Economics, Extra Sauce

The Palate Translation Department

Sam Ate It.
Would Jon Love It?

A reproducible taste prior, a modest regression, and absolutely no imaginary meals.

What This Predicts

Estimated Jon fit on the historical 100-point Discovery Score scale. It is not Jon’s personal star rating, a probability of enjoyment, or a verified recreation of the old Jon 50 calculation. The original project documented chosen weights and a shortlist, but no fitted coefficient table or reproducible input matrix.

Jon’s Taste Prior

The original Atlanta notes favor plump, juicy, crisp wings; savory lemon pepper, tang and heat; and less sweetness or gooey texture. Chicago’s positive ordinal position is retained as preference context, never converted into a made-up star rating. The four documented numeric personal ratings do not provide paired Sam scores sufficient to fit a personal-rating regression.

Sam base = 10 × (0.40 texture + 0.25 meatiness + 0.25 sauce + 0.075 experience + 0.025 appearance).

These are declared modeling choices based on Jon’s emphasis, not weights learned from tasting labels. Texture is a proxy for execution, not an invented crispiness or juiciness score. Sauce quality says nothing definitive about sweetness, tang or heat.

Known seed descriptors add a maximum ±12-point adjustment. Crispness, juiciness and low sweetness each have weight 3; savory and tangy each 2; heat and traditional style each 1. Match = 100 − |observed code − target|. Adjustment = 12 × Σ[weight × (match − 50)/50] ÷ 15. Unknown attributes contribute no adjustment and remain explicitly unknown.

The Regression Test

A one-predictor ridge regression maps that prior score to the 21 historical Jon Discovery Scores with reviewed Sam name/state associations. Both axes are centered at 50 and scaled by 25. Ridge penalty 10 shrinks intercept toward zero and slope toward one. It is fixed before leave-one-overlap-out testing. Each test fits on 20 pairs and predicts the omitted historical score. Comparisons include the uncalibrated prior and the training-fold mean.

This calibrates one provisional model to another. It does not validate taste predictions. Descriptor coding may reflect the same research as the old targets; name/state matches are not exact-branch verification. The small difference between candidate models is exploratory, and the selected model’s error is not an independent final-test result.

Calibration Pairs and Held-Out Predictions

RestaurantHistorical JonPriorHeld-Out RidgeAbsolute Error