a better name. Temperature and humidity weight the local model first while its own lead is under six hours, equally after. Precipitation and clear sky ship as agreement — the share of members over the threshold — because the members’ disagreement is the one signal no upstream hands over. Where a single member covers a cell, that cell keeps its value and its spread stays undefined.

Nothing is smoothed to look finished. A layer with no data draws nothing, an hour a model does not carry is null, and a scoring factor with no input drops out with its coverage saying so.

The map: blended wave height across the Bay of Biscay drawn as a field, the value printed at every town, wave propagation as particles, and the legend, layer stack and timeline beside it. — Blended wave height over the Bay of Biscay — the field, its value at every town, and wave propagation drawn as particles. Captured from the running app, not composed.

Bars it passes

Every figure is a critic-checked measurement; the full ledger, losses included, is in the repository’s progress log.
Bar	Measured
Point probe against the upstream API	12/12 points, \|probe − API\| = 0.0
Rendered pixel against an independent LUT decode	40/40 canvas pixels within 2 channel steps
Temperature field against Windy’s own decoded field	corr 0.996, bias −0.06 °C, RMSE 0.57 °C over 66k px
Wave height against Windy’s own decoded field	corr 0.998–0.9998, RMSE 0.012–0.036 m
Cloud blend against its best single member	CRPS 53–60 % better at 3/3 stations
Precipitation agreement against its best member	Brier 0.066 against 0.098
Clear-sky agreement against its best member	Brier 0.126 against 0.177
Render, four seconds of continuous pan	median frame 11.3 ms, p95 27.1 ms, with raster and particles

The Windy comparison is field against field, not screenshot against screenshot: their tiles carry the scalar with a per-tile range map, so both sides decode to physical units before anything is differenced. At the cell pitch their pyramid stops at zoom 4; this one serves measured data to zoom 8.

Open the map

Forecast data from Open-Meteo’s AWS Open Data mirror. Basemap from Protomaps and OpenStreetMap. Bathymetry from EMODnet DTM 2024 and GEBCO 2026; night sky from the Lorenz Light Pollution Atlas 2025.