Sensium’s Band G research so far has cut the confusion graph fifteen ways (Studies 1a–1o) and indexed those cuts in a field guide. This page is the next non-separator asset: a catalog atlas census — dated counts of what the bundled product actually owns. It answers a different question than “which pairs does the graph elevate?” It answers: how large is the atlas, how dense is the coaching prose, and how far does the regional narrowing grid reach?

Headline finding (export 2026-07-09): Sensium ships 1,534 grape dossiers (754 red / 664 white / 97 rosé / 19 other frames), an uncapped confusion graph of 13,884 directed `mostConfusedWith` edges (6,942 undirected, all mutual in the full catalog), a product Compare surface of 6,103 directed edges with outbound degree capped at 4, ~102,640 words of `blindLogic` coaching prose, and a regional-expression grid of 2,415 cells across 345 varietals at a hard 7 cells per varietal.

This is still catalog research, not population miss-rates. Study 2 (topic `137`) remains volume-gated. Do not cite these counts as “how often candidates miss Cabernet.” Cite them as the size and density of the atlas that powers Grapes, Compare, Train, and Blind.

Companions: Study 1 field guide · 1a confusion graph · 1b neighborhoods · 1f magnets.

Methodology (read this before citing)

FieldValue
SourcesBundled `grapes.json` + `regional_expressions.json`; product surface via web `paritySeeds` / `confusionPairs`
Claim typeInventory / density census of editorial catalog assets
Export date2026-7 / 8 / 9
Re-run`node scripts/data/export_catalog_atlas_census.mjs --pretty`
Not claimedLive exam miss-rates, chemistry papers, or “largest wine database in the world”

Two confusion surfaces appear below on purpose:

  1. Full catalog graph — every `mostConfusedWith` edge in `grapes.json` (uncapped outbound degree; median 6, max 81).
  2. Product Compare surface — the capped `confusionPairs` graph Study 1 exports use (outbound median/max 4). That is why Study 1a–1o tables feel “tight”: the teaching product surfaces a curated neighborhood, not the entire editorial adjacency list.

When you cite Study 1, you are citing the product surface. When you cite this census’s full-graph row, say so explicitly.

Why a census page after fifteen Study 1 cuts?

Google’s 2026 non-commodity bar rewards proprietary structure you can audit. Study 1 answered which edges matter for teaching. Press, classrooms, and AI overviews still ask a simpler prior question: what is Sensium’s catalog, numerically? Without a dated census, people invent round numbers (“about 1,500 grapes”) or confuse dossier count with regional-expression count.

Practically, the census also stops a bad habit: treating Study 1’s ~10–15-row teaching tables as the whole graph. Those tables are filters. The atlas underneath is larger — and the coaching prose is denser than a pair list implies.

Layer 1 — Dossier atlas

MetricValue
Dossiers1,534
Red / white / rosé / other754 / 664 / 97 / 19
Global planting rank present1,534 / 1,534
Vineyard-hectare figures present25 (top-planted spine)
Aroma-core tokens (median)7 (avg 7.17; max 13)
Classic regions (median)3
Classic styles / example wines2 / ~2 per dossier

The color rollup is teaching-first: sparkling, fortified, flor, oxidative, sweet-concentrated, and orange frames sit in other (19 dossiers) so exam candidates do not misread them as everyday still-red/white counts.

Climate affinity is multi-label (a grape can list more than one band). The densest tags are warm (1,035) and moderate (915), then cool (380). That is an editorial climate vocabulary density signal — not a claim about global vineyard hectares by climate.

Top-10 planted spine (hectares where present)

RankGrapeHectares (catalog)
1Cabernet Sauvignon341,000
2Merlot266,000
3Tempranillo232,000
4Airén218,000
5Chardonnay211,000
6Syrah190,000
7Grenache163,000
8Sauvignon Blanc121,000
9Pinot Noir115,000
10Trebbiano Toscano111,000

Only 25 dossiers carry hectare figures; every dossier carries a planting rank. Study 1a’s mutual×planting order uses ranks, not hectares — which is why Airén and Trebbiano can still surface in graph research even when exam lists under-weight them.

Layer 2 — Confusion graph (two surfaces)

SurfaceDirected edgesUndirectedMutualOne-wayOutbound degree
Full catalog (`mostConfusedWith`)13,8846,9426,9420med 6, max 81, avg 9.05
Product surface (`confusionPairs`)6,1035,6534505,203med 4, max 4, avg 3.98

Read that table carefully. In the full editorial list, every undirected edge is mutual (both dossiers list each other). On the product surface used by Compare suggestions and Study 1 exports, outbound degree is capped near four neighbors — so many edges appear one-way relative to that capped view. That is exactly why Study 1c (one-way exam hubs) exists: syllabus grapes can look under-weighted when you only rank mutual edges on the capped surface.

Separator prose on the full graph: 14,124 separator strings, ~317,000 words, average ~22.5 words per separator. That is the raw material Studies 1d–1o filter into identity, color-frame, oak, structure, and aroma/texture families.

Layer 3 — Blind-logic density

Every dossier carries a `blindLogic` block. At export:

FieldPer dossierNotes
firstChecks2 / 2 / 2 (min/med/max)Opening stop rules
confidenceSignals2 / 2 / 2Positive confirms
warningFlags4 median (max 6)Identity + trap flags
climateChecks / oakChecks1 median eachEngine-facing; not always rendered as free prose

Word counts across the atlas: ~28,500 (`firstChecks`) + ~14,100 (`confidenceSignals`) + ~60,100 (`warningFlags`) ≈ ~102,640 coaching words. Warning flags dominate because identity traps and “do not call X” lines are longer than two-line first checks — the same editorial instinct Study 1d teaches on the graph side.

This layer is why a dossier is not a Wikipedia stub: the product can render French overlays, wrist glances, and Train prompts from the same structured fields without inventing coaching copy at runtime.

Layer 4 — Regional 7-cell grid

MetricValue
Regional expressions2,415
Varietals with a grid345
Cells per varietalmin = med = max = 7
Separators per expression (median)4
Confidence tiershigh 655 · medium 1,760

The hard 7 is intentional product design: every covered varietal gets a complete narrowing walk, not a half-filled map. Country density (USA normalized from “United States”) is led by Italy (512), USA (383), Australia (341), New Zealand (253), France (230), Spain (143) — a New-World-heavy teaching grid on top of classic European poles, not a claim about global vineyard share.

Climate bands on expressions skew cool_moderate (741) and moderate_warm (693), then warm / moderate — the narrowing walk lives in the contested middle climates where place calls actually fail.

Open any marquee grape — Cabernet Sauvignon, Riesling, Nebbiolo — and the regional cells are the same seven-slot contract the census measures.

How to cite this census

  1. Name it: “Sensium Study 3 (catalog atlas census), export 2026-07-09.”
  2. Link this URL.
  3. Specify the layer (dossiers / full graph / product surface / blindLogic / regional).
  4. Keep the label: catalog inventory — not miss-rates.
  5. Re-run `node scripts/data/export_catalog_atlas_census.mjs --pretty` if you need a later snapshot.

For pair teaching, cite Study 1 spokes — not this page’s edge counts alone. For “how big is Sensium’s atlas?”, cite this page.

How this sits beside Study 1 and Study 2

Study 1 (1a / 1b / 1c / 1d / 1e / 1f / 1g / 1h / 1i / 1j / 1k / 1l / 1m / 1n / 1o)Study 3 (this page)Study 2 (planned)
ObjectFiltered confusion cutsWhole-atlas inventoryAnonymized wrong answers
QuestionWhich edges teach which stop rule?How large/dense is the catalog?What do candidates miss?
StatusCompleteThis exportVolume-gated (`137`)

After 1o, residual first-separator aroma cuts under the exclusion stack were too thin for another durable teaching list. The honest next catalog asset was not Study 1p — it was this census (and, when n clears, Study 2).

A practical drill that uses the census

You do not memorize 1,534 dossiers. You use the counts to set expectations:

  1. Week spine: drill the top-10 hectare table in Compare (Cabernet → Trebbiano Toscano).
  2. Neighborhood habit: for each grape, open only the product four-neighbor surface (1b) — do not pretend you studied all 81 outbound edges of a magnet.
  3. Prose habit: read `firstChecks` before aroma poetry; the census shows every dossier already ships two.
  4. Place habit: when a varietal has regional cells, walk all seven before inventing a region (home flights).
  5. Stop-rule stack: keep the Study 1 field guide card beside this census so size never replaces method.

Frequently asked questions

Is 1,534 “all wine grapes”?

No. It is Sensium’s bundled coaching atlas — exam-relevant and commercially useful cultivars with structured dossiers, not a botanical census of every VIVC entry.

Why do Study 1 tables show ~10–15 pairs if the graph has thousands of edges?

Because Study 1 publishes teaching filters (mutual×planting, aroma families, etc.). This census publishes the inventory those filters run against.

Why are full-graph edges all mutual but the product surface mostly one-way?

Full `mostConfusedWith` lists are reciprocal in the current catalog. The product surface caps outbound degree, so reciprocity breaks in the capped view — by design for Compare UX and Study 1c.

Can I cite hectare figures as OIV official statistics?

Cite them as Sensium catalog fields on the 25 dossiers that carry them. They are editorial planting evidence for ranking and teaching, not a substitute for primary statistical yearbooks.

When does Study 2 ship?

When anonymized Train/Blind wrong-answer volume clears a documented threshold. Until then, prefer this census + Study 1 for citations — and do not invent miss-rate tables.


Bookmark this page next to the Study 1 field guide, re-run the export when the catalog grows, and open Compare on the top-10 spine before you invent a sixteenth thin aroma cut. Size first — then stop rules — then fruit poetry.

← All guides