The Instrument Room
Measurement, not opinion.
Palate turns recipes into numbers. Most of the time those numbers power the app — but some of what they show is interesting on its own. These pieces are the findings we can stand behind: each one is a single measured claim about food, with the script that computed it, the chart it drew, and a section on what the data can’t say.
No number here came from a model’s imagination. If we couldn’t compute it or cite it, we left it out.
Piece 01
When does it go in? — pick an ingredient
The interactive version of the timing chart. Choose any of 161 ingredients, compare two, and split the curve by the kind of dish — cinnamon sits at 0.28 in mains and 0.48 in desserts.
Read the piecePiece 02
Nationality tells you little about how food tastes
236 recipe-title words scored by how similar their dishes actually taste, on 18 measured flavor dimensions. Ingredient words carry real information; most nationality words are indistinguishable from chance.
Read the piecePiece 03
Where does the flavor actually come from?
Every dish's flavor signal, attributed to its ingredients with the engine's own arithmetic. In the median dish, ingredients under 5% of the mass carry half the signal — paella's rosemary is 0.06% of the mass and 40% of the aroma.
Read the piecePiece 04
The order of operations
Four taught sequencing rules checked against 14,362 real recipe steps. Fresh herbs at the end holds; garlic-after-onions collapses; 37% of recipes never mention salt in a single step.
Read the piecePiece 05
What actually goes with X
A pairing reference computed from 63,696 recipes — pointwise mutual information with the n printed on every value, and an honest accounting of what co-occurrence can and cannot claim.
Read the piecePiece 06
Nobody has measured how sour things are
USDA publishes sodium for 410 of the 432 ingredients that can move a dish's acid reading. The number of them with a published measurement of how much acid they contain: 16.
Read the piecePiece 07
Flavor provinces
Ingredient communities found in the co-occurrence graph rather than assumed from cuisine labels — and an honest account of why the partition is not unique, published as the stable core instead of a tidy map.
Read the piecePiece 08
Every cooking verb has its hour
66 techniques placed by where in a recipe they happen. Garnishing is the literal last step 92% of the time; seasoning is the one verb with no hour. Plus: a four-hour recipe is fifteen minutes of cooking and a long wait.
Read the piecePiece 09
What each kind of dish is made of
Ten kinds of dish across 18 measured flavor dimensions — and only three of them have a fingerprint at all. What marks a dessert is the missing savory as much as the sugar.
Read the piecePiece 10
A well-stocked pantry is not a sufficient one
Own the 50 most common ingredients and you cover half of every ingredient line in the corpus — and can cook one recipe out of 2,345. Plus a dosage reference: how much of an ingredient is actually a lot.
Read the piecePiece 11
Nobody timed the recipe
77% of stated cook times are a multiple of five minutes — four times what you'd see if the last digit meant anything. Recipe times are estimates wearing a number's clothes.
Read the piece