Where does the flavor actually come from?

Pick up a plate of paella. By weight it is rice, chicken, beans, water. The rosemary in it — the thing you smell from the doorway — is 0.06% of the dish’s mass. This piece is about that gap: the distance between where a dish’s mass is and where its flavor signal is, measured across every dish in Palate’s corpus.

The arithmetic is the engine’s own aggregation weight — log1p(grams) × (potency + 1)^1.5 — applied to 27,229 linked ingredient lines across 2,343 dishes (0.7% of lines were unlinked or massless and skipped; 1,665 of 1,729 distinct ingredient names link exactly). An ingredient’s signal share is its weight over the dish’s total; its mass share is grams over grams.

Every linked ingredient line in the corpus: share of the dish's mass against share of its flavor signal, both log-scaled. The diagonal is 'contributes exactly its weight.' Aromatics live far above it; rice, plain pasta and water far below.

Three dishes, opened up

Spaghetti carbonara. The richness is 37% pancetta, 34% egg yolk, 20% pecorino. The spaghetti — 44% of the mass — contributes 0%.

Spaghetti Carbonara — richness 6.8/10
ingredient	share of richness	share of mass
pancetta	37%	22.0%
yolk	34%	12.9%
pecorino romano cheese	20%	7.3%
eggs	9%	12.9%

French onion soup. The umami is 43% beef broth and 38% gruyère. The onions the dish is named for: 4%, from 36% of the mass. (What caramelization does to those onions is real, and is exactly what a composition model undersells — see the limits.)

French Onion Soup (Soupe à l'oignon) — umami 9.2/10
ingredient	share of umami	share of mass
beef broth	43%	37.6%
gruyere	38%	12.5%
dry white wine	10%	6.3%
fresh yellow onions	4%	36.0%
season salt	2%	0.08%
unsalted butter	2%	1.4%

Paella Valenciana. The aromatic channel is 52% garlic and 40% rosemary — rosemary being 0.06% of the dish by mass. A 600-fold mismatch between where the weight is and where the smell is.

Paella Valenciana — aromatics 1.2/10
ingredient	share of aromatics	share of mass
garlic	52%	0.38%
rosemary	40%	0.06%
sweet paprika	9%	0.15%

The corpus-wide picture

Measured as the median across all recipes each appears in: cayenne punches 30× over its mass share (n=183 recipes), rosemary 35× (n=36), thyme 23× (n=132), saffron 26× (n=25), shrimp paste 39× (n=19). At the other end, white rice returns about 0.1× its mass share in signal (n=68) and water — 19% of a typical dish it appears in — carries essentially nothing (n=325).

Aggregated: in the median dish, the ingredients that are each under 5% of the mass together carry 49% of the flavor signal (interquartile range 33%–60%, n=2,343 dishes). Half of what a dish is, by this model, comes from things you could hold in one cupped hand. None of this will shock a cook — nobody seasons by the pound. What the measurement adds is how big the effect is, and which ingredients live furthest from the diagonal.

The most leveraged aromatics (salt rows omitted — salt behaves the same way but is its own story). Only ingredients appearing in ≥15 recipes.
ingredient	leverage (signal ÷ mass)	median mass share	n recipes
shrimp paste	39×	0.26%	19
cayenne	36×	0.10%	27
rosemary	35×	0.11%	36
cayenne pepper	30×	0.11%	183
marjoram	29×	0.08%	15
saffron	26×	0.08%	25
cardamom	25×	0.22%	19
thyme	23×	0.06%	132
nutmeg	20×	0.07%	58
dark sesame oil	20×	0.32%	124
pepper	19×	0.08%	65
coriander	19×	0.09%	107

Honest limitations

This is a composition model, not a perception model. Real flavor perception runs on concentration ÷ odor threshold (the odour-activity-value framework of Grosch and colleagues), not gram share with a potency correction. A mass-weighted model systematically under-weights high-impact aromatics — saffron’s true perceptual leverage is likely higher than shown, not lower. Treat the leverage numbers as floors on the phenomenon.
The potency term is a hand-labeled partial correction, not a measurement. No human palate has ever validated one of these vectors against tasting the dish — the attribution is exact arithmetic on the engine’s model; the model itself is the untested part.
Cooking transforms contributions. Caramelized onions give more than raw-onion labels suggest; volatiles cook off; fat carries flavor it doesn’t own. The attribution here is computed on raw composition.
~40% of the corpus is project-authored (934 of 2,345 dishes), which may use more regular quantities than recipes in the wild.

Reproduce this

Deterministic, offline, no model calls. Every number on this page is read from the script’s results file — the page cannot state a number the script didn’t produce.

python analysis/instrument_room/scripts/piece2_attribution.py
python analysis/instrument_room/scripts/piece2_figures.py

Inputs: prototype/web/src/data/sample_corpus_v17.json + data/recipe_corpus/ingredient_vectors_v2.json (read-only). Every number on this page is in analysis/instrument_room/results/piece2_attribution.json.