A Federal Definition Won't Make Ultra-Processed the Right Number

Two issues ago I wrote that ultra-processed was never the number that mattered, that a 200,000-person, four-decade Harvard study found diet quality dominates processing level when you’re predicting heart disease, diabetes and death. On August 3rd, five days after I filed that piece, the FDA sent its long-awaited definition of “ultra-processed food” to the White House’s Office of Information and Regulatory Affairs for review. The federal government is about to make official the exact category I spent 700 words arguing isn’t the right one to build around.

I don’t think that makes me wrong, and I don’t think it makes the FDA wrong either. It makes clear that regulation and personalization are two different jobs, and a single category can’t do both well.

Here’s what’s actually in the document, as far as the reporting on it goes. FDA and USDA spent the past year fielding public comment on three competing ways to define “ultra-processed”: a processing-based approach built around industrial techniques, an ingredient-based approach built around specific additives, and a nutrient-based approach built around added sugar, sodium, saturated fat and energy density. The white paper reportedly lands on a hybrid: an initial screen on ingredients or processing method, nutrition criteria layered on top, and exemptions for foods everyone agrees are fine despite qualifying on a technicality, whole-grain bread, canned beans, fortified cereal. It’s a notice, not a rule. Nothing changes on a package tomorrow. But it’s the clearest signal yet that “ultra-processed” is about to become a legal category with real downstream consequences: front-of-pack labeling, school lunch procurement, maybe eventually what a manufacturer is allowed to call a food at all.

That’s a completely reasonable thing for a regulator to build. A government needs one enforceable line a court can apply to a factory production line, and processing method or ingredient list is about the only thing you can audit at scale. You cannot regulate “this person’s overall diet quality over the past year.” You can regulate “does this product contain more than a defined threshold of these additives.” The category has to be legible to inspectors and lawyers, not accurate to physiology.

That’s precisely the trap I flagged for an app like ours. Mr BITE isn’t writing labeling law. It’s trying to tell one person, today, whether the plate in front of them is good for them. On that job, the Harvard data was blunt: within the same quality tier, factory-made and home-cooked diets carried about the same risk. A federal definition, however carefully hybridized, is still answering “was this made a certain way,” not “is this good for the person eating it.” Those two questions correlate, loosely, which is exactly why processing-based proxies survive as long as they do: they’re right often enough to feel earned. Loose correlation is a bad foundation for the one number an app puts in front of someone every time they log a meal.

If anything, watching the FDA spend a year testing three different definitions and landing on “combine all three, plus exceptions” is the best argument I’ve seen for how genuinely hard this category is to pin down cleanly, even with a legal team and a public comment period behind you. If it takes that much machinery to make “ultra-processed” hold together as a regulatory line, it was never going to hold together as a badge on a single meal photo. The two problems aren’t the same size, and treating them as if they were is how you end up shipping a green checkmark on a fortified cereal and a red flag on a bowl of white rice and butter, which is exactly the failure mode I described in Issue 3.

So no, this doesn’t change anything about the argument. It sharpens it. A category built for one job, drawing an enforceable line for inspectors, is arriving with a mandate. That’s worth watching. It’s still the wrong number for the other job, telling a person what to eat today, and no amount of regulatory rigor changes what the data already said about that.

BITE of the Week: The Meal Plan That Dropped a Snack, and Everything Rebuilding It Found

Version status first: 1.6.0 (iOS build 98, Android build 129), cut August 4th and covered last issue, is still the build sitting in the review queue as far as anything since then shows. The live app in your hands is 1.5.0.

The real story of this cycle started with one of us just trying to build a meal plan on August 5th and getting back four dinners, two of them the same dish on non-adjacent days, no snack despite asking for one, and a wall of warnings pasted in verbatim instead of actual advice. Worse: a library of more than 45 eligible recipes sat behind that plan, unused. The first diagnosis blamed a protein filter quietly emptying the candidate pool. Wrong: replaying the real ranker over the real rows showed 41 to 47 candidates reaching it every time, none dropped by that filter at all. The actual cause, found only after that wrong turn, was almost funny: the model asked for 2 candidate recipes per meal slot, an unrelated safety floor clamped that up to 3, and a full week got built from 3 recipes because nothing in its own instructions distinguished “how many recipe cards to show a user” from “how many options to plan from.”

Fixing the real cause took more than one patch. The candidate shortlist now scales with how many days a plan actually needs to cover, instead of a flat floor. A recipe spanning multiple days now has to run a contiguous stretch from its cook day, no more the same dish landing on Monday and Friday, tied to whether the user actually preps ahead. Three rounds of the model re-checking and revising its own draft got replaced with draft once, check deterministically for free, repair once. And model “thinking” got turned off for plan drafts entirely: the original failure happened with reasoning at max effort, which is a pretty direct proof that more deliberation wasn’t going to rescue bad inputs. A separate, narrower change lets Mr Bite compose a simple recipe, plain chicken and rice, scrambled eggs, when the catalog genuinely has nothing to offer, reversing a blanket “never invent a recipe” rule for exactly that one gap, verified before it’s ever shown and clearly labeled as Mr Bite’s own.

Rebuilding the recipe supply also surfaced two allergy bugs that had nothing to do with any of this and had been sitting in production the whole time. A profile allergy of “seafood” wasn’t expanding to the actual fish and shellfish names in a dish, so a user allergic to seafood could still be served a recipe literally named Grilled Salmon; of 11 profiles carrying any allergy, 5 used a term that resolved to nothing. And an allergen match at the final save step used to be a warning the model could clear on its own, with nothing verifying a person was ever shown it; it now refuses outright instead. Neither made it into a version number yet. Both are merged to main, safety fixes that outrank whatever feature ships around them next.

Also in this window, quieter but worth a line: a step-tracking bug where a day with no recorded steps reported a confident zero instead of being left out, making your own past habit history look worse than it actually was. That’s fixed too, along with widening the Progress tab’s data window so the 3-month and 6-month views actually have data behind them.

Three Quick Bites

1. The FDA sent its ultra-processed food definition to the White House for review. On August 3rd, FDA and USDA forwarded a white paper defining “ultra-processed food” to the White House’s regulatory review office, the clearest sign yet that a federal definition is coming after a year of public comment weighing processing method, ingredient list and nutrient content as competing approaches. It’s a notice, not a rule, so nothing changes on a label tomorrow, but it’s the news behind the essay above. FoodNavigator

2. A Rockefeller University study found a diet amino acid helps the immune system spot cancer and viruses. Published July 30th in Cell, the research found that arginine restriction stalls production of MHC-I, the protein that flags a mutated or infected cell to the immune system, while a higher-arginine diet suppressed colon tumors and improved flu and COVID resistance in mice. It’s animal data, and the researchers are explicit that nobody should start supplementing arginine on the strength of it, but the mechanism is a genuinely new piece of “what food does to your immune system” science. ScienceDaily

3. One 75-minute workout a week matched three shorter sessions for cutting belly fat. A 315-person, four-month randomized trial out of the University of Hong Kong found that adults with central obesity doing one weekly session of brisk interval walking lost about as much visceral fat and gained about as much fitness as a group doing the identical total time split across three days a week. Worth remembering next time an app, including possibly this one, tells you consistency has to mean daily. ScienceDaily

Tool of the Week: FDA Substances Added to Food

Given what this issue is about, the pick writes itself. Substances Added to Food, formerly EAFUS, is the FDA’s free, public, searchable inventory of roughly 4,000 ingredients legally added to the U.S. food supply: food additives, color additives, and substances Generally Recognized As Safe, each with its regulatory status and the rule that governs it. For readers: it’s the fastest way to look up what an unfamiliar ingredient on a label actually is and why it’s allowed to be there, instead of trusting whatever a scary-sounding name implies. For builders: it’s a genuine primary source if you ever want your app to explain an ingredient rather than just flag it, which is a more useful feature than a red dot next to a word nobody recognizes. fda.gov. Free, public, U.S. government data.

Thanks for reading. If you’d rather track what’s actually in your food than wait for Washington to agree on a definition, Mr BITE is free on iOS and Android.

Collins, building Mr BITE