In this LAB note
Question · Hypothesis · Method · Results · What Failed · What We Learned · What Changed at MOMS
Question
MOMS describes personalization as three inputs — body type, program week, and goal. Are those three independent axes, or does the third one fall out of the first two?
Hypothesis
We assumed goal was a real axis. The working model was that a customer's body type sets a baseline, the program week advances it, and the goal steers it — so two people with the same body type in the same week could receive materially different shakes depending on what they were after. If that were true, the database should contain body-type-and-week cells holding more than one goal. We also expected week to carry a progression: week 4 should look further along than week 1.
Method
The precision-nutrition literature is largely a literature about variance partitioning, which is the same question in a different vocabulary. Zeevi and colleagues (Cell, 2015) monitored week-long glucose in 800 people across 46,898 meals and found that identical meals produce highly variable responses — no single dietary rule predicted well, and their model only worked once it combined blood markers, anthropometrics, activity and microbiome together. Berry and colleagues (Nature Medicine, 2020) ran PREDICT 1 in 1,002 adults and decomposed the variance explicitly: person-specific factors such as the gut microbiome accounted for 7.1% of postprandial triglyceride variance against 3.6% for meal macronutrients, while for glycemia the ordering reversed at 6.0% and 15.4%. The lesson both studies carry is that in personalization, the main effects are small and the interaction is where the signal lives. Internally we treated our own database as the object of study. We took all 201 recipes and asked, first, a structural question: how many distinct goals appear inside each body-type-and-week cell? Then we ran a one-way variance decomposition for each of body type, week, and goal against fat, fiber, carbohydrate and Score, reporting omega-squared rather than eta-squared because our three factors have very different level counts (8, 4 and 28) and eta-squared rewards the factor with more levels.
Results
The structural answer settled it before the statistics did. Eight body types by four weeks gives 32 cells. All 201 recipes fall into those 32 cells, and every single cell contains exactly one goal — zero cells contain two. There are 28 goal names covering 32 cells. Goal is not an input to the system; it is a name the system applies after body type and week have already decided where you are. The variance decomposition agrees. Week on its own explains almost nothing: omega-squared of 0.032 for fiber, 0.036 for fat, 0.082 for carbohydrate and 0.011 for Score. Body type carries more — 0.164, 0.136, 0.159 and 0.106 respectively. Goal, which is simply a relabeling of the body-type-and-week intersection, carries roughly twice what body type does alone: 0.222 for fiber, 0.196 for fat, 0.221 for carbohydrate and 0.339 for Score. The combination is worth about double either dial by itself, which is the same shape Berry's decomposition found. The progression assumption failed outright. Mean fiber runs 7.70 g in week 1, 7.51 g in week 2, 8.40 g in week 3 and 6.76 g in week 4 — week 4 is the lowest of the four. Calories go 375.4, 373.4, 383.7, 367.0. Mean Score goes 15.03, 15.41, 14.56, 14.77. There is no ramp in any direction. Week 4 is not a more advanced shake than week 1; it is a different shake.
What Failed
Our tidy "goal equals cell" model has a hole in it, and it is a real one. Two goal names are reused across cells. "Recovery & Calm" covers both EY-W4 and MI-W4, which is defensible — those two cells average 380.5 and 378.1 kcal with identical mean protein at 32.08 g. But "Light & Vital Renewal" is applied to four different cells that do not resemble each other: WY-W4 averages 330.8 kcal, 26.0 g carbohydrate and 4.91 g fiber; WI-W4 averages 392.5 kcal, 40.4 g carbohydrate and 8.38 g fiber; WAY-W4 averages 362.9 kcal and scores 18; WAI-W4 averages 37.1 g protein and 14.2 g fat and scores 16.83. That is a 62 kcal spread and a 71% spread in fiber under one goal name, with Score ranging from 13 to 18. So the honest reading cuts against us twice. Goal is not an independent axis — but it is also not a reliable name for the cell it sits on. When the same words appear on four formulations that differ by 62 calories, the word is doing no work at all. We cannot claim goal is a clean naming layer when a quarter of the week-4 library shares one name and nothing else.
What We Learned
The system has three labels and two dials. Body type and week are the real inputs; goal is derived, and derived imperfectly. That is not automatically a design flaw — Zeevi and Berry both found that personalization power comes from combinations of inputs rather than from any single one, and our own numbers say the same thing: the intersection is worth about twice either dial alone. The flaw is in how we described it. We had been presenting goal as something a person selects, which implies a degree of freedom that does not exist in the database. We also learned that week is a rotation, not a progression. Nothing ramps across the four weeks, and week 4 has the least fiber. That may be the right design — variety has its own justification — but it is not the design our language implied.
What Changed at MOMS
Goal is being demoted from an input to a derived display field. It is now computed from body type and week rather than stored alongside them, which makes it impossible for a goal name to drift away from the formulation it describes. The four cells currently sharing "Light & Vital Renewal" are being renamed so that no goal name spans formulations more than 30 kcal or 2 g of fiber apart. Week is being described internally and externally as a rotation cycle rather than a progression, until we can show a ramp we actually built on purpose. We have opened a separate question — whether week 4 should differ from week 1 on fiber and fat, and in which direction — and we are not going to answer it by relabeling what we already have.
References
1. Zeevi D, Korem T, Zmora N, et al. Personalized Nutrition by Prediction of Glycemic Responses. Cell. 2015;163(5):1079-1094. 2. Berry SE, Valdes AM, Drew DA, et al. Human postprandial responses to food and potential for precision nutrition. Nature Medicine. 2020;26(6):964-973.
Keep reading
Previous experiment
Next experiment
Not sure which shake fits you? We match the formulation to your body type.
Find your body type