Where this goes next

The pipeline so far reproduces one good idea faithfully: canadarasp's windgram, ported constant-for-constant, published openly. That parity was the right first move — pilots calibrate against the original, and trust is earned by agreeing with it before diverging from it. But a pipeline that already speaks six models, keeps every forecast it has ever published, and costs nothing to run is a foundation, not a finish line. This page is the honest version of the roadmap: what looks promising, what it would take, and what we would refuse to ship without labelling.

Multi-model disagreement as an uncertainty signal

Once RDPS, GDPS, HRRR, and GFS profiles publish alongside the two HRDPS legs, the same site and hour exists in up to six independent forecasts. Deterministic models agree on easy days and diverge on hard ones — which makes disagreement itself a forecast product. A renderer that overlays six usable-lift tops tells a pilot something no single model can: whether tomorrow is a confident forecast or a coin flip. The first version of this is pure presentation — the data will already be there.

The history archives make the retrospective version possible too: every run of every model is preserved (data/history/, one JSON line per run), so "which model actually calls this valley best?" is a zcat | jq study away — per site, per season, per quantity. Verification against pilot-recorded IGC tracks is the gold standard here, and would also finally test the forty-year-old Lenschow–Stephens constants the lift model rests on (see how a windgram is computed).

Ensemble windgrams

ECCC publishes true ensembles — GEPS (global, 20+ members) and REPS (regional) — designed to quantify exactly the uncertainty the multi-model overlay only gestures at. An ensemble windgram would draw bands, not lines: the boundary-layer top as a 10th–90th percentile envelope, thermal strength as a distribution. The derivation is already pure and per-column, so running it per-member is mechanical; the real work is presentation (what does an honest uncertainty band look like on a time–height chart?) and cost (members multiply fetches — the subsetting patterns in engineering the pipeline become load-bearing).

A long-range outlook with proxy physics

The ECMWF IFS investigation (bestiary) ended in a "no" for windgrams: no published heat fluxes, three usable pressure levels. But a 15-day outlook is a different product with a different honesty bar — "which days next week are worth watching?" rather than "how high at 2 pm?". A radiation-proxy flux with an assumed Bowen ratio is defensible for that question, provided the product is named and labelled as estimated physics and never rendered in the windgram's clothes. This is the most likely next adoption; CC-BY-4.0 attribution is the only licensing cost.

Gridded and map windgrams

Everything today samples single launch points. The models are grids; the derivation is a pure function of one column. Run it over every column in a crop and you get a map of usable lift top or thermal velocity — RASP-style regional views, or a transect along a ridge line. The WCS-crop and byte-range transports already fetch small regions, so the marginal cost is derivation compute (cheap) and output size (the real design problem: a 48-hour gridded product is no longer a 60 KB JSON file).

Richer sounding-derived parameters

Nine pressure levels is a coarse sounding, but it is still a sounding, and the profile JSON already carries it. Candidate derivations, roughly in order of value per effort: CAPE and convective inhibition (over-development risk — the windgram currently says nothing about storms), freezing level, wind shear through the boundary layer (a safety number pilots actually brief on), and a buoyancy-based top-of-lift that could one day be compared against the Lenschow–Stephens hcrit rather than replacing it silently. Any change to the soaring numbers themselves ships as a new, named quantity next to the old one — parity with canadarasp stays a feature, not an accident.

Your site belongs here

The catalogue is sites.json: slug, name, coordinates, launch elevation. That's the entire onboarding surface. The West Kootenays is the founding case study, not a boundary — HRDPS covers nearly all of Canada, HRRR the continental US, and GFS the planet, so any launch on Earth is at most a pull request away from a published, permanently archived, freely consumable windgram. Each site costs roughly 60 KB per model run. Bring coordinates; the pipeline does the rest.