Flavor provinces: the map that replaces cuisine labels

Full membership for every province is in the results file; the cards above show the anchors. What the map says, qualitatively: the savory European pantry refuses to split into “Italian” and “French” (one 76-ingredient fabric); the sweet/savory line is the strongest border in the data; the East Asian pantry (n=29) is the tightest province by more than double; the statistic rediscovered mid-century American convenience cooking (worcestershire · paprika · green pepper, n=26) without being told; lime · cilantro · cumin ingredients that a cuisine atlas files under both “Mexican” and “Indian” share one province (n=16); and a pickling-and-brine counter (mustard · tamarind · tabasco, n=15 — dill, horseradish, vinegar, beet, salmon, crab) emerged that no cuisine name covers.

The 37 seed-unstable ingredients are not noise — they are the pantry’s most cosmopolitan citizens, and their refusal to sit in one province is evidence the method is working. A few travel in packs: broccoli + cheddar cheese + cottage cheese + egg noodle; lemongrass + peanut + thai pepper; brown rice + cauliflower; canola oil + garlic; cucumber + vegetable oil; heavy cream + shallot; lamb + saffron; pork + scallop. The rest travel alone: black bean, blue cheese, cabbage, cardamom, chicken, chickpea, chicory, cod, cornstarch, grapefruit, lemon juice, lentil, lettuce, pea, sherry, starch, watercress, yogurt. All are excluded from every published cut (reporting floor n ≥ 15).

Mean PPMI edge weight within provinces runs 0.30–1.08; between provinces, 0.002–0.20. The weakest coupling in the whole map is dessert shelf ↔ casserole aisle; the strongest bridges are exactly where a cook would guess — the East Asian pantry ↔ the lime-cilantro province, and the brine counter ↔ the European savory pantry.

• The corpus is Western-skewed and one source is 90% of it. 63,696 recipes = 5,296 parsed + 1,081 hand-curated + 57,319 from the dataset published with Ahn et al.’s 2011 paper, which is largely North American and European. A West African or Levantine province is absent because the corpus barely wrote those recipes down, not because those pantries lack structure.
• Community detection is resolution-dependent and partitions are not unique. We ran Louvain at resolution 1 — the default, disclosed, not tuned. Other resolutions would give coarser or finer maps. “Six provinces” is a property of this graph at this resolution, not a law of cooking.
• The seeds disagree, and we measured it instead of hiding it. ARI 0.71–0.85 across 5 seeds; 5 communities under some seeds, 6 under others. Everything published here survives all runs; the 37 ingredients that don’t are named and excluded. A same-seed rerun is byte-identical (verified in tests), so the instability is the algorithm’s, not the code’s.
• PPMI favors the rare end. Even above the inherited floors (pair count ≥ 20, both ingredients in ≥ 100 recipes, plus Piece 4’s staple and near-duplicate exclusions), rarer ingredients score proportionally higher, which can pull niche tokens toward the center of small provinces. Every member ships with its df in the results file.
• Co-occurrence is convention plus popularity, not deliciousness. The worcestershire province is the proof: the statistic faithfully mapped 1955’s casserole aisle. These provinces describe what recipe authors do together, not what belongs together.
• The vocabulary carries its sources’ quirks. “Celery oil” and “lovage” in the brine province are compound-database tokens standing in for things a cook would name differently. We publish tokens as-is.

Every number on this page is a deterministic measurement over the same recipe corpus — nothing here was estimated or embellished. Want to check the work, or run your own analysis? A 250-dish sample of the corpus — ingredients, 18-note flavor profiles, timing, macros — is free to download.