Foudre is a young product, and we'd rather be plainly honest about that than dress a preview up as the finished thing. Some of what you see is built on real, authoritative data today; some is an illustrative preview of a feature we're still wiring to live data. Here is exactly which is which — and how the app itself tells you, every time.
What is real today
The climate engine is real. We use the European Copernicus programme's ERA5 reanalysis for the 1990–2024 daily history and its SEAS5 seasonal-forecast ensemble for the year ahead — the same datasets used in published climate research — downscaled to each château's location and elevation.
The estates and terroir are real. Coverage is the 61 Left-Bank 1855-classed châteaux, with location, elevation, distance to the Gironde, slope and aspect curated from estate sources and public elevation models. The scoring itself runs against a structured vintage-quality model built from the viticulture literature.
The backtest evidence is real, with one caveat. For past vintages, the critic and market scores Foudre compares itself against are read from public web pages at the time of the request, each kept with a link to its source. They are extracted by AI from those pages, not pulled from a licensed ratings feed — and the report says so.
What is still illustrative
A few visuals on the Trade dashboard — the market-sentiment donut, the regional risk comparison, and the sample weather line — are illustrative previews. They show the shape of the feature, not yet a live computation, and each carries an amber “Illustrative” badge so it can't be mistaken for a measured result.
Document uploads are read for filename and type only today; their contents are not yet parsed and do not change the score. And when a demo mode is enabled, or a live data channel is unavailable, the whole result is produced from hand-tuned fixtures or fallbacks rather than a full live run.
How the app tells you
Every analysis opens with a provenance banner. When the run is fully live — real climate, terroir, retrieval and model — it shows a quiet green “Real data” mark. When any channel fell back to a fixture or a heuristic, it turns amber and lists exactly which channel degraded, so a preview can never pass for a complete result.
Charts that are previews carry their own amber “Illustrative” badge. If you don't see a badge and the banner is green, what you're looking at is the real thing. As we wire the remaining previews to live data, those badges come off — and we'll note it here.