Why aren't calories and macros enough?

Because two meals with identical calories and macros can have radically different effects on insulin response, satiety, inflammation, and long-term disease risk. A 2,000-calorie day from salmon, lentils, olive oil, and mixed vegetables is not biologically equivalent to 2,000 calories from a frozen pizza and a soda, even if the protein, carb, and fat grams match. NOVA processing level, fiber matrix, polyphenol content, glycemic response, and micronutrient density all move independently of the macro sheet.

What are NOVA groups and why do they matter?

NOVA is a four-tier classification developed by a research group at the University of Sao Paulo that sorts foods by processing level rather than nutrient content. Group 1 is unprocessed or minimally processed (apple, rice, plain yogurt), Group 2 is culinary ingredients (olive oil, salt, sugar), Group 3 is processed foods (bread, cheese, canned fish), and Group 4 is ultra-processed (chips, soda, most cereals, frozen dinners). A growing body of evidence links Group 4 intake to adverse health outcomes even after adjusting for total calories, suggesting processing itself matters.

What is polyphenol diversity, and why track it?

Polyphenols are a broad family of plant compounds that include flavonoids, phenolic acids, stilbenes, and lignans. The health literature increasingly shows that variety across classes matters more than total milligrams — a diet with coffee, green tea, berries, olive oil, and cocoa hits five different polyphenol families, while a diet hitting a single class at high dose does not replicate the effect. Tracking diversity rather than a single total is closer to what the underlying biology rewards.

Why compute glycemic index at the meal level?

Because you almost never eat a single food alone. White rice has a high GI on its own, but paired with salmon and vegetables the mixed-meal glycemic load is meaningfully lower. Carb quantity, fiber, fat, protein, and cooking method all shift the response. Tracking GI per ingredient and then summing ignores these interactions. Tracking glycemic load at the meal level is closer to what the glucose curve actually does.

Does meal timing really change outcomes independent of what you eat?

Yes, for several outcomes. Eating window duration, circadian alignment of the last meal, and meal spacing all associate with metabolic, sleep, and body composition endpoints in the literature — even when total daily intake is matched. The exact magnitude is population-dependent and the research is still maturing, but the direction is consistent enough that an eating window of 14 hours ending at midnight is not interchangeable with a 10-hour window ending at 7 pm, even at identical calories.

Is tracking IARC carcinogen exposure from food a form of scaremongering?

Only if it's done badly. The IARC classification system is a consensus evaluation of evidence, not a dose-response model — Group 1 ('carcinogenic to humans') and Group 2A ('probably carcinogenic') are the strongest categories. Weekly cumulative exposure to processed meats, charred meat HAAs and PAHs, alcohol, and aflatoxin-prone grains is worth surfacing as a surveillance signal so users can make informed choices. The framing matters: this is awareness, not diagnosis, and single meals rarely move the needle.

How does AI read a meal photo?

A modern vision pipeline combines a multimodal LLM that identifies items and estimates portions from a single photo with a perceptual hash that detects near-duplicate meals so the same lunch photographed twice does not hit the model twice. Output includes per-item macros, confidence scores, and NOVA/GI estimates, which are then enriched against reference databases like USDA FoodData Central, Open Food Facts, and a glycemic index table. The user confirms or edits, and low-confidence items are flagged for review rather than silently accepted.

Can a tracker tell me my dietary pattern?

With a week of reasonably complete logs, yes. Patterns like mediterranean, ketogenic, high-protein, plant-based, high-processed, balanced, low-carb, and high-carb can be inferred from the distribution of foods logged over a week. The point is not labeling you — it's showing you what your diet actually is, which is often different from what you think it is.

Nutrition Intelligence: What's Actually In Your Food Beyond Calories and Macros

You eat 2,000 calories. 120 grams of protein, 220 grams of carbs, 70 grams of fat. Your tracker shows four numbers and a green checkmark.

So does your friend who ate frozen pizza, a protein shake, and a candy bar that hit the same macros.

Are those two days the same? Your tracker thinks so.

Why Calories and Macros Is the Floor, Not the Ceiling

The calorie-plus-macro model has been the default for a decade because it is simple, measurable, and mostly adequate for managing weight in the short term. If you eat fewer calories than you burn and get enough protein, you will lose fat. That part is not in dispute.

What is in dispute is whether four numbers are enough to describe the nutritional content of what you ate. The biological answer is no, and the gap between what macros capture and what food actually does to your body is the space where the last fifteen years of nutrition research has lived.

Consider a concrete example. Two lunches, both 650 calories, both 40 grams of protein, 65 grams of carbs, 22 grams of fat. Lunch A is grilled salmon, quinoa, olive oil, and a mixed salad with berries. Lunch B is a fast-food grilled chicken sandwich with fries and a diet soda. The macro line is identical. Almost nothing else is.

Lunch A has NOVA-1 and NOVA-2 ingredients only. It contains flavonoids from the berries, phenolic acids from olive oil, fiber from the whole grain and vegetables, and omega-3 fatty acids from the fish. Its glycemic response is blunted by the fat and protein matrix. Its micronutrient profile covers roughly a third of daily magnesium, vitamin K, folate, and potassium needs. It leaves you satiated for four hours.

Lunch B is roughly half NOVA-4 by calorie weight. The bun has emulsifiers and added sugar. The fries have acrylamide and oxidized seed-oil residues. The diet soda contains artificial sweeteners whose long-term effects on gut microbiome and glucose homeostasis are still under active investigation. Fiber content is around a quarter of Lunch A’s. Polyphenol diversity is essentially zero. Micronutrient density is a fraction of Lunch A’s, especially for magnesium, potassium, and vitamin K. Satiety tends to be shorter.

The dashboards call these meals equivalent. They are not. The question this pillar asks is: what would a tracker look like if it captured the real differences?

The answer has seven parts: processing level, polyphenol diversity, meal-level glycemic response, meal timing and eating window, carcinogen exposure, micronutrient adequacy, and the quality of the analysis pipeline that extracts all of this from a meal photo in the first place. Plus a final layer — dietary pattern classification — that gives you language for what your diet actually is. Each section below unpacks one of these, and the related spokes in this cluster go much deeper into each.

For the lifting-and-recovery side of this story, see the adaptive training intelligence guide. For the underfueling side — where nutrition and training collide — see the energy availability guide. And for the accuracy-and-trust side of any composite metric, see when to trust your health score. This pillar focuses on the nutritional dimensions themselves.

NOVA Food Processing Groups: Four Tiers That Change Health Outcomes

NOVA is a classification system developed by a research group at the University of Sao Paulo, led by Carlos Monteiro, that sorts foods by processing rather than by nutrient content. It has become the de facto framework in population nutrition research because it predicts health outcomes — cardiovascular disease, type 2 diabetes, all-cause mortality — better than macro-only frameworks do, even after adjusting for calories.

The four groups are:

Group 1 — Unprocessed or minimally processed. Foods in their natural state or with processing limited to drying, crushing, grinding, pasteurizing, or freezing. Fresh fruit, vegetables, legumes, grains, fresh or frozen meat, plain yogurt, eggs, coffee beans. The vast majority of daily calories in a traditional Mediterranean or Okinawan diet fall here.

Group 2 — Culinary ingredients. Substances extracted from Group 1 foods and used to season or cook — olive oil, butter, salt, sugar, vinegar. Rarely eaten alone, used in small amounts to prepare Group 1 foods.

Group 3 — Processed foods. Products made by adding Group 2 ingredients to Group 1 foods and then preserving or modifying — bread, cheese, canned fish in oil, home-style pickles, cured meats without novel additives. Recognizable as food, made in a kitchen that could be scaled up or down.

Group 4 — Ultra-processed. Industrial formulations of substances extracted or synthesized from foods plus additives (emulsifiers, colorings, flavorings, bulkers) engineered to be hyperpalatable, shelf-stable, and usually branded. Sodas, packaged snacks, most breakfast cereals, instant soups, frozen dinners, many “protein” bars, most commercial breads. The ingredient list often contains items you would not find in a home kitchen.

The outcome data on Group 4 intake is where NOVA earns its place in a serious tracker. Large prospective cohorts consistently find that higher Group 4 consumption associates with higher incidence of cardiometabolic disease and higher mortality, even after adjusting for total calories, macronutrients, and body mass index. The effect survives adjustment for what you would expect to be obvious confounders, which suggests processing is doing something independent of energy content — possibly through the disruption of the food matrix, possibly through additives, possibly through hyperpalatability driving overconsumption in a different eating regime. The literature is still sorting out the mechanism. The association itself is robust.

A serious nutrition tracker assigns a NOVA group to every logged item. The per-meal quality indicator should weight meals with more than half their calories from Groups 1 and 2 as high-quality, meals with 20 to 50 percent from those groups as moderate, and meals below 20 percent as low. The daily summary should carry a calorie-weighted NOVA average and a weekly trend. Users don’t need a lecture — they need the number visible so they can act on it.

For a deeper treatment of NOVA, the common misinterpretations (“all processed food is bad” is not what the literature says), and how to interpret a weekly NOVA-4 percentage, see NOVA Groups: Why ‘Ultra-Processed’ Isn’t the Same as ‘High-Calorie’.

Polyphenol Diversity: Why Variety of Antioxidant Classes Matters

Polyphenols are a vast family of plant compounds with antioxidant, anti-inflammatory, and microbiome-modulating effects. The common classes include flavonoids (subdivided into flavonols, flavanones, anthocyanins, isoflavones, and catechins), phenolic acids (chlorogenic acid in coffee, hydroxycinnamic acids in olive oil), stilbenes (resveratrol in wine and berries), and lignans (flaxseed, sesame).

The naive approach to polyphenol tracking is to sum total polyphenol milligrams across the diet and report one number. This is better than ignoring them but misses the more interesting signal. The research that most reliably separates “protective” diets from “neutral” ones does not rely on total polyphenol dose — it relies on diversity across classes. A diet with coffee, green tea, red wine in moderation, berries, cocoa, and extra-virgin olive oil hits five or six distinct polyphenol families. A diet hitting a single class at high dose does not replicate the effect in population studies.

The biological intuition is that different polyphenol classes hit different targets. EGCG from green tea inhibits different enzymes than anthocyanins from berries do. Chlorogenic acid from coffee affects glucose kinetics differently than resveratrol does. A diverse dietary polyphenol intake hits a broader set of pathways, and the microbiome metabolizes each class into distinct downstream compounds. Redundancy does not substitute for breadth.

Per-item polyphenol enrichment using a reference like the Phenol-Explorer database, resolving each food to canonical polyphenol contents and scaling by portion, produces a per-meal polyphenol profile. The weekly summary is not “you ate X milligrams” — it is “you hit 6 of 8 canonical polyphenol classes this week, and class 3 (flavonols) is consistently absent.” That is a more useful signal because it maps directly to a dietary change: eat more kale, apples, or onions.

The flipside is measurement uncertainty. Polyphenol content of a given food varies with variety, ripeness, storage, and cooking. A strawberry is not a fixed polyphenol bomb — its anthocyanin content differs by a factor of two or three depending on cultivar and freshness. Any tracker that reports polyphenols to three decimal places is overclaiming. The right framing is tier-based — high, medium, low, or absent for each class — with explicit uncertainty.

For the full treatment, including the eight canonical polyphenol classes the research most consistently rewards, see Polyphenol Diversity: Why Variety Beats Milligrams for Antioxidant Intake.

Glycemic Index at the Meal Level, Not the Ingredient Level

Glycemic index is a measure of how quickly a carbohydrate-containing food raises blood glucose in a standardized setting, relative to pure glucose or white bread. It is usually reported per food, which is how it was first developed and how most reference tables present it. The problem with per-food GI is that almost nobody eats a single food.

Meal-level glycemic load captures what the food mixture actually does. Glycemic load is roughly (GI × grams of available carbohydrate) / 100 per food, summed across the meal. The sum matters because the interactions matter. Fat and protein slow gastric emptying. Fiber blunts the absorption curve. Cooking method changes starch digestibility — al dente pasta has a lower effective GI than overcooked pasta. Freezing and reheating bread increases resistant starch. The GI of a single ingredient in isolation is not what shows up in your glucose curve when that ingredient is eaten with everything else on the plate.

A practical pipeline resolves each logged item to a GI estimate (from a reference database or an LLM-assigned low/medium/high tier mapped to nominal values like 40 / 62 / 80) and computes per-meal glycemic load. The meal-level number is what matters for day-to-day decisions: a meal with GL above 20 tends to produce a stronger glucose excursion, GL between 10 and 20 is moderate, GL under 10 is mild.

The important caveat — and why GI is a context signal rather than a strict rule — is individual variability. Continuous glucose monitor data over the last five years has consistently shown that personal glycemic responses vary substantially across people for the same food. The same banana can spike one person to 160 and barely move another. For people with no CGM, meal-level GL is still a useful population-averaged proxy. For people with CGM, the right move is to shift from population GI tables to their own measured responses.

For the full methodology, including how mixed meals are computed, how the research on individualized response changes the picture, and the limits of per-food tables, see Glycemic Index Per Meal (Not Per Food): How Mixed Eating Changes the Response. For comparisons to continuous glucose monitors as a measurement layer, see best Levels CGM alternatives for metabolic health tracking.

Chrono-Nutrition: Eating Window, Meal Spacing, and Circadian Alignment

Chrono-nutrition is the study of how the timing of food intake — independent of what is eaten — affects metabolic, sleep, and body composition outcomes. The field is still maturing, but the direction of the evidence is consistent enough to include in any serious tracker.

Three measurable quantities matter at the tracker level:

Eating window. The hours between the first and last calorie-containing intake of the day. Shorter windows (around 8 to 12 hours) associate in the literature with improved insulin sensitivity, lower fasting glucose, and modest effects on body weight, independent of total caloric intake. Windows longer than about 14 hours, especially when extending past 10 pm, show the inverse pattern.

Meal spacing. The average gap between meals. Very short gaps (constant grazing) and very long gaps (one or two large feedings) each have their own metabolic signatures. Meal spacing of 3 to 5 hours is associated with more stable glucose and satiety regulation for most people, though individuals with specific metabolic contexts may do better at the extremes.

Circadian alignment. The clock position of the last meal relative to biological night. Eating near habitual bedtime, regardless of total eating window length, associates with worse sleep quality, attenuated fasting lipolysis, and a larger glucose response to the late meal than the same meal would produce earlier in the day. The ISSN and ACSM position statements on nutrient timing both recognize this as an underappreciated variable.

The tracker implementation is straightforward if the meal logs have accurate timestamps. Compute nutrition_eating_window_hours as the hours between first and last meal, nutrition_meal_spacing_hours as the mean gap, and flag when the last meal of the day lands within two hours of typical sleep onset. A weekly summary that says “your eating window averaged 13.2 hours this week and your last meal averaged 9:47 pm” is more actionable than any macro breakdown for the specific outcome of metabolic health.

One caveat worth repeating: chrono-nutrition effects are real but smaller in magnitude than total intake and food quality. Do not let eating-window optimization become an excuse for ignoring what goes into the window. The order of priority is: food quality first, total intake second, timing third.

For the full treatment — ISSN and ACSM position statements on nutrient timing, the research on time-restricted eating for non-weight outcomes, and how to measure whether a tighter window is actually changing anything for you — see Chrono-Nutrition: Eating Window, Meal Spacing, and Why Timing Changes Outcomes.

Carcinogen Exposure: IARC Groups, Weekly Cumulative Counters, Realistic Interpretation

The International Agency for Research on Cancer (IARC) classifies substances and exposures into four groups based on the strength of evidence for carcinogenicity in humans. Group 1 is “carcinogenic to humans,” Group 2A is “probably carcinogenic,” Group 2B is “possibly carcinogenic,” Group 3 is “not classifiable.” The classification is about the evidence of carcinogenicity, not the magnitude of risk — Group 1 includes processed meat, alcohol, tobacco, and asbestos, and nobody thinks a sandwich equals asbestos exposure.

The relevant food-system exposures are limited but worth tracking:

Processed meat (Group 1). Bacon, ham, salami, hot dogs, most deli meats. The IARC Working Group’s 2015 review estimated roughly an 18 percent increase in colorectal cancer incidence per 50 grams per day of processed meat. The effect is dose-dependent and cumulative over years. Occasional consumption is not a five-alarm fire; daily consumption at above-50g levels is a surveillance signal.

Red meat (Group 2A). Beef, pork, lamb when not processed. Evidence for carcinogenicity is weaker than for processed meat. The association exists but is more variable across studies and more confounded by cooking method and overall dietary context.

Alcohol (Group 1). A significant population-level carcinogen, with clearly established associations with cancers of the mouth, throat, esophagus, liver, colon, and breast. Dose-dependent with no well-established safe level for cancer specifically, though cardiovascular interactions complicate the full picture.

High-temperature cooking byproducts. Heterocyclic amines (HAAs) and polycyclic aromatic hydrocarbons (PAHs) form when meat is cooked at high temperatures, especially grilled or charred. IARC classifies several as probable or possible carcinogens.

Aflatoxins (Group 1). A family of fungal toxins that contaminate improperly stored grains and nuts in humid climates. In wealthy countries with regulated food supplies, background exposure is low; in regions without such controls it is a significant liver cancer driver.

The right tracker framing is weekly cumulative exposure with context, not a per-meal alarm. Processed meat servings over the trailing seven days is a useful surveillance counter. Weekly alcohol units is a useful surveillance counter. Grilled or charred meat frequency is a softer counter but worth showing. The interpretation has to be clear: these are awareness signals so the user can make informed choices, not diagnostic signals, and weekly totals matter far more than any single meal.

For the full methodology and the most common misinterpretations — including the “processed meat equals asbestos” misread that the 2015 IARC report triggered and why it is wrong — see IARC Group 1 and 2A Carcinogens in Food: Weekly Exposure, Not Single Meals. If you have specific cancer-risk concerns, that content is a primer for an informed conversation with an oncologist or dietitian — not a substitute for one.

Micronutrients: 35-Nutrient Tracking, Supplement Integration, and Percent Daily Value Reality

The standard macro view — protein, carbs, fat, calories — ignores roughly 30 micronutrients that move independently of the macro sheet and that most people are deficient or marginal in for at least a subset of them. Tracking the major vitamins (A, B-complex, C, D, E, K), the major minerals (calcium, iron, magnesium, potassium, sodium, phosphorus, selenium, zinc, copper, manganese, chromium, molybdenum, iodine), and a handful of conditionally essential quantities (choline, omega-3 EPA and DHA, fiber subtypes) is what gets you to 35.

The implementation is not complicated in principle — the USDA FoodData Central database carries per-100g micronutrient profiles for most foods, and Open Food Facts fills in for branded and international products. The complications are at the edges:

Bioavailability is variable. Iron from spinach is absorbed at 2 to 20 percent depending on context (vitamin C presence, phytate content, other minerals competing for transport). Iron from beef is absorbed at 15 to 35 percent. A raw “15 mg iron” log line is not equivalent to a bioavailable-15-mg actual delivery. Any tracker reporting micronutrient totals without caveats is overclaiming.

Percent Daily Value is a regulatory construction, not a personal target. The FDA %DV is set to cover 97.5 percent of healthy adults. For many nutrients (vitamin D, magnesium, omega-3), the %DV is below what current research suggests is optimal rather than just adequate. Hitting 100% of %DV is not the same as hitting your personal optimal, and for some nutrients the gap is material.

Supplements are a separate input stream. Supplements contribute micronutrients on top of food intake and should be tracked in a separate system that aggregates into the daily total. A serious nutrition platform logs branded supplements against a reference database like the NIH Dietary Supplement Label Database, stores per-serving micronutrient profiles, and adds to the food-derived totals for a unified view. A supplement tracker that doesn’t integrate with the food tracker is giving you two half-pictures.

The useful framing for users is not “hit 100% of each %DV daily,” because that’s both unrealistic and not quite the right target. It is “over the trailing 14 to 30 days, which nutrients are consistently below 70% of DV on average?” That identifies the real gaps. A three-day spinach kick does not offset thirty days of iron below 40%.

For the full treatment — including which micronutrients are most commonly deficient by population, which are commonly over-supplemented, how %DV compares to the DRI and to research-based targets, and when to treat a gap as a signal to escalate to bloodwork — see Tracking 35 Micronutrients: Catching Deficiencies Before They Become Symptoms. For the bloodwork side of the same conversation, see what your blood work tells you about fitness.

How AI Reads a Meal Photo: Dual-Model Vision, Perceptual Hashing, Confidence

Photo-based meal logging is the single biggest compliance lift in nutrition tracking over the last five years. Barcode scanning is great for packaged food and useless for a plate of homemade curry. Text entry is slow and produces bad data when users are rushed. Photos are fast and generate structured data — if the vision pipeline is good enough.

A modern meal-photo pipeline has four stages:

1. Photo intake and quality assessment. When a user uploads a photo, compute a quality score (resolution, aspect ratio, file size) and compute a perceptual hash. The quality score gates whether the photo is eligible for training-data collection. The hash is a 64-bit fingerprint that lets the system recognize near-duplicates of photos from the last N days.

2. Near-duplicate detection. If the perceptual hash is within a small Hamming distance (typically 5 bits out of 64) of a recent photo, the pipeline can reuse the previous analysis instead of calling the vision model again. This catches the case where the user photographs the same meal twice, or where a partial re-upload happens after a network hiccup. Cost savings aside, it also ensures consistency — the user doesn’t see the same plate produce different macros on re-log.

3. LLM vision analysis. For new photos, a multimodal LLM (Claude Sonnet, GPT-4o, or similar-grade models) identifies items, estimates portions using plate-size and hand-size anchors, and produces structured output — per-item macros, NOVA group estimate, GI tier (low/medium/high), and a confidence score per item. A dual-provider architecture with A/B testing lets the pipeline compare two different models on the same photo stream and keep the one that’s currently more accurate for a given food class. When the primary fails, a fallback provider handles the request so the user doesn’t see an error.

4. Confidence-gated review. Items above a confidence threshold (typically 0.85) auto-confirm. Items below the threshold are flagged for user review. Pending items are not silently dropped into the daily total — they require an explicit confirm, because propagating a low-confidence guess into micronutrient and NOVA summaries contaminates the very signals the whole pipeline exists to produce.

After the user confirms, the item goes through an enrichment pipeline: USDA macro validation (correcting LLM estimates if they deviate by more than about 25 percent from the FDC reference), Open Food Facts NOVA and Nutri-Score enrichment for packaged items, glycemic index lookup, and polyphenol resolution. Each enrichment carries a data-source label — usda_verified, off_matched, llm_only — so downstream analytics know how much to trust each field.

This is also where training-data collection lives. Meals with photos of sufficient quality, where the user either confirmed or corrected the LLM output, are eligible (with explicit opt-in consent) to be copied to a training bucket with provenance metadata. Over time that corpus becomes the input to a nutrition-specific model that will, in principle, outperform general-purpose vision LLMs on the exact task. The ethics layer is not optional: consent is explicit, withdrawal deletes the training copies, and the collection is gated on quality so noisy inputs do not degrade the training set.

For a deeper technical treatment, including dual-model A/B testing, perceptual hash collision behavior, and why confidence-gated review is the non-obvious part that keeps the system honest, see How AI Reads a Meal Photo: Dual-Model Vision, Perceptual Hashing, and Confidence. For the broader reason this matters for nutrition-tracker evaluation, see best MyFitnessPal alternatives that actually understand your diet and best Cronometer alternatives for serious nutrition tracking.

Dietary Pattern Classification: The Eight Patterns That Emerge From Real Logs

Once a week of reasonably complete logs is in the system, a classifier can assign a dietary pattern from a small set of canonical categories. The patterns that most consistently emerge from real user data are: mediterranean, ketogenic, high-protein, plant-based, high-processed, balanced, low-carb, and high-carb.

The classifier is not just a macro calculator. Ketogenic is not “carbs under 50g” — it requires the overall food composition to look ketogenic (high fat from whole sources, adequate protein, minimal processed carbs). High-processed is not a calorie threshold — it’s when NOVA-4 calories dominate the week regardless of macros. Mediterranean is not “has olive oil” — it requires the pattern of fish, vegetables, legumes, whole grains, and olive oil to hold across most meals.

The practical output is a weekly label with a short description (“Your pattern this week looks mediterranean — most meals emphasized fish, legumes, vegetables, and olive oil”) and an eight-week history so the user can see whether their pattern is stable or drifting. A stable pattern with good biomarkers is fine. A drift from plant-based to high-processed over three weeks is a signal worth surfacing.

The classification is weekly, not daily, because daily variance is too high — Sunday brunch doesn’t redefine your dietary pattern, and a tracker that tells you “your pattern today is high-processed” because of a single pizza is overreacting. A rolling week smooths the noise.

The reason this matters is that dietary pattern, more than any single nutrient or macro, is what the long-term outcome research tracks. “Mediterranean adherence” is a meta-variable that subsumes NOVA, polyphenol diversity, fiber intake, and omega-3 balance into one label. Giving users that label is more useful than giving them a 35-row nutrient table they cannot integrate into a single decision.

For how the eight patterns are defined, what drift between patterns looks like, and why the classification is better done by an LLM than by a rules engine, see Your Diet Has a Pattern — Here Are the 8 Common Ones.

Putting It Together

A nutrition tracker that does the work described above does not produce a single number — it produces a structured, uncertainty-aware picture:

Daily calories and macros (the floor)
A calorie-weighted NOVA group average, with daily NOVA-4 percentage
A polyphenol-class diversity counter (classes hit this week, classes missed)
Per-meal glycemic load with daily total, and the trailing-week glucose-curve-friendly meal count
Eating window hours, meal spacing, and late-meal count per week
Weekly IARC Group 1 and 2A surveillance counters (processed meat servings, alcohol units, charred-meat count)
A 35-nutrient trailing-30-day percent Daily Value grid with gap identification
A weekly dietary pattern label with an eight-week trend
Per-meal photo analysis with confidence scores and enrichment provenance

None of this replaces a registered dietitian. What it does is give the user the raw material to ask better questions of one, and to run shorter-cycle personal experiments between visits. The tracker’s job is surveillance and pattern recognition — not diagnosis, not prescription, not treatment. If the 35-nutrient view shows persistent iron below 40% for three months, that is a signal to get a ferritin test, not to self-supplement blindly. If the NOVA-4 percentage is 60% across twelve weeks and energy is persistently low, that is a signal to change something in the food environment, not to add a pre-workout.

For the related conversation on training: the adaptive training intelligence pillar covers how per-muscle volume tolerance and readiness interact with nutrition. For the underfueling and RED-S side, which sits at the nutrition-plus-training boundary, see the energy availability guide. For the accuracy layer — which inputs deserve how much trust, and why every score should carry a confidence value — see when to trust your health score.

Omnio is built around this stack. The food photo analysis feature handles the vision pipeline described above. The supplement tracking feature handles the branded-supplement side of the micronutrient total. The nutrition page documentation is the shortest version of how it all fits together in the product.

More in This Series

This is the pillar post for the nutrition intelligence cluster. The full set of spokes goes much deeper into each section above:

NOVA Groups: Why ‘Ultra-Processed’ Isn’t the Same as ‘High-Calorie’ — coming soon
Polyphenol Diversity: Why Variety Beats Milligrams for Antioxidant Intake — coming soon
Chrono-Nutrition: Eating Window, Meal Spacing, and Why Timing Changes Outcomes — coming soon
Glycemic Index Per Meal (Not Per Food): How Mixed Eating Changes the Response — coming soon
Your Diet Has a Pattern — Here Are the 8 Common Ones — coming soon
IARC Group 1 and 2A Carcinogens in Food: Weekly Exposure, Not Single Meals — coming soon
Tracking 35 Micronutrients: Catching Deficiencies Before They Become Symptoms — coming soon
How AI Reads a Meal Photo: Dual-Model Vision, Perceptual Hashing, and Confidence — coming soon

Existing posts in the same cluster: best MyFitnessPal alternatives that actually understand your diet, best Cronometer alternatives for serious nutrition tracking, best MacroFactor alternatives for adaptive nutrition tracking, best Levels CGM alternatives for metabolic health tracking, and most nutrition trackers count calories — ours understands your diet.