How a verdict is reached.

We do not score wines and do not rank wineries by how good they are. We collect verifiable facts about the producer, the price and the awards, and derive a category from them by published rules. The rules are the same for everyone: two people with this methodology and the same sources arrive at the same result.

The category is computed by a deterministic function, not by a language model. Feature weights and thresholds live in a versioned config in the repository, and the version is stored next to the producer's category — it is always visible which edition of the methodology produced the result. Weights cannot be changed quietly: a set of reference wineries with known categories is checked on every build, and a change that breaks it fails the build.

Silence in the sources is not evidence. Where data is missing, the producer never gets a lower category by default: the absence of confirmation says something about us and our sources, not about the company. But silence about one thing is not silence about everything — a winery whose vineyard area is not published is not thereby a winery we know nothing about, and we say what we do know.

Every conclusion opens up to the facts it rests on: each fact carries a value, a source, a confidence level and the date we recorded it. That is both a product feature and our legal position — we publish what can be observed, not guesses about what is hidden.

There is no single «wine score» here and there will not be one. Provenance, recognition and the position of the price are shown separately, each with its own explanation, and we give no tasting scores of our own: a single taster would be exactly the matter of taste this project is built away from.

A significant fact is not stored as a field on a card but as a statement: a value, a source, the date it was observed and a confidence level — state registry, producer statement, inference from catalogue data, or manual verification. On the producer's page that list opens under «What this is based on», with the source and the date next to each fact.

Classification uses features from outside sources — production address, own vineyard area, full-cycle production, the estate-bottled mark, a named winemaker — together with features counted from our own catalogue: the number of appellations and regions in the range, spirits in the lineup.

Sources are ranked: registry first, then the label, the producer's own site, the retailer, everything else last. An individual source can be raised or lowered relative to its type — a fact checked by a person, for instance, should not lose to a shop scraper.

When sources disagree we do not quietly pick the right one. Both versions stay in the database with their sources and dates, the higher-ranked one wins, and the discrepancy is flagged for review — and, when it is worth the reader's attention, shown openly. «The label says since 1920, the company was registered in 2019» is content, not dirt in the data.

Re-importing the same value from the same source changes nothing; a new value from that source supersedes its own previous one, while statements from other sources are left untouched — their disagreement is settled by rank, not by overwriting.

From shop catalogues we take facts only: the name of the listing, the price, availability and the link. Retailers' own descriptions we do not copy.

The verdict answers one question — is this wine expensive relative to comparable ones — and says nothing about its quality. It is computed from price observations: for each shop we take the latest recorded price among listings that were in stock, and the median of those is the wine's price.

That median is compared with the median of a cohort. The cohort is the wines of the same producer category and the same style class, and the style class is currently colour plus sweetness. If either is unknown, the style class is undefined and no verdict follows: «no data» is never treated as a value.

Sparkling and semi-sparkling wines form one style class of their own and are compared only with each other, never with still wines of the same colour. We do not split them by sugar or colour: a sparkling wine's price is set by its method and ageing, not its sugar, and seven sugar levels would leave cohorts too small to compare. We cannot yet tell methods apart (traditional, tank, pét-nat), so a traditional-method wine sits in one range with tank-method ones and usually comes out pricier than the range. Sugar for sparkling wines is shown with the EU label terms — brut, extra dry, demi-sec — not the still-wine scale: brut is not «dry».

A deviation within ±15% of the cohort median reads «in line with similar wines», below it «cheaper», above it «pricier». Neither word is a judgement of the wine: a dearer bottle from a small producer is not a worse one. The tolerance and the other thresholds are configuration rather than a constant written into this text; on the wine page an expandable block shows the actual numbers — the wine's median, the cohort's median and the size of the cohort.

Vintage weather is computed from the reanalyses of the Copernicus Climate Change Service: temperatures from ERA5-Land (a grid of about 0.1°, roughly 9 km), precipitation from ERA5 (a 0.25° grid). A reanalysis is a model reconciled with observations, not a weather station in the vineyard: it describes the area around a grid node. We take the data directly from the Copernicus Climate Data Store under the CC BY 4.0 licence: «Contains modified Copernicus Climate Change Service information 2026». Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus information or data it contains.

The season is compared not with the climate normal but with this zone's own mean over the previous ten years: against the 1991–2020 normal almost every recent year would read as “warmer than usual”, which says more about warming than about the season. On a wine page only the unusual is coloured: warmth and September rain against that decade; a frost means a minimum of −2 °C or lower after 10 April. An uncoloured indicator means “as usual”, a dashed one means no data.

The weather describes an area, not a particular vineyard. If a wine has no appellation on record, the weather is taken for the region and describes the wider area. If a zone has no village of its own in open maps, its point is approximate and the season is computed at the nearest landmark. Several zones can share one grid cell; their weather does not differ in this data. Warmth and rain explain what the season was like; they do not rate the wine or tell you when to open it.

The series were recomputed in full in October 2026. The same models used to arrive through the Open-Meteo archive; they now come from Copernicus itself and follow the rules below, the same for every year and country. For most years the figures did not change. Rainfall in 1996 and 2022–2025 came out 5–10% lower in many zones: the old archive processed those years differently from the rest, so they looked wetter than their decade because of processing, not weather. In some zones individual marks changed — mostly September rain, occasionally a hot day or a frost that sat right at the threshold.

The data comes in four layers: public registries, producers' own sites and labels, shop catalogues, and the official winners' lists of competitions. From shops we take the price, availability, the name of the listing and the link. Which registry stands behind an Armenian appellation is stated on the page of that appellation, next to the fact itself.

Different data moves at different rhythms. Prices are collected in a full weekly pass; once a week we also look for changes — listings that appeared or vanished in shops, winery pages that were edited; producers' sites are crawled monthly. The pace is deliberately gentle, over public pages, without working around any restriction.

Price observations are written once and kept: every price stays in the database with its own date and shop. History therefore accumulates from the very first run, and a wine catalogued later inherits all of it instead of starting from zero.

Part of the work is automatic. Linking a source record to a catalogue record is applied on its own once the confidence computed from weighted signals passes a threshold; where the match is ambiguous nothing is decided at all, and anything decided can be undone. That relaxation does not extend to classification or to publishing facts about companies — there the decision stays with a person.

The database is assembled from open sources and it contains errors — that follows from how it is built, not from carelessness. We would rather show «no data» than fill a gap with a guess, so an empty spot on a card is itself a piece of information.

Found a mistake — tell us. A correction that names its source is the fastest kind to verify, and the handling is the same as for the right of reply described in the section above.