Working research document, copied from the author’s notes on 2026-09-06. Rough, long, and unedited apart from removing personal details. The finding page summarizes it.

Whole Foods Evidence Base for Meal Scoring

Compiled: 2026-09-05 Purpose: a machine-encodable evidence base on whole food classes and ingredients — the actual foods, not additives — for judging whether a meal is healthy. Designed to be applied to ~800 real meal-delivery ingredient lists.

Subject the scores are calibrated for: a healthy general-population adult. [personal profile and health goals removed]

A. How to read the evidence tiers

Tier	Meaning
beneficial	Converging evidence across designs that including this food improves outcomes or displaces something worse.
neutral	No credible evidence of net benefit or harm at realistic intake; score reflects only macronutrient/energy-density considerations.
harmful	Consistent evidence of a net negative signal at realistic intake.
contested	Competent researchers reading the same literature reach opposite conclusions, and the disagreement is methodological rather than resolvable by citing more studies.

confidence is separate from tier and describes how much the score could move if better evidence arrived, not how big the effect is.

B. The evidence hierarchy actually used here

Ordered by how much weight a claim gets:

Meta-analysis of RCTs with hard endpoints — vanishingly rare in whole-food nutrition. Exists for saturated-fat replacement and essentially nothing else in this document.
RCTs with validated intermediate endpoints (LDL-C from controlled feeding, blood pressure, MRI-measured visceral fat, HOMA-IR). This is the best evidence most food classes have. LDL-C in particular has a well-characterised causal link to events via Mendelian randomisation and statin trials, so an LDL change can be translated into a defensible risk estimate.
Objective-biomarker prospective cohorts (circulating fatty acids, not FFQ). Materially better than standard cohorts because they remove the dominant measurement-error term.
FFQ-based prospective cohorts. The workhorse of nutritional epidemiology and the weakest link in almost every popular nutrition claim. See §C.
Case-control studies. Recall bias makes these systematically worse for diet.
Mechanistic / in vitro / animal. Hypothesis-generating only. Nutrition has an unusually bad track record of mechanism failing to predict outcome (β-carotene, vitamin E, folate, antioxidant supplements — several of which caused net harm in trials).

C. Why FFQ-based nutritional epidemiology is weak — the four failure modes

Because most of the claims in this document rest on it, and because the correct posture is neither credulity nor dismissal:

Healthy-user confounding. Eating any food that is culturally coded as healthy is one of the strongest available proxies for not smoking, exercising, higher education, higher income, more preventive care, and lower alcohol. Covariate adjustment uses error-laden measures of these, so residual confounding survives adjustment. This biases “healthy” foods toward apparent benefit and “unhealthy” foods toward apparent harm — i.e. it inflates effects in the direction everyone already expects.
Measurement error. Food frequency questionnaires ask people to recall average intake over a year. Validation studies routinely find correlations of ~0.4–0.6 with reference methods. This attenuates true associations toward null and creates spurious ones through differential misreporting (under-reporting of energy intake is systematic and correlates with BMI).
The noise floor. For observational nutrition, relative risks between roughly 0.8 and 1.25 should be treated as inside the noise floor — plausibly generated by residual confounding and measurement error alone. This single heuristic disposes of a large fraction of nutrition headlines. It is why “18% higher risk” per 50 g/day of processed meat deserves a careful look at absolute risk rather than reflexive alarm, and equally why “20% lower mortality” for nuts or olive oil should not be taken at face value.
Multiplicity and flexible analysis. Large cohorts support thousands of food-outcome comparisons with many defensible covariate sets. Publication and analytic flexibility select for significant findings.

What survives this critique: effects that are large (RR <0.6 or >1.7), dose-responsive, consistent across populations with different confounding structures, corroborated by an intermediate-endpoint RCT, and mechanistically coherent. Very few food-outcome claims clear all five bars. Trans fat does. Most do not.

D. Absolute-risk translation — the arithmetic used throughout

A relative risk is uninterpretable without a baseline. The baselines used in this document:

Outcome	Approximate baseline for a US adult	Note
Lifetime risk of any cancer diagnosis	~40%	dominates all “cancer risk” translations
Lifetime risk of colorectal cancer	~4–5%	the relevant baseline for processed/red meat claims. [Correction 2026-09-05: this band is imprecise. ACS puts lifetime colorectal cancer risk at 5.3%; SEER’s 2021–2023 figure is 3.9%. The two differ by source and method, not by error — the meal-scoring work uses ~4% (SEER-anchored, rounded) throughout, so pick one and say which.]
Lifetime risk of cardiovascular disease	~30–40%
10-year ASCVD risk, healthy adult male	<1%	near-term event prevention is essentially irrelevant for this subject
Regulatory de minimis cancer risk	1 in 1,000,000 (1×10⁻⁶)	a threshold for involuntary population exposure, NOT a personal decision threshold

The translation rule: relative change × baseline = absolute change. A “18% increased risk” of colorectal cancer moves a ~4.5% lifetime risk to ~5.3% — about 0.8 percentage points, or roughly 1 additional case per 125 people exposed for life. A “18% increased risk” of all-cause mortality would be an entirely different magnitude of claim. The percentage alone tells you nothing.

The corollary for this subject specifically: the subject is a healthy, active young adult. The adult’s near-term absolute risk of essentially every outcome in this literature is very low. The meaningful levers for the adult are (a) body composition and energy balance, (b) sodium and blood pressure trajectory, (c) protein sufficiency for training, (d) fibre and micronutrient adequacy, and (e) the 40-year cumulative trajectory of LDL exposure. A scoring system that ranks meals on cancer-hazard classifications rather than on those five things will rank them wrongly.

E. The scoring rubric

score_adjustment is an integer −10 to +10 answering: how much does the presence of a normal serving of this food in a meal move the meal’s health score, for this subject?

Range	Meaning
+8 to +10	Strong converging evidence of substantial benefit
+4 to +7	Solid benefit
+1 to +3	Mildly positive, or better than the thing it displaces
0	Neutral
−1 to −3	Mildly negative in context
−4 to −7	Consistent harm signal
−8 to −10	Strong harm signal

Five rules for applying these scores correctly

These are modifiers, not the whole model. They sit on top of a macro/energy assessment. Total calories, protein grams, fibre grams, and sodium milligrams should carry more weight in the final meal score than the sum of the ingredient adjustments. A meal of “good” ingredients at 1,100 kcal and 1,900 mg sodium is worse for this subject than a plain chicken-rice-broccoli tray.
Quantity is not encoded. An ingredient list tells you a food is present, not how much. A meal listing “olive oil” as the 12th ingredient and one that is 30% olive oil by weight score identically. Treat the scores as presence signals, and expect the model to be systematically imprecise on any calorie-dense item. Where possible, use ingredient-list ORDER as a weak proxy for quantity (regulatory ordering is by descending weight).
Do not double-count. Sodium appears as a property of soy sauce, cured meat, cheese, canned beans, olives and pickles. Cooking method appears as acrylamide, AGEs, HCAs, and added fat. Pick one place to count each thing. The recommendation in this document is: count sodium once as a nutrient total, count cooking method once as a preparation term, and let the ingredient scores handle food identity only.
Displacement is the real question. In a fixed-calorie meal, adding one food removes another. Coconut oil scores −4 not because it is toxic but because it occupies fat calories that olive oil could have occupied. This is why almost nothing scores below −5: at one meal’s scale, the counterfactual is another ordinary food, not a poison.
Do not let phytochemical stories generate large between-food differentials. The evidence that vegetables beat no vegetables is much stronger than the evidence that kale beats zucchini. Compressed between-vegetable scoring is more defensible than spread-out scoring.

F. Detection-string engineering — the practical failure modes

The table columns include detection_strings because the bottleneck in applying this to 800 ingredient lists is string matching, not evidence. The recurring problems:

Substring collisions are the dominant error source. The worst offenders documented in this file: "butter" (matches peanut butter, almond butter, butternut squash, buttermilk, butter lettuce), "pepper" (bell pepper vs black pepper vs pepperoni vs pepperoncini), "ham" (matches hamburger), "pea" (matches peanut, peach, pear, pea protein), "corn" (matches corn syrup, cornstarch, popcorn), "olive" (matches olive oil), "rice" (matches rice vinegar, rice flour), "celery" (matches celery powder, which is a meat-curing nitrate source, not a vegetable).
Match longest/most-specific first. "olive pomace oil" before "extra virgin olive oil" before "olive oil" before "olive". A greedy shortest-match pass will systematically mis-score the fats.
Word boundaries are mandatory for short strings. Regex \b on both sides, not naive in.
Absence of a string is not absence of the thing. Ingredient lists rarely say “canned,” rarely specify oil type when a blend is used, and rarely disclose cooking method. Model absence as unknown, not as zero.
Ingredient-list order carries information. US and EU labelling both require descending order by weight. The first 3–5 ingredients typically account for the large majority of the mass. Weighting by position is a cheap and substantial accuracy improvement over binary presence.
Compound dishes hide their components. “Chicken tikka masala” may not enumerate cream, ghee, or sugar. Cuisine-level priors may be needed as a fallback.

G. What this document deliberately does NOT cover

Additives (emulsifiers, sweeteners, preservatives, colours, gums) — out of scope by instruction; handled elsewhere.
Ultra-processing as a construct (NOVA classification, UPF epidemiology) — handled elsewhere.
Portion size, total energy, and macronutrient totals — these dominate the outcome for this subject and must be handled by the meal-level model, not by this ingredient table.

H. Verification standard applied

Every citation in this document was either (a) verified in this session by fetching the source, the PubMed record, or the abstract via the NCBI E-utilities API, or (b) explicitly marked as unverified recall. No citation, sample size, effect estimate, or DOI has been invented. Where a specific number could not be verified, the number is flagged and the qualitative claim is stated separately from it so the reader can discount the two independently.

Sections written under a partially exhausted web-search budget relied more heavily on the PubMed API; this affected breadth of discovery, not the reliability of what is reported.

Section 01 — PROTEINS

Subject the scores are calibrated for: a healthy general-population adult. [personal profile and health goals removed]

Purpose: evidence base for programmatic scoring of ~800 meal-delivery ingredient lists.

Verification standard used here: every citation below was pulled in-session from a primary index (Europe PMC REST API, PubMed/PMC full text, USDA FoodData Central API, USDA/FDA published tables, SEER) unless explicitly flagged. Anything I could not verify is marked [UNVERIFIED RECALL] and should not be used as a load-bearing citation.

0. Cross-cutting epistemics (applies to every subsection)

Before any food-specific claim, four structural facts about this literature that determine how much weight each number can carry:

Almost all diet–disease evidence here is prospective cohort, not RCT. There is essentially no long-term hard-endpoint RCT for any whole food in this section. The few RCTs are surrogate-endpoint (LDL, apoB, TMAO, body composition) or supplement trials (omega-3 capsules), which are not the same exposure as the food.
The noise floor. In nutritional epidemiology, RRs in the ~1.10–1.30 band are within the plausible magnitude of residual confounding and measurement error. Food-frequency questionnaires misclassify intake substantially; adjustment models are analyst-chosen; and healthy-user confounding is severe for foods with strong cultural valence (bacon, red meat, eggs). A 1.15 HR from an FFQ cohort is not the same grade of evidence as a 1.15 HR from a randomized trial. Two named critiques of the field, both now verified: Ioannidis JP, “Implausible results in human nutrition research,” BMJ 2013;347:f6698; and Schoenfeld JD & Ioannidis JP, “Is everything we eat associated with cancer? A systematic cookbook review,” AJCN 2013;97(1):127–134. The latter took 50 common cookbook ingredients and searched for cancer associations: 40 of 50 (80%) had published cancer-risk articles; of 264 single-study assessments, 191 (72%) concluded the food was associated with increased (n=103) or decreased (n=88) risk, and 75% of those risk estimates had weak (0.05 > P ≥ 0.001) or no statistical significance. Median RR was 2.20 (IQR 1.60–3.44) for studies concluding increased risk and 0.52 (0.39–0.66) for decreased — yet the corresponding meta-analytic RRs were on average null (median 0.96, IQR 0.85–1.10). Statistically significant results were far more likely to be promoted into the abstract (P < 0.0001). This is the single best empirical demonstration that single-study nutrition effect sizes are systematically inflated and shrink toward null when pooled — and it is the reason every number in this document is reported with its design label and, wherever possible, from a meta-analysis rather than a single cohort.
Healthy-user confounding runs in a specific direction in US cohorts. In the US, “eats bacon and eggs” clusters with smoking, low physical activity, low fiber, higher BMI, and lower education. In Asian and some European cohorts, that cluster dissolves — and, strikingly, so do several of the associations (see eggs/T2D §3.4 and eggs/CVD §3.3). Geographic heterogeneity of the same exposure–outcome pair is a strong tell for confounding rather than causation.
Absolute risk is what matters for scoring. US lifetime risk of colorectal cancer is 3.9% (SEER, 2021–2023 data; ~1 in 26). [Correction 2026-09-05: baseline ~4% per ACS lifetime risk; ~4% → ~4.7% at the 18%-per-50 g figure.] US lifetime risk of diagnosed diabetes for men is 40.2% (95% CI 39.2–41.3) (Gregg EW et al., Lancet Diabetes Endocrinol 2014, based on 2000–2011 NHIS data). Those two baselines drive almost every absolute-risk calculation below. Note that this subject — lean, active, normoglycemic in early adulthood — sits far below the population baseline for the diabetes number, so the absolute effect of any HR applied to the adult is proportionally smaller.
A note on applying RRs to lifetime cumulative risk. Multiplying a lifetime risk by a hazard ratio is an approximation that ignores competing risks and assumes lifelong exposure at the stated dose. I do it below because it is the most interpretable framing, but every such number should be read as “order of magnitude,” not a point estimate.

1. PROCESSED MEAT

(bacon, sausage, deli/lunch meat, ham, andouille, chorizo, pepperoni, prosciutto, salami, hot dogs, pancetta, capicola, kielbasa, mortadella, corned beef, pastrami, Spam)

1.1 What IARC Group 1 actually means

IARC classified processed meat as Group 1, carcinogenic to humans in October 2015 (Bouvard V, Loomis D, Guyton KZ, et al., Lancet Oncology 2015 — IARC Monographs Volume 114 Working Group). The WHO Q&A accompanying the monograph is explicit about two things that are routinely garbled in public discussion:

Group 1 is a statement about strength of evidence that the agent causes cancer, not about magnitude of risk. WHO: Group 1 means “there is sufficient evidence of carcinogenicity in humans… convincing evidence that the agent causes cancer.” Tobacco smoking, asbestos, and processed meat are all Group 1. That does not make a hot dog equivalent to a cigarette. WHO’s own comparison, from the Global Burden of Disease Project: ~34,000 cancer deaths/year worldwide attributable to diets high in processed meat, versus ~1,000,000/year from tobacco, 600,000/year from alcohol, and >200,000/year from air pollution. Processed meat is a real but ~30-fold smaller population signal than tobacco.
The dose-response figure: WHO states that “every 50 gram portion of processed meat eaten daily increases the risk of colorectal cancer by about 18%.”

1.2 Translating 18% into absolute risk — the arithmetic

Baseline: US lifetime CRC risk 3.9% (SEER 2021–2023). [Correction 2026-09-05: baseline ~4% per ACS lifetime risk; ~4% → ~4.7% at the 18%-per-50 g figure.]

Daily dose	RR (linear extrapolation from 18%/50 g)	Lifetime CRC risk	Absolute change	Per 1,000 people	Approx NNH
50 g/day (≈2 hot dogs or 4–6 slices cooked bacon)	1.18	3.9% × 1.18 = 4.60%	+0.70 pp	+7	~143
25 g/day	1.09	4.25%	+0.35 pp	+3.5	~286
~16 g/day (2 slices cooked bacon, every single day, for life)	~1.058	4.13%	+0.23 pp	+2.3	~435

[Correction 2026-09-05: baseline ~4% per ACS lifetime risk; ~4% → ~4.7% at the 18%-per-50 g figure.]

So: a person who eats two slices of bacon every day for their entire life moves their lifetime colorectal cancer risk from about 3.9% to about 4.1%. That is a real, non-zero, and defensible reason to prefer other proteins. It is not a reason to treat a single bacon-containing meal as a health emergency, and any scoring system that assigns bacon a catastrophic penalty on cancer grounds alone is misrepresenting the magnitude.

The cardiometabolic numbers are bigger than the cancer numbers, and this is under-appreciated.

Coronary heart disease. Micha R, Wallace SK, Mozaffarian D, Circulation 2010 — systematic review + meta-analysis, 20 studies (17 prospective cohorts, 3 case-control), 1,218,380 individuals, 23,889 CHD / 2,280 stroke / 10,797 diabetes cases. Processed meat: RR 1.42 per 50 g/day for CHD (95% CI 1.07–1.89) and RR 1.19 per 50 g/day for diabetes (1.11–1.27). Unprocessed red meat: RR 1.00 per 100 g/day for CHD (0.81–1.23) and 1.16 for diabetes (0.92–1.46), both null. This is the single cleanest demonstration that “processed” and “red” are different exposures with different risk profiles.
Type 2 diabetes, the best-powered modern estimate. Li C, Bishop TRP, Imamura F, et al. (InterConnect consortium), Lancet Diabetes Endocrinol 2024;12(9):619–630 — individual-participant federated meta-analysis, 1,966,444 adults, 107,271 incident T2D cases, 31 cohorts, 20 countries, median 10 y follow-up. Processed meat HR 1.15 per 50 g/day (1.11–1.20); unprocessed red meat HR 1.10 per 100 g/day (1.06–1.15); poultry HR 1.08 per 100 g/day (1.02–1.14). The federated design (analysis code shipped to each cohort, harmonised protocol) removes a lot of the publication-bias and analyst-choice objections that dog conventional meta-analysis. This is the strongest observational evidence in the whole section.
- Absolute: applying HR 1.15 to the US male lifetime diabetes baseline of 40.2% gives ~46.2%, i.e. +6 percentage points for lifelong 50 g/day. [Correction 2026-09-05: cumulative-risk arithmetic gives about +4 points (1−(1−0.40)^1.15), not +6; the +6 comes from naive multiplication.] That is roughly ten times the absolute magnitude of the colorectal cancer effect. For this subject, whose actual baseline is far below 40%, scale down proportionally — but the ordering (metabolic > cancer) holds.
US cohort with absolute risk differences. Zhong VW, Van Horn L, Greenland P, et al., JAMA Internal Medicine 2020 — pooled US cohorts, 29,682 adults, median 19.0 y, 6,963 CVD events, 8,875 deaths. Per 2 servings/week: processed meat HR 1.07 (1.04–1.11), ARD 1.74% for incident CVD; unprocessed red meat HR 1.03 (1.01–1.06), ARD 0.62%; poultry HR 1.04 (1.01–1.06), ARD 1.03%; fish HR 1.00 (0.98–1.02), ARD 0.12%. All-cause mortality: processed meat HR 1.03 (1.02–1.05); poultry HR 0.99 (0.97–1.02); fish HR 0.99 (0.97–1.01).
Burden of Proof re-analysis. IHME’s Burden of Proof study on processed meat, SSBs and trans fats (Nature Medicine, published 30 June 2025) reports a conservative “burden of proof risk function” of at least ~11% average increase in T2D risk and ~7% increase in CRC risk across the observed exposure range — i.e., when you force the analysis to report the most conservative estimate consistent with the data across all studies, the signal shrinks but does not vanish. I verified the publication, journal and date and the two headline percentages via IHME/press summaries; I could not retrieve the paper’s own star ratings and exposure ranges directly (Nature paywall/redirect) — treat the star-rating detail as [UNVERIFIED RECALL]. Note also that the Burden of Proof method itself has published critics (Nature Medicine 2023, “Concerns about the Burden of Proof studies,” with an authors’ reply) — this is a contested methodology, not a settled arbiter.

1.3 Mechanism: what is actually doing the damage?

Four candidate mechanisms, in descending order of evidential support:

(a) Endogenous N-nitroso compound (NOC) formation, driven by haem iron — strongest mechanistic evidence. This is the one mechanism with human interventional support. Controlled human feeding studies (Bingham and colleagues, Cambridge) showed volunteers on a high red-meat diet (420 g/day) excreted significantly more faecal N-nitroso compounds than on a low-meat diet (60 g/day), and that an isonitrogenous vegetable-protein diet did not produce the increase. Follow-up work established that haem, not inorganic iron and not meat protein, drives the nitrosation (Cross AJ et al., Cancer Research 2003, “Haem, not protein or inorganic iron, is responsible for endogenous intestinal N-nitrosation arising from red meat”). Downstream, O⁶-carboxymethylguanine adducts have been detected in exfoliated colonocytes from volunteers eating red meat, correlating with faecal apparent-total-NOC. Design labels: human controlled feeding (n small, tens), plus mechanistic/in-vitro and animal models. It is a coherent chain from exposure → biomarker → promutagenic DNA lesion, but no human trial has ever run it to a cancer endpoint.

Critically, this mechanism is not specific to added nitrite. Haem itself catalyses nitrosation from nitrogen oxides in the gut. That is why unprocessed red meat also raises faecal NOC, and it is why the “if I just remove the added nitrite it becomes safe” argument is weaker than it sounds.

(b) Added nitrite / nitrate — supported, but the human evidence is thinner than the confidence with which it is asserted. The best direct human data are from NutriNet-Santé (France):

Chazelas E, Pierre F, et al., International Journal of Epidemiology 2022 — prospective cohort, 101,056 adults, median 6.7 y. Additives-origin nitrites were associated with prostate cancer risk (sodium nitrite E250: HR 1.58, 1.14–2.18, P = 0.008); “no association was observed for natural sources” (vegetables, water).
Srour B, Chazelas E, et al., PLoS Medicine 2023 — prospective cohort, 104,168 adults, median 7.3 y. Additives-origin nitrites vs non-consumers: HR 1.53 (1.24–1.88) for T2D; sodium nitrite specifically 1.54 (1.26–1.90); natural-source nitrate tertile 3 vs 1: 1.26 (1.03–1.54). These are single-cohort, self-reported-additive-exposure estimates in the range where residual confounding is a live explanation (additive-nitrite intake is a near-perfect proxy for “eats a lot of charcuterie”). They are suggestive, not decisive.
IARC separately evaluated ingested nitrate and nitrite “under conditions that result in endogenous nitrosation” as Group 2A (probably carcinogenic) in Monographs Volume 94. Note the conditional clause — it is not a blanket statement about nitrate.

(c) Sodium — underrated, and the most directly relevant mechanism for THIS subject. See §1.5.

(d) Confounding — the honest residual. The InterConnect and Micha estimates are adjusted for BMI, smoking, activity, energy, and other diet, but processed meat intake in Western cohorts is a marker for a whole dietary and behavioural pattern. It is not possible to fully separate “bacon” from “the life in which bacon is eaten daily” with FFQ data. The fact that the processed vs unprocessed contrast is so sharp (Micha: 1.42 vs 1.00 for CHD) is modest evidence against pure confounding, since both track the same lifestyle cluster — but processed meat also tracks convenience-food patterns more tightly, so this is not airtight.

Also relevant but usually lumped in wrongly: high-temperature cooking products (heterocyclic aromatic amines, polycyclic aromatic hydrocarbons) are a cooking-method exposure, not a processing exposure. They apply equally to charred chicken and grilled steak. Do not attribute HCA/PAH risk to the processed-meat class specifically.

1.4 “Uncured,” “no nitrate or nitrite added,” celery-powder-cured — does it change anything?

Short answer: no, and the label is arguably actively misleading. Celery powder is a nitrate source; the resulting cured-meat chemistry is essentially the same reaction.

The precise facts, from the American Meat Science Association fact sheet Alternative Curing (author: Amanda Gipe McKeith, PhD, Western Kentucky University; reviewers Joe Sebranek, Iowa State, and Jeff Sindelar; dated 6/2014; PIG 12-05-10), retrieved and text-extracted in-session:

Conventional curing adds sodium nitrite directly. USDA limits: up to 200 ppm in-going nitrite for pumped/massaged cured meats (except bacon), 156 ppm for comminuted, 120 ppm minimum for refrigerated products, 120 ppm in-going nitrite plus 550 ppm erythorbate for pumped/massaged bacon, and 625 ppm for dry-cured products.
“Natural”/”uncured” products use vegetable nitrate — usually celery juice powder — plus a nitrate-reducing starter culture and an incubation step to convert nitrate → nitrite. “Vegetable juice powders may contain as much as 2.5% nitrite or more than 25,000 ppm” and “Commercial celery juice powder has approximately 27,462 ppm nitrate (~2.75%)” (citing Sindelar et al. 2007). Some vegetable juice powders “can exceed 4.0% (40,000 ppm) nitrate and 2.5% (25,000 ppm) nitrite.”
The fact sheet also gives raw vegetable juice nitrate for context (Sebranek 2006): carrot 117 ppm, celery 2,114 ppm, beet 2,273 ppm, spinach 3,227 ppm nitrate.
On outcomes: “meat products containing natural nitrate/nitrite and a natural antimicrobial can have similar residual nitrite levels, as compared with conventionally-cured products” (Jackson et al. 2011). Residual nitrate is generally higher in vegetable-powder products because the powder starts with more nitrate.
USDA requires these products to be labelled “uncured” and “no nitrate or nitrite added” — a labelling artefact of the definition of “cure” as directly added nitrite, not a statement about chemistry.

Independent retail survey confirming equivalence: Nuñez De González MT, Osburn WN, Hardin MD, Longnecker M, Garg HK, Bryan NS, Keeton JT, Journal of Agricultural and Food Chemistry 2012;60(15):3981–3990, DOI 10.1021/jf204611k — survey of 470 retail cured-meat products across six categories in five US metro areas, HPLC. Result: nitrite concentrations were comparable between conventional and organic/natural/uncured/indirectly-cured products; some ONC products showed lower nitrate in some cities. Design: cross-sectional retail survey.

Secondary sources (Consumer Reports / CSPI 2019 FSIS labelling petition, and a Journal of Food Protection analysis reported alongside it) state that “natural” hot dogs ranged from half to ten times the nitrite of conventional hot dogs, and natural bacon from about a third to more than twice conventional. I verified that the CSPI/Consumer Reports petition to FSIS exists (submitted August 2019) but could NOT verify the first author, volume, or page of the underlying Journal of Food Protection paper — treat those specific fold-ranges as [UNVERIFIED RECALL].

One genuine chemical nuance, stated honestly: vegetable-derived cures are typically accompanied by ascorbate/erythorbate-equivalent reducing agents (cherry powder, ascorbic acid), which inhibit nitrosamine formation. But conventional bacon is legally required to carry 550 ppm erythorbate for exactly this reason. So the antioxidant advantage is not unique to “uncured” products either.

Scoring implication: “uncured,” “nitrate-free,” and “celery-powder-cured” processed meats get the SAME penalty as conventional. The detection layer should treat “uncured bacon,” “no-nitrate bacon,” and “nitrate-free turkey” as processed meat, not as a mitigated class. The only thing that materially varies between brands is sodium, and that is on the label.

1.5 The sodium load — the decisive consideration for this subject

This subject is actively reducing sodium. Processed meat is, by weight, one of the densest sodium sources in a normal diet, and this is a deterministic, measurable property — not an epidemiological inference.

From Micha 2010 (Circulation) Table 2, extracted from full text in-session: processed meats averaged 622 mg sodium per 50 g serving vs 155 mg per 50 g for unprocessed red meats — roughly 4-fold. The same table reports processed meats carry approximately 50% higher non-salt preservatives (nitrates/nitrites/nitrosamines). (These figures were read out of the paper’s Table 2 by full-text extraction; the Discussion refers to them qualitatively as “substantially higher sodium and nitrate preservative levels.”)

USDA FoodData Central values retrieved in-session, mg sodium per 100 g:

Item (USDA description)	Na (mg/100 g)	Protein (g/100 g)
Pepperoni, beef and pork, sliced	1,580	19.2
Salami, cooked, beef	1,140	12.6
Salami, cooked, turkey	1,110	19.2
Chicken breast, oven-roasted, fat-free, sliced (deli)	1,090	16.8
Ham, turkey, sliced, extra lean, prepackaged/deli	1,040	19.6
Pork bacon, cooked — reduced sodium variety	1,030	37.0
Chorizo, pork, cooked, pan-fried	983	19.3
Canadian bacon, pan-fried	993	28.3
Roast beef, deli style, sliced	853	18.6
Frankfurter, beef, heated	852	11.7
Chicken breast roll, oven-roasted (deli)	883	14.6
Beef, ground, 90/10, raw (for contrast)	66	20.0

USDA SR Legacy sodium-by-household-measure table (nal.usda.gov), retrieved in-session:

Pork frankfurter, 1 link: 620 mg
Cured ham, shank, lean only, 3 oz: 719 mg
Italian salami, pork, 1 oz: 529 mg
Bologna, meat and poultry, 1 slice: 455 mg
Whole egg, raw, 1 large: 71 mg
Ground beef 90/10 patty, cooked, 3 oz: 58 mg

Arithmetic for this subject. US Dietary Guidelines limit: 2,300 mg/day. AHA “ideal”: 1,500 mg/day. A single 50 g portion of processed meat at ~620 mg is 27% of the DGA limit and 41% of the AHA ideal, from one component of one meal. Two eggs plus 4 oz of chicken breast delivers roughly the same protein for under 200 mg. For someone deliberately cutting sodium, this is the largest single, most controllable, most certain effect of processed meat on the adult’s stated goals — larger and far better evidenced than the cancer question.

Effect size worth having in mind: Sacks FM, Svetkey LP, Vollmer WM, et al., DASH-Sodium, NEJM 2001 — randomized feeding trial, n = 412, three sodium levels × two diet patterns, 30 days each. Going from intermediate to low sodium lowered systolic BP by 4.6 mm Hg on the control diet and 1.7 mm Hg on the DASH diet; low-sodium DASH vs high-sodium control was 7.1 mm Hg lower systolic in normotensives and 11.5 mm Hg in hypertensives. Design: RCT (feeding trial) — the strongest design in this entire section, and it happens to be the mechanism most relevant to the subject’s stated goal.

1.6 Processed poultry as a sub-class

Turkey bacon, chicken sausage, deli turkey, turkey pepperoni: lower saturated fat and lower haem iron than pork/beef versions, but the same or higher sodium (USDA: turkey salami 1,110 mg/100 g; extra-lean deli turkey ham 1,040 mg/100 g; fat-free oven-roasted deli chicken breast 1,090 mg/100 g) and the same curing chemistry. IARC’s processed-meat definition is about the processing (salting, curing, fermentation, smoking), not the species. Give a somewhat smaller penalty than red processed meat (less haem, less saturated fat, better protein-per-calorie) but do not treat it as clean.

1.7 Verdict and score

score_adjustment = −6, tier = harmful, confidence = high.

Reasoning: four independent evidence streams converge (IARC Group 1 for CRC; Micha CHD RR 1.42/50 g; InterConnect T2D HR 1.15/50 g in ~2M people; deterministic sodium arithmetic that directly opposes a stated subject goal). The absolute cancer effect alone would justify only about −2 or −3. The sodium load and the T2D signal push it to −6. It is not −9: the absolute lifetime cancer increment from realistic delivery-meal doses is ~2 per 1,000, and a scoring system that treats bacon as poison loses credibility and utility.

Sub-class: uncured/celery-cured processed meat = −6 (identical). Processed poultry = −4.

2. UNPROCESSED RED MEAT

(beef, steak, ground beef, pork loin/chop/tenderloin, lamb, bison, venison, veal)

This is the genuinely contested one, and the honest answer is closer to “small and uncertain” than either camp’s public position.

2.1 The two positions

WCRF/AICR (2018 Continuous Update Project / Third Expert Report): evidence that red meat increases colorectal cancer risk is graded “probable” (processed meat: “convincing”). Recommendation: limit red meat to moderate amounts — roughly <500 g cooked per week (~350–500 g) — and eat little, if any, processed meat. IARC placed unprocessed red meat in Group 2A, “probably carcinogenic,” citing limited evidence in humans plus strong mechanistic evidence; WHO’s stated figure is “risk of colorectal cancer could increase by 17% for every 100 gram portion of red meat eaten daily.”

Absolute translation: 3.9% × 1.17 = 4.56%, i.e. +0.66 percentage points, ~7 per 1,000 lifetime, for 100 g/day every day for life. [Correction 2026-09-05: baseline ~4% per ACS lifetime risk; ~4% → ~4.7% at the 18%-per-50 g processed-meat figure. Note this row is unprocessed red meat at 17% per 100 g, a different exposure.]

NutriRECS (2019): Johnston BC, Zeraatkar D, Han MA, Vernooij RWM, Valli C, El Dib R, Marshall C, Stover PJ, Fairweather-Taitt S, Wójcik G, Bhatia F, de Souza R, Brotons C, Meerpohl JJ, Patel CJ, Djulbegovic B, Alonso-Coello P, Bala MM, Guyatt GH. Unprocessed Red Meat and Processed Meat Consumption: Dietary Guideline Recommendations From the Nutritional Recommendations (NutriRECS) Consortium. Annals of Internal Medicine 2019;171(10):756–764. A 14-member panel from seven countries (including three community representatives) issued weak recommendations, based on low-certainty evidence, that adults continue their current consumption of both unprocessed and processed red meat. Environmental and animal-welfare considerations were explicitly excluded.

Its four supporting systematic reviews (all Ann Intern Med 2019;171(10)):

Zeraatkar D, Han MA, Guyatt GH, et al., pp. 703–710 — 55 cohorts, >4 million participants (61 articles): all-cause mortality and cardiometabolic outcomes. Low-certainty evidence that reducing unprocessed red meat by 3 servings/week produces very small reductions in cardiovascular mortality, stroke, MI and T2D. “Associations are very small, and the evidence is of low certainty.”
Han MA, Zeraatkar D, Guyatt GH, et al., pp. 711–720 — 56 cohorts, >6 million participants (118 articles): cancer mortality and incidence. “The possible absolute effects of red and processed meat consumption on cancer mortality and incidence are very small, and the certainty of evidence is low to very low.”
Vernooij RWM, Zeraatkar D, et al., pp. 732–741 (DOI 10.7326/M19-1583) — dietary patterns.
Plus a values-and-preferences review (Valli C et al.).

2.2 The methods argument — and it is a real argument, not just a food fight

NutriRECS applied GRADE, which starts observational evidence at “low certainty” and requires large effects, dose–response, or plausible-confounding arguments to upgrade. GRADE was built for clinical interventions. The nutrition-epidemiology objection (Harvard Chan’s Nutrition Source and many others) is roughly:

GRADE’s automatic downgrade of observational evidence is inappropriate when RCTs are impossible in principle (you cannot randomize 100,000 people to eat steak for 30 years). Under a strict GRADE reading, the evidence that smoking causes lung cancer would also be “low certainty.”
NutriRECS translated small relative risks into “3 servings per week” absolute changes over a lifetime and then judged them trivial — but the same framing applied to any single dietary factor makes every dietary factor look trivial, while collectively they are not.
The panel’s own vote was close: the recommendations passed on a bare majority, and several panellists dissented publicly.

The counter-argument, which is also real:

GRADE was applied transparently and consistently; the critics’ objection amounts to wanting a lower evidentiary bar for a conclusion they already hold.
NutriRECS did not claim red meat is harmless. It claimed the certainty is low and the magnitude is small — both of which the underlying meta-analyses support and which the Burden of Proof re-analysis (§1.2) largely corroborates.
Guideline bodies had been issuing “strong” recommendations off “low-certainty” evidence, which is a genuine methodological violation under GRADE.

My read: the substantive claim (small effect, low certainty for unprocessed red meat) is well supported. The communication was reckless — “continue current consumption” is a recommendation, and framing a null-certainty finding as a positive recommendation was a category error the authors were warned about.

2.3 The conflict-of-interest dispute — what is documented and what is not

Bradley Johnston (first author) was, while the papers were being drafted, negotiating a tenured position at Texas A&M AgriLife, an institution that runs meat research/marketing/promotion programs; the adult accepted the position on 1 August 2019; the adult’s Annals COI form was dated 23 August 2019 and did not disclose it.
Patrick Stover (co-author) heads Texas A&M AgriLife; the adult’s COI form dated 10 September 2019 did not disclose that role.
Johnston had previously led a similarly-structured GRADE review concluding that evidence for reducing sugar intake was low-quality, funded by the International Life Sciences Institute (ILSI), an industry group; this prior funding was reportedly not disclosed on the meat papers either.
Annals of Internal Medicine published a correction: “Correction: Nutritional Recommendations (NutriRECS) on Consumption of Red and Processed Meat,” Ann Intern Med 2020;172(3):228, DOI 10.7326/L19-0822. I confirmed this correction exists, its journal, volume, issue, page and year, but could NOT retrieve its full text this session — the exact wording of what was corrected is [UNVERIFIED — retrieved metadata only].
An accompanying editorial: Carroll AE, Doherty TS, “Meat Consumption and Health: Food for Thought,” Ann Intern Med 2019, DOI 10.7326/M19-2620.

How much should the COI change your posterior? Some, but less than the coverage implied. The underlying meta-analyses are reproducible, use standard methods, and their point estimates broadly agree with WCRF’s — the two camps mostly disagree about how to characterise certainty and how to translate small effects into advice, not about the numbers. An undisclosed COI is a serious process failure and a reason for heightened scrutiny; it is not itself evidence that the estimates are wrong.

2.4 Mendelian randomization — the newest and most under-cited evidence

MR uses genetic variants as instruments for lifelong exposure, which sidesteps most confounding and reverse causation. For red meat this is imperfect (genetic instruments for food preference are weak and pleiotropic — they may proxy for taste, socioeconomic status, or general appetite rather than meat per se), but it is a genuinely independent line of evidence.

Hoang T et al., BMC Cancer 2024, DOI 10.1186/s12885-024-12923-1 — one-sample MR in UK Biobank; 374,001 participants after exclusions, GWAS on 408,093; 4,686 CRC cases (3,131 colon, 1,555 rectal). Genetically-predicted red meat: HR 0.72 (0.40–1.28), null. Processed meat: HR 0.57 (0.29–1.11), null overall; a nominally inverse association for rectal cancer (0.29, 0.09–0.93) which the authors themselves caution against over-reading given small case numbers. Critically: the same dataset’s conventional observational analysis DID show positive associations, which disappeared under MR.
Two-sample MR studies of red/processed meat and cardiovascular disease using UK Biobank exposure GWAS and CAD outcome consortia have likewise reported no causal association.

Interpretation, stated carefully: MR null results here are suggestive but not decisive. The instruments are weak (dietary-intake GWAS explain a tiny fraction of variance), power is limited (4,686 CRC cases is not a lot for MR), and the confidence intervals are wide enough to contain the observational point estimates. “MR found no causal effect” and “MR could not detect an effect of the size observational studies report” are both true statements about these data. But the fact that the observational association vanished within the same cohort when the instrument changed is a meaningful red flag about confounding.

2.5 TMAO — a mechanism that is real as a biomarker response and unproven as a cause

Feeding-trial evidence that red meat raises TMAO: yes. Wang Z, Hazen SL, et al., European Heart Journal 2019 (DOI 10.1093/eurheartj/ehy799) — randomized crossover controlled feeding, 113 volunteers, red vs white vs non-meat protein at 25% of calories, in a 2-arm (high/low saturated fat) design. Chronic red meat — but not white meat or non-meat — increased plasma and urine TMAO more than two-fold (P < 0.0001); press reporting describes ~3-fold average with >10-fold in some individuals. It also reduced fractional renal TMAO excretion. Design: RCT (crossover feeding).
But the RCT record overall is inconsistent. Jafari F et al., Advances in Nutrition 2025 — systematic review of 13 unique RCTs across 15 publications, 553 participants, median 28 days. Of the comparisons: 6 showed higher TMAO with more red meat (71–420 g/day), 7 showed no significant difference, and 2 showed LOWER TMAO with red meat than with seafood-rich diets. The authors note many participants had baseline TMAO <6.2 μM, below the concentration considered risk-enhancing, and that “causality cannot be inferred.”
Is TMAO causal for CVD? Probably not straightforwardly. Bidirectional Mendelian randomization analyses have found genetically-predicted TMAO not associated with T2D or other cardiometabolic disease, with evidence instead consistent with reverse causation — kidney dysfunction and diabetes raise TMAO rather than the other way round (reviewed in Nutrition & Diabetes 2025, “TMAO and diabetes: from the gut feeling to the heart of the problem”). TMAO is heavily renally cleared, so it rises with any decline in kidney function; a large fraction of the observational TMAO–CVD association may be renal function in disguise. Also note fish — universally recommended — is the single largest dietary source of preformed TMAO, which is awkward for the “TMAO is the mechanism” story.
Verdict: TMAO is a well-documented biomarker response to red meat and a poorly-supported causal mediator. Do not score red meat on TMAO grounds.

2.6 What red meat gives this subject

Positives that a purely risk-focused analysis misses: highest-quality complete protein with a strong leucine content; haem iron (~2.2 mg/100 g in 90/10 ground beef, absorbed several-fold better than non-haem); B12; zinc; creatine and carnosine (relevant for a power sport like recreational sport); very high satiety per calorie for lean cuts; ~66 mg sodium/100 g (essentially nothing). For a adult in a fat-loss phase trying to preserve lean mass, lean red meat is an efficient tool.

Negative that is RCT-grade and often ignored by red-meat defenders: Bergeron N, Chiu S, Williams PT, King SM, Krauss RM, American Journal of Clinical Nutrition 2019 (APPROACH trial) — randomized crossover controlled feeding, 113 participants (61 high-SFA arm, 52 low-SFA arm completed all three diet periods). “LDL cholesterol and apoB were higher with red and white meat than with nonmeat, independent of SFA content (P < 0.0001 for all, except apoB: red meat vs nonmeat P = 0.0004).” Red and white meat did not differ from each other. This is a genuinely important and widely-misreported result: swapping beef for chicken does not improve atherogenic lipoproteins; swapping either for plant protein does.

2.7 Verdict and scores

Lean unprocessed red meat (sirloin, pork loin, 90/10+ ground beef, tenderloin, bison, venison): score_adjustment = 0, tier = contested, confidence = moderate. It is a nutrient-dense, low-sodium, high-satiety protein that supports this subject’s lean-mass goal. The harm signals are small (CRC +0.7 pp lifetime at 100 g/day; T2D HR 1.10/100 g), rated low-certainty by GRADE, and not reproduced by MR. But it does raise apoB relative to plant protein in RCT, and WCRF’s “probable” grading is not nothing. Net: neutral in a mixed diet, mildly negative at high frequency. Score 0 with the frequency caveat handled at the meal-plan level, not the meal level.

Fatty unprocessed red meat (ribeye, short rib, 80/20 ground beef, pork belly, lamb shoulder, brisket): score_adjustment = −2, tier = contested, confidence = moderate. Same evidence, plus meaningfully higher saturated fat and energy density, which matters for a [calorie target removed] fat-loss budget.

3. POULTRY AND EGGS

3.1 Poultry — the evidence

Chicken and turkey are the closest thing to a null food in this literature, which for scoring purposes means “better than what it displaces.”

Zhong VW et al., JAMA Internal Medicine 2020 (design/n above): poultry, per 2 servings/week, incident CVD HR 1.04 (1.01–1.06), ARD 1.03%; all-cause mortality HR 0.99 (0.97–1.02), ARD −0.28%. The CVD signal is at the very edge of the noise floor, and the authors themselves noted it was partly attributable to fried chicken in these cohorts.
InterConnect (Li C et al., Lancet Diabetes Endocrinol 2024, n = 1,966,444): poultry HR 1.08 per 100 g/day (1.02–1.14) for T2D — smaller than processed meat (1.15/50 g) and unprocessed red meat (1.10/100 g), but not zero. The authors note the poultry association was less robust in sensitivity analyses.
Bergeron 2019 APPROACH RCT (above): white meat raised LDL and apoB as much as red meat, versus non-meat protein. This is the strongest reason not to score chicken as “healthy” in an absolute sense — only as “healthy relative to red and processed meat.”
A 2025 outlier worth flagging honestly: Bonfiglio C, Tatoli R, Donghia R, Pesole PL, Giannelli G, Nutrients 2025 — prospective cohort, 4,869 participants (MICOL and NUTRIHEP, southern Italy). “For GCs [gastrointestinal cancers], the SHR for weekly poultry consumption above 300 g was 2.27 (1.23–4.17), a risk that for men increased to 2.61 (1.31–5.19)”; >300 g/week also associated with 27% higher all-cause mortality (HR 1.27, 1.00–1.61). This got substantial press. Treat with heavy scepticism: small cohort, small number of events, a biologically implausible effect size (an HR of 2.6 for chicken would be larger than for processed meat, which is absurd on its face), no dose–response mechanism, and it contradicts the ~2 million-person InterConnect and the 29,682-person Zhong analyses. This is a textbook example of a single small cohort producing an implausibly large RR. It should not move the score.

3.2 Poultry verdict

Unbreaded poultry (chicken breast, chicken thigh, turkey, ground chicken/turkey, duck breast): score_adjustment = +3, tier = beneficial, confidence = moderate. Rationale: near-null in cohorts, very high protein-per-calorie (chicken breast ~31 g protein / 165 kcal per 100 g cooked), low sodium in unbrined form, and a strictly better displacement for processed meat. Not +6, because APPROACH shows it is not lipid-neutral versus plant protein and InterConnect shows a small T2D signal. Caveat for the detection layer: many delivery-service chicken products are brined or marinated and carry 300–600 mg sodium per serving — the sodium is real even for “clean” chicken.

Breaded/fried poultry (nuggets, tenders, katsu, popcorn chicken, fried chicken): score_adjustment = −3, tier = harmful, confidence = moderate. Energy density, refined-flour coating, frying oil, and higher sodium. Most of the poultry-CVD signal in US cohorts lives here.

3.3 Eggs and dietary cholesterol — settling it

The case for concern (prospective cohort): Zhong VW, Van Horn L, Cornelis MC, Wilkins JT, Ning H, Carnethon MR, Greenland P, Mentz RJ, Tucker KL, Zhao L, Norwood AF, Lloyd-Jones DM, Allen NB. JAMA 2019;321(11):1081–1095. Individual participant data pooled from 6 prospective US cohorts, n = 29,615, median follow-up 17.5 years, 5,400 CVD events, 6,132 deaths.

Per 300 mg/day dietary cholesterol: incident CVD HR 1.17 (1.09–1.26), adjusted ARD 3.24% (1.39–5.08); all-cause mortality HR 1.18 (1.10–1.26), ARD 4.43% (2.51–6.36).
Per half egg/day: incident CVD HR 1.06 (1.03–1.10), ARD 1.11% (0.32–1.89); mortality HR 1.08 (1.04–1.11), ARD 1.93% (1.10–2.76).
Crucially — and this is the part almost never reported: after further adjustment for dietary cholesterol, the egg associations went to null: CVD HR 0.99 (0.93–1.05), ARD −0.47%; mortality HR 1.03 (0.97–1.09), ARD 0.71%. The authors’ own analysis attributes the egg signal to the cholesterol it carries, not to eggs as a food.

The case against concern (larger prospective cohorts + meta-analysis): Drouin-Chartier JP, Chen S, Li Y, Schwab AL, Stampfer MJ, Sacks FM, Rosner B, Willett WC, Hu FB, Bhupathiraju SN. BMJ 2020;368:m513. Three US cohorts — NHS (83,349 women), NHS II (90,214 women), HPFS (42,055 men) = 215,618 participants, up to 32 years, >5.54 million person-years, 14,806 CVD events — plus an updated meta-analysis of prospective cohorts totalling 1,720,108 participants.

≥1 egg/day vs <1/month: HR 0.93 (0.82–1.05) — no association.
Meta-analysis, per additional egg/day: RR 0.98 (0.93–1.03) for CVD; CHD 0.96 (0.91–1.03); stroke 0.99 (0.91–1.07).
Geographic heterogeneity: Asian cohorts showed an INVERSE association, RR 0.92 (0.85–0.99); US and European cohorts null.

Why do these disagree? Zhong is smaller (29,615 vs 215,618), used mostly single-baseline dietary assessment in several constituent cohorts, and the effect is carried by dietary cholesterol rather than eggs. Drouin-Chartier used repeated FFQs over decades and had far more events. The methodological edge is with Drouin-Chartier. But note both are cohorts and both are subject to the same healthy-user problems in opposite directions — the Harvard cohorts are unusually health-conscious populations, and in that population an egg is not a proxy for a bacon-and-hash-browns breakfast.

The RCT evidence, which is what actually settles the mechanism:

Li MY et al., Nutrients 2020 — meta-analysis of 17 RCTs, ~885 participants, interventions 21–84 days. Egg consumption raised LDL-C by 8.14 mg/dL (95% CI 4.46–11.82) and the LDL-C/HDL-C ratio by 0.14 (0.05–0.22); HDL-C change was not significant (+1.27 mg/dL, −0.28 to 2.83). So dietary cholesterol from eggs does raise LDL, on average, measurably. Design: meta-analysis of RCTs. This is a real effect and the “dietary cholesterol doesn’t matter” wellness line is wrong as stated.
PROSPERITY trial (2024) — RCT, n = 140 adults ≥50 with or at high risk for CVD, ≥12 fortified eggs/week vs <2 eggs/week, 4 months. No adverse lipid effect: at 4 months, HDL-C −0.64 mg/dL and LDL-C −3.14 mg/dL in the fortified-egg group vs the non-egg group. Presented at ACC.24 (April 2024); published in American Heart Journal. Caveats: small, short, used fortified eggs (omega-3 enriched, lower saturated fat), older/higher-risk population, and industry-adjacent.

Hyper-responders. Roughly one third of people show a substantially larger plasma cholesterol rise per unit dietary cholesterol. A common operational cut is >2.2 mg/dL rise per additional 100 mg dietary cholesterol = hyper-responder. Fernandez and colleagues’ work reports that hyper-responders raise both LDL-C and HDL-C, with a shift toward larger LDL particles. Two caveats: (a) these are small mechanistic studies, and (b) Maria Luz Fernandez has an extensive history of egg-industry funding and has had papers retracted — her interpretive framing (“hyper-responders get larger, less atherogenic LDL, so it’s fine”) should be treated as a hypothesis, not a finding. The particle-size-protects argument is also weakened by the fact that apoB particle number tracks risk better than particle size.

Eggs and type 2 diabetes — a clean natural experiment in confounding. Multiple independent meta-analyses find the egg–T2D association exists only in US cohorts:

Drouin-Chartier JP et al., AJCN 2020: 3 US cohorts + meta-analysis of 16 studies (589,559 participants). “Each 1 egg/d was associated with higher T2D risk among US studies (RR 1.18; 1.10–1.27)” but not in European or Asian cohorts.
Wallin A et al., Diabetologia 2016: Swedish cohort (39,610 men) null; meta-analysis of 12 studies: per 3×/week increment, HR 1.18 (1.13–1.24) in five US studies vs 0.97 (0.90–1.05) in seven non-US studies.
Djoussé L et al., AJCN 2016: 12 cohorts, 219,979 subjects, 8,911 cases — 39% higher risk in US studies, RR 0.89 (0.79–1.02) elsewhere.
Tamez M et al., BJN 2016: dose–response meta-analysis, 251,213 individuals, 12,156 cases — “for studies conducted in the USA, we observed a stronger association (RR 1.47; 1.32–1.64), whereas results were null for studies conducted elsewhere.” An egg is chemically identical in Stockholm and in Ohio. A four-fold difference in association by continent is very hard to explain by biology and very easy to explain by what the egg is eaten with and by whom. This is the single strongest piece of evidence in this whole section that the American egg literature is confounded.

Whole eggs vs whites for this subject. van Vliet S, Shy EL, Abou Sawan S, Beals JW, West DWD, Skinner SK, Ulanov AV, Li Z, Paluska SA, Parsons CM, Moore DR, Burd NA. AJCN 2017 — crossover RCT, 10 resistance-trained men, isonitrogenous whole egg vs egg white after resistance exercise: “whole-egg ingestion increased the postexercise myofibrillar protein synthetic response to a greater extent than did the ingestion of egg whites (P = 0.04).” Small n, acute surrogate endpoint (myofibrillar FSR, not hypertrophy), single study — do not over-read it — but combined with the micronutrients in the yolk (choline, lutein/zeaxanthin, vitamin D, B12, selenium), there is no good reason for a healthy adult to prefer whites except for calorie budgeting. Egg white is a clean protein-per-calorie play (~11 g protein, ~52 kcal, ~166 mg sodium per 100 g) and is fine; it is just not superior.

3.4 Eggs verdict

Whole eggs: score_adjustment = +2, tier = contested, confidence = moderate. Rationale: the largest and best-designed cohort evidence (Drouin-Chartier BMJ 2020, 215,618 participants + 1.72M meta) is null; the geographic heterogeneity of the T2D signal is strong evidence of confounding; and the food itself is dense in protein, choline, lutein and vitamin D at 71 mg sodium per egg. Offsetting: the RCT meta-analysis is unambiguous that eggs raise LDL-C by ~8 mg/dL at 1–2/day, which is a real, measurable, non-trivial effect on a causal risk factor, and ~1/3 of people respond more. So not +5. For a lean, active adult with (presumably) good baseline lipids, +2 is the right call, with a note that if the adult’s apoB or LDL-C is elevated, eggs are one of the first things to test-and-reduce.

Egg whites / liquid egg whites: +2, tier = beneficial, confidence = moderate. No cholesterol effect, excellent protein-per-calorie for a fat-loss phase, but no micronutrients.

4. FISH AND SEAFOOD

4.1 The supplement RCTs vs the whole-fish cohorts — these are DIFFERENT questions

This is the most commonly conflated distinction in the whole omega-3 literature. Handle them separately.

Supplement RCTs (this is the strong-design evidence, and it is mostly disappointing):

VITAL. Manson JE et al., NEJM 2019;380(1):23–32. RCT, n = 25,871, 1 g/day marine n-3 (Omacor) vs placebo, median 5.3 years, primary prevention. Primary major CV events HR 0.92 (0.80–1.06), P = 0.24 — null. Invasive cancer HR 1.03 (0.93–1.13) — null. All-cause mortality HR 1.02 (0.90–1.15) — null. But: total myocardial infarction HR 0.72 (0.59–0.90) — a 28% reduction in MI specifically, with stroke null (1.04) and CV death null (0.96). The MI finding is a secondary endpoint and should be treated as hypothesis-generating, but it is not nothing, and it was largest in low-fish-consumers.
Cochrane. Abdelhamid AS, Brown TJ, Brainard JS, et al., Cochrane Database of Systematic Reviews 2020 — 86 RCTs, 162,796 participants. All-cause mortality RR 0.97 (0.93–1.01), high-certainty; cardiovascular mortality RR 0.92 (0.86–0.99), moderate-certainty; coronary events RR 0.91 (0.85–0.97), low-certainty. So: essentially no effect on total mortality, and a small, moderate-certainty effect on CV mortality. This is the fairest single summary of the supplement question.
REDUCE-IT. Bhatt DL, Steg PG, Miller M, Brinton EA, Jacobson TA, Ketchum SB, Doyle RT Jr, Juliano RA, Jiao L, Granowitz C, Tardif JC, Ballantyne CM; REDUCE-IT Investigators. NEJM 2019;380(1):11–22 (online 10 Nov 2018), DOI 10.1056/NEJMoa1812792. RCT, n = 8,179 statin-treated patients (70.7% secondary prevention) with fasting TG 135–499 mg/dL and LDL-C 41–100 mg/dL, 4 g/day icosapent ethyl (purified EPA ethyl ester) vs placebo, median 4.9 years. Primary composite endpoint: 17.2% vs 22.0%, HR 0.75 (95% CI 0.68–0.83), P < 0.001. Key secondary: 11.2% vs 14.8%, HR 0.74 (0.65–0.83). Cardiovascular death 4.3% vs 5.2%, HR 0.80 (0.66–0.98), P = 0.03. Harms: hospitalization for atrial fibrillation/flutter 3.1% vs 2.1% (P = 0.004); serious bleeding 2.7% vs 2.1% (P = 0.06). Funded by Amarin Pharma. This is a genuinely positive, large, well-powered RCT — but the drug is a 4 g/day purified pharmaceutical EPA in a high-risk statin-treated population, which has essentially nothing to do with whether a healthy adult should eat salmon.
STRENGTH. Nicholls SJ, Lincoff AM, Garcia M, et al., JAMA 2020 — RCT, n = 13,078, 4 g/day omega-3 carboxylic acid (EPA+DHA) vs corn oil, in high-CV-risk patients. Primary composite HR 0.99 (0.90–1.09), P = 0.84; event rates 12.0% vs 12.2% — completely null. GI adverse events 24.7% vs 14.7%.
The REDUCE-IT vs STRENGTH controversy is unresolved and matters. REDUCE-IT used a mineral oil placebo; the mineral-oil arm showed rises in LDL-C, non-HDL-C, apoB and CRP by end of study, while STRENGTH’s corn-oil arm did not. Two live hypotheses: (a) purified high-dose EPA is genuinely beneficial and EPA+DHA is not; (b) part or all of REDUCE-IT’s benefit is an artefact of a harmful comparator. Nobody has settled this. Design note: both are large, well-run RCTs; the disagreement is about the comparator, which is exactly the kind of problem RCTs are supposed to avoid.
ASCEND. ASCEND Study Collaborative Group (Bowman L, Mafham M, Wallendszus K, et al.), “Effects of n-3 Fatty Acid Supplements in Diabetes Mellitus,” NEJM 2018;379(16):1540–1550, DOI 10.1056/NEJMoa1804989. RCT, n = 15,480 people with diabetes and no evident CVD, 1 g/day n-3 vs placebo, mean 7.4 years (76% adherence). Serious vascular event: 8.9% vs 9.2%, rate ratio 0.97 (0.87–1.08), P = 0.55. Serious vascular event or revascularization: 11.4% vs 11.5%, RR 1.00 (0.91–1.09). All-cause death 9.7% vs 10.2%, RR 0.95 (0.86–1.05). Completely null in a large, long, high-risk primary-prevention population.

Bottom line on supplements: 1 g/day fish oil capsules do essentially nothing for hard endpoints in unselected people. Anyone claiming otherwise is over-reading VITAL’s MI subgroup or REDUCE-IT’s contested comparator.

Whole-fish cohorts (weaker design, different exposure, more favourable results):

Mohan D, Mente A, Dehghan M, et al., JAMA Internal Medicine 2021;181:631–649. Pooled analysis of 191,558 individuals from 4 cohorts — 147,645 from PURE (21 countries) plus 43,413 patients with vascular disease from 3 studies (40 countries), 58 countries total, 9.1 y follow-up in PURE. Key result: ≥175 g/week (2 servings) was associated with lower major CVD and total mortality among high-risk individuals and those with existing vascular disease — but NOT in general populations without vascular disease. In PURE’s general population, ≥350 g/week vs ≤50 g/month: major CVD HR 0.95 (0.86–1.04), total mortality HR 0.96 (0.88–1.05) — both null.
Zhong 2020 JAMA IM (US, n = 29,682): fish, per 2 servings/week — incident CVD HR 1.00 (0.98–1.02); all-cause mortality HR 0.99 (0.97–1.01). Null.

Honest synthesis: for a healthy adult primary-prevention subject, the evidence that eating fish will lower the adult’s cardiovascular risk is weak. PURE says the benefit concentrates in secondary prevention. The supplement RCTs say n-3 alone does little. The case for fish in this scoring system rests mainly on (a) nutrient density — EPA/DHA, vitamin D, selenium, iodine, B12; (b) exceptional protein-per-calorie and satiety, which directly serve the visceral-fat/lean-mass goal; (c) very low sodium in unprocessed form; and (d) displacement value — every fish meal is a meal that isn’t processed or red meat. That is a solid case. It is not “fish prevents heart attacks in healthy young adults,” and this section should not claim that.

Additional plausible-but-unsettled benefit relevant to this subject’s joint limits: EPA/DHA and exercise-induced muscle soreness / joint symptoms. The literature is small, heterogeneous, and mostly acute-surrogate. Label as hypothesis, not finding.

4.2 Farmed vs wild salmon — the popular belief is now roughly backwards

The origin of the belief: Hites RA, Foran JA, Carpenter DO, Hamilton MC, Knuth BA, Schwager SJ, Science 2004, “Global Assessment of Organic Contaminants in Farmed Salmon.” Analysed over 2 metric tons of farmed and wild salmon worldwide for 14 organochlorine contaminants. PCBs, dioxins, toxaphene and dieldrin were consistently and significantly higher in farmed than wild salmon; European farmed salmon was worse than North/South American farmed. The paper identified fish feed as the source and recommended restricted consumption of farmed salmon. This paper is the ancestor of essentially every “farmed salmon is toxic” claim still circulating.

What changed: the aquaculture industry substituted plant/vegetable oils for marine fish oils in feed. Norwegian salmon feed went from predominantly marine ingredients to roughly 70% plant-derived. Because the organochlorines rode in on the fish oil, contaminant levels fell sharply. Monitoring data show dioxins + dl-PCBs in fish oil falling from a mean of 9.6 ng TEQ05 in 2003 to 4.7 by 2015, and in Norwegian farmed salmon from 2.0 ng TEQ05 in 1999 to 0.6 in 2011, plateauing from about 2014.

Current measured levels — the key citation: Jensen IJ, et al., Foods 2020 (an update on fatty acids, dioxins, PCBs and heavy metals in farmed, escaped and wild Atlantic salmon in Norway; 20 farmed, 23 wild, 17 escaped):

	Farmed	Wild	EU maximum
Sum dioxins + dl-PCBs	0.51 ng TEQ/kg	1.48 ng TEQ/kg	6.5 ng TEQ/kg
Mercury	0.018 mg/kg	0.056 mg/kg	0.5 mg/kg
EPA + DHA (per 100 g fillet)	1.4 g (0.5 EPA + 0.9 DHA)	1.2 g (0.4 + 0.8)	—
Total fat	18%	6%	—

So, on current Norwegian data: farmed salmon has ~3× LOWER dioxins/dl-PCBs, ~3× LOWER mercury, and HIGHER absolute EPA+DHA per 100 g than wild. Both are far below EU limits.

The honest caveats, because this is not a clean win:

Farmed salmon’s omega-3 is proportionally much lower as a fraction of total fatty acids (7.5% vs 21.1%) — it is diluted by plant-oil-derived linoleic acid. The absolute EPA+DHA per serving is still higher only because farmed fish are fatter. If you eat to a calorie target (this subject does), farmed salmon delivers more calories per gram of EPA+DHA.
Farmed salmon has a substantially higher n-6:n-3 ratio than wild, from vegetable oil in feed. Whether that matters clinically is contested and I would not score on it.
This is Norwegian data. Chilean and other farmed sources are not necessarily equivalent; I did not verify non-Norwegian current contaminant data.
Antibiotic use, sea lice treatments, escapee/ecological impacts are real issues — but they are not human-nutrition issues for this scoring system.
Wild salmon is leaner (6% vs 18% fat), which for a calorie-budgeted fat-loss phase is a genuine advantage.

Verdict: “Is farmed salmon actually worse?” — No, not on contaminants, and not on absolute omega-3. It is worse on omega-3 density and calorie efficiency, and better on mercury and dioxins. Score them the same; if anything prefer wild/sockeye for the calorie budget, not for safety.

4.3 Mercury by species — concrete numbers and a serving ceiling for THIS subject

Source: FDA, “Mercury Levels in Commercial Fish and Shellfish (1990–2012)” — mean and maximum total mercury, ppm (= mg/kg), retrieved in-session.

Species	Mean Hg (ppm)	Max (ppm)
Tilefish (Gulf of Mexico)	1.123	3.73
Swordfish	0.995	3.22
King mackerel	0.73	1.67
Bigeye tuna (fresh/frozen)	0.689	1.816
Yellowfin tuna (fresh/frozen, “ahi”)	0.354	1.478
Albacore tuna (canned)	0.350	0.853
Halibut	0.241	1.52
Skipjack / canned light tuna	0.126	0.889
Cod	0.111	0.989
Chub mackerel (Pacific)	0.088	0.19
Herring	0.078	0.56
Trout (freshwater)	0.071	0.678
Atlantic/N. Atlantic mackerel	0.05	0.16
Pollock	0.031	0.78
Salmon (fresh/frozen)	0.022	0.19
Anchovies	0.016	0.049
Sardine	0.013	0.083
Tilapia	0.013	0.084
Shrimp	0.009	0.05
Scallop	0.003	0.033

(Mussels were not in the FDA table retrieved; bivalves are generally in the same very-low band as scallops. Treat as low but [UNVERIFIED].)

Serving-ceiling arithmetic for an ~80 kg adult male. EPA reference dose for methylmercury = 0.1 µg/kg body weight/day. 80 kg × 0.1 = 8 µg/day = 56 µg/week. Grams of fish per week that reach the RfD:

Species	g/week to hit RfD	≈ 85 g (3 oz) servings/week
Swordfish	56 g	0.7
Bigeye tuna	81 g	~1
Ahi/yellowfin steak	158 g	~1.9
Albacore (“solid white”) canned	160 g	~1.9
Halibut	232 g	~2.7
Cod	505 g	~5.9
Skipjack / “chunk light” canned	444 g	~5.2
Salmon	2,545 g	~30
Sardines / anchovies	~3.5–4.3 kg	effectively unlimited
Shrimp	6.2 kg	effectively unlimited

Important honesty about what the RfD is. The EPA RfD is anchored on fetal neurodevelopment (Faroe Islands / Seychelles cohorts) and carries an uncertainty factor of roughly 10. It is not a threshold above which a adult man is harmed; it is a conservative bound designed to protect the most sensitive subpopulation. There is no good evidence of neurological harm in healthy adult males at 2–3× the RfD. Practical guidance for this subject: don’t cap salmon, sardines, shrimp, cod, tilapia at all. Treat albacore/solid-white tuna and tuna steak/ahi/poke as ~2 servings/week, and swordfish/king mackerel/tilefish/bigeye as occasional. Canned light/skipjack is the default canned tuna — it is 2.8× lower in mercury than albacore for the same protein and cost.

Also worth stating as a hypothesis, not a fact: fish selenium may bind and offset methylmercury toxicity (“selenium health benefit value”). Mechanistic and animal evidence; not established in humans. Do not score on it.

4.4 Oily vs white fish, and the shellfish

Oily fish (salmon, sardines, mackerel, anchovy, herring, trout, arctic char, sablefish): the EPA+DHA, vitamin D and selenium carriers. Salmon at ~1.2–1.4 g EPA+DHA per 100 g (Jensen 2020) delivers in one serving what a 1 g fish oil capsule cannot. Sardines and anchovies additionally supply calcium (edible bones) and are among the lowest-mercury, lowest-contaminant, most sustainable options. Sodium caveat: canned sardines/anchovies in brine and all smoked fish are high-sodium — anchovies in particular are effectively a salt product.

White fish (cod, haddock, pollock, tilapia, halibut, flounder, sole, barramundi, hake, swai/basa): almost no omega-3, but the single best protein-per-calorie in the whole food supply — roughly 20 g protein per ~90 kcal per 100 g cooked, near-zero saturated fat, very high satiety index, negligible sodium unformed. For someone cutting fat while protecting lean mass, white fish is close to an ideal macro tool even with no omega-3 story at all. Halibut is the only one with a mercury consideration (0.241 ppm).

Shellfish:

Shrimp/prawns — 0.009 ppm Hg (lowest measured), very high protein-per-calorie, high selenium and B12. The dietary-cholesterol content (~190 mg/100 g) is irrelevant given everything in §3.3. Watch sodium: many commercial shrimp are STPP-brined and can carry several hundred mg sodium per serving.
Scallops — 0.003 ppm Hg, lean, high protein. Also frequently STPP-treated (“wet” scallops) with added sodium.
Mussels, clams, oysters — outstanding micronutrient density (mussels/clams are among the richest food sources of B12 and iron; oysters for zinc), very low trophic level so very low mercury and low PCB. Some of the best-value protein in the section. Frequently cooked in wine/broth with added salt in delivery meals.
Imitation crab / surimi / “krab” — do NOT score as shellfish. It is a processed, starch-bound, high-sodium, low-protein product. Score near processed meat.

Canned tuna: light/skipjack is the workhorse (+4). Albacore and tuna steak/ahi carry the mercury ceiling above (+2). Tuna packed in oil adds calories; tuna in water is the fat-loss-appropriate form. Sodium: USDA gives canned white tuna in oil, 3 oz = 337 mg; “no salt added” versions exist and are meaningfully lower.

Smoked/cured fish (lox, gravlax, smoked trout/mackerel, kippers): retains the omega-3, but is a cured, high-sodium product. Score it down toward neutral for a sodium-cutting subject. Note it also falls under the processed-meat processing definition in some frameworks, though IARC’s Group 1 evaluation was of processed meat, not processed fish, and the cohort evidence for processed fish is thin.

4.5 Fish verdict

Oily fish: +7, beneficial, high confidence. Nutrient density + protein quality + very low sodium + displacement + a plausible (not proven) CV benefit + relevance to this subject’s joint/recovery concerns.
White fish: +5, beneficial, moderate confidence. Best protein-per-calorie in the section, near-zero downside.
Shellfish (shrimp/scallop/mussel/clam/oyster/crab/lobster/squid/octopus): +4, beneficial, moderate confidence. Docked from +5 only for the added-sodium processing common in commercial supply.
Canned light/skipjack tuna: +4, beneficial, moderate.
Albacore / tuna steak / ahi / poke: +2, beneficial, moderate. Real food, real mercury ceiling.
Smoked/cured fish (lox, smoked salmon, kippers, anchovies-as-condiment): +1, contested, moderate. Omega-3 intact, sodium load significant.
Imitation crab/surimi: −3, harmful, moderate.

5. PLANT PROTEINS

5.1 Soy — settling the isoflavone question

Soy isoflavones (genistein, daidzein, glycitein) are selective estrogen receptor modulators, with preferential affinity for ERβ over ERα. They are not estrogen, and the ERβ-preference is why the mechanistic prediction of “feminization” was wrong from the start.

(a) Male feminization / testosterone — settled, no effect. Reed KE, Camargo J, Hamilton-Reeves J, Kurzer M, Messina M. Reproductive Toxicology 2021;100:60–67, DOI 10.1016/j.reprotox.2020.12.019. Meta-analysis of 41 clinical studies: n = 1,753 men for total testosterone, 752 for free testosterone, 1,000 for estradiol, 239 for estrone, 967 for SHBG. Result: “Regardless of the statistical model, no significant effects of soy protein or isoflavone intake on any of the outcomes measured were found”; dose and duration sub-analyses were likewise null. Conclusion: “neither soy protein nor isoflavone exposure affects TT, FT, E₂ or E₁ levels in men.” COI flag, stated openly: Mark Messina is the author most associated with defending soy and is affiliated with the Soy Nutrition Institute. That is a genuine conflict and should raise scrutiny. However: (i) this is a meta-analysis of clinical trials, not a narrative review — the underlying data are public and the effect estimates cluster on null; (ii) the result is consistent with independent RCT literature; (iii) no well-conducted trial has ever shown soy lowering testosterone. The COI justifies scepticism about framing, not about the null itself.

Where the feminization claim came from: case reports, at absurd doses.

Martinez J & Lewi JE, Endocrine Practice 2008 — 60-year-old man with gynecomastia and breast tenderness consuming 3 quarts of soy milk per day; resolved on discontinuation. Three quarts is roughly 360 mg isoflavones/day, on the order of 9× typical high Asian intake.
Siepmann T et al., Nutrition 2011 — 19-year-old type 1 diabetic vegan male with hypogonadism attributed to large quantities of soy; testosterone normalized within a year of dietary change.
Also: Sea et al., J Pediatr Endocrinol Metab 2021 (8-year-old boy, prepubertal gynecomastia); Widjajahakim et al., Int Med Case Rep J 2025 (33-year-old man, 0.5–1 L/day homemade soy milk). Design label: n = 1 case reports. Four case reports over ~17 years, at intakes 5–10× normal, against 41 clinical studies showing nothing. This is as close to settled as nutrition gets. Normal soy consumption does not feminize men.

(b) Thyroid — a real but clinically trivial effect in iodine-replete people. Otun J et al., Scientific Reports 2019 — systematic review and meta-analysis of RCTs, 18 articles. Soy supplementation produced no significant change in free T3 or free T4, and a statistically significant but very small rise in TSH (WMD 0.248 mIU/L, P = 0.049). Authors: “Soy supplementation has no effect on the thyroid hormones and only very modestly raises TSH levels, the clinical significance, if any, of the rise in TSH is unclear.” Corroborating RCT: Liu ZM et al., Phytotherapy Research 2021, double-blind placebo-controlled, n = 270 postmenopausal equol-producers: “consumption of whole soy and purified daidzein at the provided dosages are safe and have no detrimental effect on thyroid function.” Where a real caution exists: iodine deficiency, and (separately) soy interfering with levothyroxine absorption if taken simultaneously. Neither applies to a healthy iodine-replete US active adult. Do not penalise soy on thyroid grounds.

(c) Breast cancer — the data run the OPPOSITE direction from the folk belief.

Shu XO, Zheng Y, Cai H, Gu K, Chen Z, Zheng W, Lu W. JAMA 2009 — population-based prospective cohort, 5,042 breast cancer survivors, median 3.9 y. Highest vs lowest quartile of soy protein: total mortality HR 0.71 (0.54–0.92), recurrence HR 0.68 (0.54–0.87).
Nechuta SJ, Caan BJ, Chen WY, et al. AJCN 2012 — pooled prospective analysis of 2 US + 1 Chinese cohort, 9,514 breast cancer survivors, mean 7.4 y. Isoflavone ≥10 mg/day: recurrence HR 0.75 (0.61–0.92); breast cancer mortality HR 0.83 (0.64–1.07).
Qiu S & Jiang C, Eur J Nutr 2019 — meta-analysis, 12 articles, 37,275 women: pre-diagnosis soy and overall survival HR 0.84 (0.71–0.98). These are cohorts, so causality is not established and residual confounding (soy intake tracks with Asian ancestry, which tracks with other exposures) is live. But the direction is consistent across US and Chinese populations, and there is no cohort showing harm. The “breast cancer survivors should avoid soy” advice is not supported by human data; it descends from rodent studies in ovariectomized athymic mice, a model whose relevance to humans is poor.

(d) What soy actually does for THIS subject.

Protein quality: soy protein isolate and tofu are complete proteins with DIAAS in the same range as animal proteins; soy is the only major plant protein that is not limiting in lysine. Leucine content is lower than whey per gram, and splanchnic extraction is higher, so per-gram anabolic efficiency is modestly lower — but in trials with adequate total protein, the difference disappears (§5.3).
LDL: Blanco Mejia S, Messina M, Li SS, Viguiliouk E, Chiavaroli L, Khan TA, Srichaikul K, Mirrahimi A, Sievenpiper JL, Kris-Etherton P, Jenkins DJA. Journal of Nutrition 2019 — meta-analysis of 46 controlled trials (43 with usable data): soy protein lowered LDL-C by −4.76 mg/dL (95% CI −6.71 to −2.80, P < 0.0001). Design: meta-analysis of RCTs. Small but real, and in the opposite direction from eggs and from meat (APPROACH).
Sodium: tofu and edamame are essentially sodium-free. Soy sauce, tamari, shoyu and miso are NOT — they are among the highest-sodium ingredients in any cuisine. The detection layer must not treat “soy sauce” as a soy-protein hit; it is a sodium hit.
Tempeh adds fermentation, more fibre and higher protein density than tofu; edamame adds fibre and is a whole-food form.

5.2 Seitan / vital wheat gluten — gluten in non-celiac people

Is gluten harmful to people without celiac disease? The best available cohort evidence says no: Lebwohl B, Cao Y, Zong G, Hu FB, Green PHR, Neugut AI, Rimm EB, Sampson L, Dougherty LW, Giovannucci E, Willett WC, Sun Q, Chan AT. BMJ 2017 — prospective cohort, 64,714 women + 45,303 men without prior heart disease, 26 years, 2,273,931 person-years, 2,431 + 4,098 CHD events. Highest vs lowest gluten intake, multivariable-adjusted HR 0.95 (0.88–1.02), p-trend 0.29 — no association. Crude event rates were 352 vs 277 per 100,000 person-years (lowest vs highest gluten). The authors’ own caveat is the important one: avoiding gluten leads to reduced whole grain intake, which may itself increase cardiovascular risk — the harm of a gluten-free diet in non-celiac people is more plausible than the harm of gluten.

Non-celiac gluten sensitivity (NCGS) exists as a clinical entity, but double-blind gluten-challenge studies repeatedly fail to reproduce symptoms with gluten specifically; FODMAPs in wheat and nocebo effects are better-supported explanations. A recent double-blind challenge study from the Biesiekierski group (published in United European Gastroenterology Journal, DOI 10.1002/ueg2.70014; 20 healthy controls and 16 people with self-reported NCGS) found that after sub-acute administration, abdominal pain (P < 0.001) and bloating (P = 0.001) were higher in the NCGS group “regardless of nutrient intake” — i.e. on placebo as much as on gluten — and no differences in biological markers. The authors’ conclusion: NCGS symptoms “are not gluten-specific… may be explained by nocebo effects, warranting… re-evaluating the NCGS definition.” [Verified: authors, journal, DOI, design and numbers. I did not separately verify the older, more famous Biesiekierski 2013 Gastroenterology FODMAP re-challenge trial.]

Nutritionally, seitan for this subject: very high protein (~25 g/100 g), near-zero fat, cheap. Two real limitations: (1) it is limiting in lysine — as a sole protein source it is inferior, though irrelevant in a mixed diet; (2) commercial seitan is almost always cooked or braised in soy sauce/tamari, making it a significant sodium source. For a sodium-cutting subject that is the operative issue, not the gluten.

5.3 Pea protein, TVP, and plant vs animal protein for lean mass

The practically important question for this subject is whether plant protein preserves lean mass as well as animal protein in a fat-loss phase. Evidence:

Davis BE, Young I, Giglotti JC, Yao J, Tou JC. Journal of Dietary Supplements 2026 — systematic review and meta-analysis of RCTs, 12 studies, 261 participants: “no significant effect of either whey or soy protein supplementation on LBM,” but whey improved bench press (MD 8.87) and squat (MD 9.60) performance while soy showed no strength advantage.
Santini MH, Erwig Leitão A, Mazzolani BC, et al. Journal of the International Society of Sports Nutrition 2025 — RCT, 44 untrained males: plant-based and animal-based protein blends produced similar gains — both ~2.4–2.5 kg whole-body lean mass and ~63–64 kg leg-press strength, no between-group difference.
Candow DG, Burke NC, Smith-Palmer T, Burke DG, IJSNEM 2006 — RCT, 27 untrained young adults: whey and soy equivalent, both > placebo.
Korzepa et al., European Journal of Nutrition 2025 — randomized crossover: whey concentrate produced greater peak plasma leucine (P = 0.032) and iAUC (P = 0.012) than pea protein isolate at the same dose; the authors note whether this translates to muscle anabolism “remains to be determined.”
Babault N, Païzis C, Deley G, Guérin-Deremaux L, Saniez MH, Lefranc-Millot C, Allaert FA. Journal of the International Society of Sports Nutrition 2015;12(1):3, DOI 10.1186/s12970-014-0064-5 — now verified. Double-blind RCT, 161 males aged 18–35, 12 weeks of upper-limb resistance training, randomized to pea protein (n = 53), whey (n = 54) or placebo (n = 54), 25 g twice daily; biceps thickness by ultrasound, strength by isokinetic dynamometry. Read this one carefully, because it is routinely over-sold: the primary analysis found a significant time effect on biceps thickness (24.9 → 27.3 mm, P < 0.0001) but only “a trend toward significant differences between groups (P = 0.09)” — i.e., the headline comparison was null. The positive result came from a post-hoc sensitivity analysis restricted to the weakest participants at baseline (+20.2% pea, +15.6% whey, +8.6% placebo; P < 0.05, pea > placebo, whey not different from either). COI: two authors (Guérin-Deremaux, Saniez, Lefranc-Millot) are employees of Roquette, the manufacturer of the NUTRALYS pea protein tested. Fair summary: pea protein was not shown to be inferior to whey, which is the useful claim; it was not shown to beat placebo in the full sample.

Synthesis: when total daily protein is adequate (≥1.6 g/kg, which any active adult tracking body composition will be), the source matters much less than the total. Plant proteins have lower leucine per gram and higher splanchnic extraction, so a modestly higher dose is sensible, but the RCT record does not support a meaningful lean-mass penalty. Pea protein specifically has a reasonable leucine content (better than most plant isolates) and is a legitimate whey substitute.

TVP / textured soy protein / soy curls: nutritionally these are soy protein concentrate — high protein, high fibre, no cholesterol, no sodium as the raw ingredient. But in real meal-delivery products they are almost always in a seasoned/sauced matrix. Score the raw ingredient positively but expect the sodium to arrive with the seasoning.

Plant-based meat analogs (Beyond, Impossible, etc.) are a distinct class — high sodium (often 350–450 mg/serving), coconut/refined oils, ultra-processed formulation — and should not be scored as “soy” or “pea protein.” Score neutral to mildly negative; they are better than processed meat and worse than whole soy.

5.4 Plant protein verdicts

Whole soy foods (tofu, tempeh, edamame, natto, bean curd, yuba): +5, beneficial, high confidence. Complete protein, LDL-lowering in RCT meta-analysis, essentially sodium-free, no credible endocrine or thyroid harm, favourable breast-cancer-survivor data, high satiety per calorie. The male-feminization claim is dead.
Soy protein isolate / concentrate / TVP: +3, beneficial, moderate. Same protein, less food matrix, usually arrives seasoned.
Pea protein / pea protein isolate: +3, beneficial, moderate.
Seitan / vital wheat gluten: +1, neutral, moderate. Excellent macros, lysine-limiting, and in practice a sodium vector.
Plant-based meat analogs: 0, contested, low-moderate.

6. DAIRY

6.1 The fermented vs non-fermented divergence

The most reproducible pattern in the dairy literature is that fermented dairy (yogurt, kefir, cheese) looks better in cohorts than non-fermented dairy (milk, butter, cream), and yogurt looks best of all.

Companys J et al., Advances in Nutrition 2020 — systematic review and meta-analysis, 20 cohort studies + 52 RCTs. Yogurt intake associated with 27% lower T2D risk (RR 0.73, 0.70–0.76); fermented milk associated with ~4% lower cardiovascular risk.
Zhang K et al., Critical Reviews in Food Science and Nutrition 2019 — meta-analysis of 10 cohorts, 385,122 participants: fermented dairy and CVD OR 0.83 (0.76–0.91); yogurt strongest, OR 0.78 (0.67–0.89).
Zhang M et al., Canadian Journal of Diabetes 2022 — meta-analysis, 15 studies, 485,992 participants, 20,207 diabetes cases: higher fermented dairy, OR 0.925 (0.856–0.999), dose-dependent for yogurt.
Kouvari M et al., European Journal of Clinical Nutrition 2024 — ATTICA prospective cohort, 1,988 participants, 20 years, 718 CVD cases: substituting 1 serving of non-fermented with 1 serving of fermented dairy, HR 0.74 (0.53–0.92).
Dehghan M, Mente A, Rangarajan S, et al. (PURE). Lancet 2018;392(10161) — prospective cohort, 136,384 individuals, 21 countries, 9.1 years. Composite of mortality or major CV events, >2 servings/day total dairy vs none: HR 0.84 (0.75–0.94); milk >1 serving vs none 0.90 (0.82–0.99); yogurt 0.86 (0.75–0.99); cheese 0.88 (0.76–1.02); butter 1.09 (0.90–1.33). Note PURE found total dairy protective including whole-fat, which cuts against the low-fat-dairy orthodoxy — but PURE’s population is heavily weighted toward low- and middle-income countries where dairy is a marker of nutritional adequacy, and that is a strong confounder. Do not over-generalise PURE to a well-fed US active adult.

How much of the yogurt effect is real vs healthy-user? Yogurt is close to a perfect healthy-user marker in Western cohorts — it correlates with fibre, fruit, exercise, not smoking, higher education. RRs in the 0.73–0.90 band are exactly in the noise range where confounding can generate the whole signal. What is NOT confounded, and is why I still score yogurt highly for this subject: plain Greek yogurt is ~10 g protein per 100 g at ~59 kcal, roughly 50 mg sodium per 100 g (USDA: plain low-fat yogurt, 6 oz container = 119 mg sodium), high in calcium, and among the highest-satiety-per-calorie foods measured. Those are deterministic properties that serve the adult’s exact goals.

6.2 Cheese and the “dairy matrix”

The puzzle: cheese is 25–35% fat, of which ~60% is saturated. Standard lipid modelling (Keys/Hegsted equations) predicts cheese should raise LDL substantially. In cohorts it doesn’t behave that way, and in feeding trials it under-performs its own saturated fat content.

The “dairy matrix” hypothesis: the physical and chemical structure of cheese — a calcium-phosphate-casein network — changes fat bioaccessibility and increases faecal fat excretion; calcium forms insoluble soaps with fatty acids in the gut; the fermentation-derived peptides and the MFGM (milk fat globule membrane) may independently modify lipid handling. So the same grams of saturated fat delivered as cheese do less to LDL than delivered as butter.

Evidence quality — honestly, weaker than the confidence with which it is asserted. What I verified in-session:

Rooney et al., Atherosclerosis 2025 — parallel RCT, n = 197, 6 weeks: “cheese was found to lower total and LDL cholesterol, compared to deconstructed cheese” (i.e., the same fat, protein, calcium delivered as separate components). Within-sex analysis found different responses in men and women. This is the strongest direct test of the matrix hypothesis I could verify.
O’Connor et al., Food & Function 2024 — parallel RCT, n = 162, 6 weeks: no LDL difference between unmelted/melted cheese vs deconstructed; melted cheese increased total cholesterol.
Feeney EL et al., European Journal of Nutrition 2023 — crossover RCT, but n = 7 (high-calcium vs reduced-calcium cheese): fasting LDL-C significantly lower on the high-calcium cheese arm (P = 0.002). n = 7 is a pilot, not evidence.
Hjerpsted J, Leedo E, Tholstrup T, “Cheese intake in large amounts lowers LDL-cholesterol concentrations compared with butter intake of equal fat content,” AJCN 2011;94(6):1479–1484, DOI 10.3945/ajcn.111.022426 — the foundational matrix trial, now verified. Randomized crossover, n = 49 men and women, two 6-week periods plus a 14-day habitual run-in; participants replaced part of habitual dietary fat with 13% of energy from hard cheese or from butter. Result: “the cheese intervention resulted in lower serum total, LDL-, and HDL-cholesterol concentrations and higher glucose concentrations than did the butter intervention,” and cheese did not raise total or LDL-C versus the run-in period despite higher total and saturated fat intake. Important negative detail that undercuts the standard mechanistic story: “Fecal fat excretion did not differ between the cheese and butter periods.” So the calcium-soap/fat-malabsorption explanation was not supported by this trial’s own data — the effect is real, the mechanism is not established. Hjerpsted & Tholstrup’s own later review (Crit Rev Food Sci Nutr 2016;56(8):1389–1403) concluded across four intervention studies that there was “no harmful effect on cholesterol concentrations when comparing fat intake from cheese with fat from butter,” while prospective results were split (four null, one increased risk, two decreased, one sex-dependent).

So: the matrix hypothesis is plausible, has some direct RCT support (Rooney 2025) and some direct non-support (O’Connor 2024), and is not settled. The cohort evidence is consistent with cheese being neutral (PURE: HR 0.88, 0.76–1.02, CI crosses 1).

Cheese’s sodium load — the part that actually decides the score for this subject. From the USDA SR Legacy sodium table (household measures), verified in-session:

Parmesan, grated, 1 cup: 1,804 mg
Feta, crumbled, 1 cup: 1,708 mg
Mozzarella, low-moisture part-skim, 1 cup diced: 879 mg
Mozzarella, whole milk, 1 cup shredded: 544 mg
Cottage cheese, lowfat 1% milkfat, 4 oz: 459 mg
Cottage cheese, lowfat 2% milkfat, 4 oz: 348 mg
Cottage cheese, creamed, 4 oz: 356 mg
Processed American cheese spread, 1 cup diced: 2,275 mg

Cheese is calorie-dense (~400 kcal/100 g for hard cheeses) and sodium-dense. For a man on [calorie target removed] actively cutting sodium and visceral fat, cheese is the one dairy item whose properties most directly conflict with the adult’s stated goals, regardless of how the matrix debate resolves. That, not saturated fat, is the reason to score it at 0 rather than positive.

Cottage cheese and ricotta are a genuinely different sub-class and should be detected separately: cottage cheese is ~11–12 g protein per 100 g at ~80–100 kcal with moderate sodium, and is one of the best fat-loss-phase protein foods available (slow-digesting casein, high satiety, cheap). Score it well above aged/salted cheeses.

6.3 Is dairy fat different from other saturated fat? The C15:0/C17:0 evidence

The biomarker findings are strong and consistent:

Imamura F et al., PLoS Medicine 2018 — pooled analysis of 16 prospective cohorts, 63,682 participants, 15,180 incident T2D cases, ~9 years. Hazard ratios per 10th–90th percentile range of circulating fatty acid biomarkers of dairy fat: 15:0 → HR 0.80 (0.73–0.87); 17:0 → HR 0.65 (0.59–0.72); trans-16:1n-7 → 0.82 (0.70–0.96); sum of all three → 0.71 (0.63–0.79).
Trieu K et al., PLoS Medicine 2021 — Swedish cohort (n = 4,150) plus systematic review/meta-analysis of 18 studies. Highest vs lowest tertile: 15:0 and CVD, RR 0.88 (0.78–0.99); 17:0 and CVD, RR 0.86 (0.79–0.93); t16:1n-7, 1.01 (0.91–1.12), null.

But are they causal? Probably not, or at least unproven.

Steffen et al., Frontiers in Nutrition (2026), CARDIA and ARIC cohorts (n = 3,196 and 3,889) with two-sample Mendelian randomization: higher C15:0 associated with lower blood pressure observationally, no CVD association, and “Two-sample MR analyses provided no evidence for a causal effect of C15:0.” (Verified via Europe PMC index summary; I did not retrieve the full abstract — treat the exact wording as partially verified.)
Lampousi et al., Diabetologia 2023 (EPIC-InterAct, 11,124 cases / 14,866 subcohort) found low 17:0 associated with higher diabetes incidence, but noted that “low dairy intake was not associated with diabetes incidence” — i.e., the biomarker predicts what the food does not, which is a strong hint that 17:0 is reflecting something other than dairy consumption.
Prada et al., Clinical Nutrition 2021 (EPIC-Potsdam, 820 cases): associations present in women, absent in men.
Santaren et al., AJCN 2014 (IRAS, n = 659): serum 15:0 associated with lower diabetes risk (OR 0.73), and 15:0 was a significant biomarker of total dairy — so it cannot distinguish “dairy fat protects” from “people who eat dairy differ.”

Additional confound rarely mentioned: odd-chain fatty acids are also produced endogenously from gut-microbial propionate, which is generated by fibre fermentation. So circulating 15:0/17:0 partly reflects fibre intake and gut microbial composition, not only dairy. That alone could generate the observed inverse associations with T2D.

Verdict: dairy fat may behave differently from other saturated fat, but the C15:0/C17:0 biomarker literature does not establish it. Those are markers, and the one MR test I could find was null. Do not score dairy fat as protective.

6.4 Whey protein for this subject

The clearest, most directly applicable evidence in the entire section, because the endpoint is the subject’s actual goal.

Morton RW, Murphy KT, McKellar SR, Schoenfeld BJ, Henselmans M, Helms E, Aragon AA, Devries MC, Banfield L, Krieger JW, Phillips SM. British Journal of Sports Medicine 2018 — systematic review and meta-analysis of RCTs, 49 studies, 1,863 participants. Protein supplementation during resistance training increased 1RM strength by 2.49 kg (0.64–4.33), fat-free mass by 0.30 kg (0.09–0.52), and muscle fibre CSA by 310 µm² (51–570). Note the honest magnitude: +0.30 kg fat-free mass is small, and the effect plateaued at total protein intakes around 1.6 g/kg/day. Supplementation helps when it closes a protein gap; it does little when total protein is already adequate.
Khalafi et al., Healthcare 2025 — systematic review/pairwise meta-analysis, 25 studies, 1,454 participants: whey + resistance training increased lean body mass; no significant change in fat mass or body weight.
Mechanism: whey has the highest leucine density of common proteins (~10–11% leucine) and the fastest aminoacidemia, which maximises the acute mTORC1/MPS response. It is also among the most satiating protein sources per calorie. Both properties directly serve “lose visceral fat while preserving lean mass.”

Score whey/casein positively — but understand what the evidence supports. It supports “convenient way to hit a protein target that itself preserves lean mass,” not “whey is magic.” Casein/micellar casein is equivalent for daily totals, slower-digesting.

6.5 Full-fat vs low-fat for THIS subject

This is where the epidemiology is genuinely unresolved and the practical answer is nonetheless clear.

The cohort evidence does not support the low-fat-dairy orthodoxy: PURE found whole-fat dairy associated with lower composite risk; the biomarker studies point the same direction; and no cohort robustly shows full-fat dairy worse than low-fat for CVD or mortality.
But the cohorts cannot settle it, and the one causal-inference test available (MR on C15:0) was null.
For this subject the question is largely moot, because the operative constraint is calories and protein density, not fat quality. On a [calorie target removed] fat-loss budget: nonfat Greek yogurt gives ~10 g protein/59 kcal; whole-milk Greek yogurt gives ~9 g protein/97 kcal. Skim milk ~8 g protein/83 kcal per cup vs whole ~8 g/149 kcal. The low-fat versions win on protein-per-calorie by a wide margin, which is what preserves lean mass in a deficit. Recommend low/nonfat versions on macro grounds, not on saturated-fat-fear grounds — and say so, because the reasoning matters for how the adult generalises it.

6.6 Dairy verdicts

Plain yogurt / Greek yogurt / kefir / skyr / labneh: +6, beneficial, moderate confidence. Best-evidenced fermented dairy signal (T2D RR 0.73–0.93 across meta-analyses), plus deterministic protein density, satiety, and ~50 mg sodium/100 g. Sweetened/flavoured versions should have their added sugar penalised by the carbohydrate section, not here.
Cottage cheese / ricotta / quark: +4, beneficial, moderate. Excellent protein-per-calorie, moderate sodium (348–459 mg per 4 oz — real but manageable).
Aged / salty cheese (cheddar, parmesan, feta, blue, halloumi, provolone, gouda, manchego): 0, contested, moderate. Matrix hypothesis is plausible but unsettled; cohorts neutral; but calorie-dense and sodium-dense against this subject’s two active goals.
Processed cheese (American slices, cheese sauce, queso, cheese spread): −3, harmful, moderate. 2,275 mg sodium per cup diced for American cheese spread, emulsifying salts, low protein per calorie.
Milk: +2, neutral, moderate. Good protein and micronutrients, ~120 mg sodium/cup, cohorts neutral-to-favourable. Low-fat preferred for the calorie budget.
Whey / casein / milk protein isolate: +4, beneficial, moderate.
Butter/cream as an added fat: −2, contested, low-moderate. PURE’s butter HR was 1.09 (0.90–1.33), null; but it is pure energy density with no protein, which is the opposite of what this subject needs. Largely belongs to the fats section.

7. DETECTION STRING NOTES — ambiguity warnings for the matcher

These are the substring collisions that will produce false positives if matched naively. Recommend word-boundary matching (\bham\b) rather than raw substring for the flagged items.

String	Collides with	Handling
`ham`	hamburger, hamachi (yellowtail), graham, Birmingham	Require word boundary; prefer `deli ham`, `smoked ham`, `black forest ham`, `ham steak`, `honey ham`, `serrano ham`
`egg`	eggplant, egg noodles, egg wash, eggnog, egg roll wrapper	Require word boundary or a qualifier; `eggplant` must be an explicit exclusion
`bacon`	bacon bits (often soy-based imitation), “bacon fat”	Usually safe; treat `bacon bits` as low-dose
`sausage`	chicken sausage (still processed), fresh/uncured breakfast sausage, vegan sausage, “sausage seasoning”	Always processed-meat class unless `vegan`/`plant-based`/`meatless` precedes it
`turkey`	turkey bacon, deli turkey (both processed); ground turkey (not processed)	Order matters — check processed qualifiers first
`chicken`	chicken broth/stock/bouillon/base/fat, “chicken of the woods” mushroom	Exclude broth/stock/base/bouillon; these are sodium hits, not protein servings
`beef`	beef broth/stock/base/bouillon/tallow, beefsteak tomato	Same handling
`soy`	soy sauce, tamari, shoyu, soybean oil, soy lecithin — none are soy protein	`soy sauce`/`tamari`/`shoyu` must route to the SODIUM class, never the soy-protein class
`milk`	coconut milk, almond milk, oat milk, soy milk, milk chocolate, milk powder	Require dairy qualifier
`cheese`	vegan cheese, cheesecake, cheese sauce, nacho cheese, cream cheese frosting	Split processed-cheese and vegan-cheese sub-classes
`crab`	imitation crab, krab, surimi	Route imitation crab to the processed class
`tuna`	albacore/solid white vs chunk light — different mercury classes	Distinguish; default unqualified `tuna` to light/skipjack
`salmon`	smoked salmon, lox, gravlax — high sodium sub-class	Detect the cured qualifiers first
`pea`	green peas (vegetable), split pea (legume)	Only `pea protein` / `pea protein isolate` counts here
`uncured`, `no nitrate`, `nitrate-free`, `celery powder`, `cultured celery`	Appear as reassurance language on processed meat	These strings should INCREASE confidence that the item is processed meat, not decrease the penalty
`bologna`	`bolognese` does not contain `bologna` (safe), but “bologna sauce” spellings exist	Low risk

8. MACHINE-READABLE SCORING TABLE

Section 02 — CARBOHYDRATES

Subject the scores are calibrated for: a healthy general-population adult. [personal profile and health goals removed]

Evidence-design key used throughout: MA-RCT = meta-analysis of randomised controlled trials · RCT = individual randomised trial · XO = crossover feeding trial · PC = prospective cohort · MA-PC = meta-analysis of prospective cohorts · MR = Mendelian randomisation · MECH = mechanistic/in-vitro · REG = regulatory risk assessment · COMP = food-composition data

Standing caveat on nutritional epidemiology. Almost every relative risk in Sections 1–3 comes from FFQ-based prospective cohorts. Three problems recur and I flag them inline rather than repeating:

Healthy-user confounding. Verified, not hypothetical: Jia, Miller & Jenkins (J Nutr 2026;156(7):101560, perspective) state directly that “whole grain consumers tend to smoke less, exercise more, and have higher levels of education, all of which independently reduce NCD risk.”
Exposure misclassification. Aune 2016 (below) explicitly names it: products classified as “whole grain” in the source cohorts ranged from 25% to 100% whole grain content. Jia 2026 found that of 12 major cohorts (>100,000 participants each), only 3 whole grain foods (popcorn, oatmeal, brown rice) appeared in 7 or more of them, and 17 of 54 whole grain foods were listed by a single cohort only.
The noise floor. RRs of 1.1–1.3 from FFQ cohorts are inside the range that residual confounding and measurement error alone can generate. I treat any RR in that band as hypothesis-generating, not as a fact about food.

Whole vs refined grains

The headline cohort numbers (verified)

Aune D, Keum N, Giovannucci E, et al. “Whole grain consumption and risk of cardiovascular disease, cancer, and all cause and cause specific mortality: systematic review and dose-response meta-analysis of prospective studies.” BMJ 2016;353:i2716. MA-PC. 45 cohort studies (64 publications), 705,253 participants. (PMC4908315)

Summary RR per 90 g/day of whole grains (≈3 servings):

Outcome	RR (95% CI)	n studies
Coronary heart disease	0.81 (0.75–0.87)	7
Stroke	0.88 (0.75–1.03) — not significant	6
Total CVD	0.78 (0.73–0.85)	10
Total cancer mortality	0.85 (0.80–0.91)	6
All-cause mortality	0.83 (0.77–0.90)	11
Respiratory disease mortality	0.78 (0.70–0.87)	4
Infectious disease mortality	0.74 (0.56–0.96)	3
Diabetes mortality	0.49 (0.23–1.05) — not significant	4

Risk continued falling up to 210–225 g/day. Whole grain bread, whole grain breakfast cereals and added bran carried the association; refined grains, white rice and total rice showed no clear benefit in the same analysis. The authors flag exactly the misclassification problem noted above.

Reynolds A, Mann J, Cummings J, Winter N, Mete E, Te Morenga L. “Carbohydrate quality and human health: a series of systematic reviews and meta-analyses.” Lancet 2019;393(10170):434–445. MA-PC + MA-RCT. 135 million person-years of prospective data; 58 clinical trials, 4,635 adults. (WHO-commissioned; verified from the peer-reviewed manuscript version at University of Dundee Discovery.)

Whole grains, highest vs lowest intake:

Outcome	Design	Cases / PY	RR (95% CI)	GRADE
All-cause mortality	PC (9)	99,224 / 10.7 M PY	0.81 (0.72–0.90)	Low
CHD mortality	PC (2)	1,588 / 2.0 M PY	0.66 (0.56–0.77)	Low
CHD incidence	PC (6)	7,697 / 2.8 M PY	0.80 (0.70–0.91)	Low
Stroke incidence	PC (3)	1,247 / 1.1 M PY	0.86 (0.61–1.21) ns	Very low
Cancer mortality	PC (5)	32,727 / 10.1 M PY	0.84 (0.76–0.92)	Low
Type 2 diabetes incidence	PC (8)	14,686 / 3.9 M PY	0.67 (0.58–0.78)	Low
Colorectal cancer	PC (7)	8,803 / 6.8 M PY	0.87 (0.79–0.96)	Moderate

Dose-response, assuming linearity, per 15 g/day more whole grain: all-cause mortality RR 0.94 (0.92–0.95); CHD incidence 0.93 (0.89–0.98); T2D incidence 0.88 (0.81–0.95); colorectal cancer 0.97 (0.95–0.99).

ABSOLUTE risk translation

Reynolds et al. did the arithmetic themselves and it is the honest number to quote: for whole grains, high vs low intake corresponds to 26 fewer deaths (95% CI 14–39) and 7 fewer CHD cases (95% CI 3–10) per 1,000 participants over the duration of the studies (typically 10–25 years, cohorts with mean age ~50–65). For dietary fibre the equivalents are 13 fewer deaths and 6 fewer CHD cases per 1,000.

Rendered as NNT-style: ~38 people would need to be lifelong high-whole-grain eaters for ~20 years for one fewer death, if the association is fully causal. It almost certainly is not fully causal, given point 1 above.

For type 2 diabetes I use the baseline incidence from the very cohorts that generate these HRs — Muraki 2016 reports 15,362 T2D cases over 3,988,007 person-years = 3.85 cases per 1,000 person-years in US health professionals (NHS/NHSII/HPFS). Applying the Reynolds whole-grain RR of 0.67:

3.85 × 0.67 = 2.58 per 1,000 PY → −1.27 per 1,000 PY → −12.7 cases per 1,000 people per decade.
For this subject specifically: the adult is lean, 29, and trains hard. The adult’s true baseline is plausibly one-third to one-fifth the cohort mean (which is driven by 50–65-year-olds with BMI ~26). Scaled: −2.5 to −4 cases per 1,000 per decade. That is a real but small number, and it is the honest size of the prize.

Now the crucial practical question: does “whole grain” on a label mean anything?

Short answer: it means something, but far less than a shopper assumes, and the RCT evidence is much weaker than the cohort evidence.

Stamp thresholds (verified from the Whole Grains Council, an industry programme run by Oldways — not a government scheme):

100% Stamp: all grain ingredients are whole grain, ≥16 g whole grain per labelled serving.
50%+ Stamp: at least half the grain ingredients are whole grain, ≥8 g per labelled serving.
Basic Stamp: ≥8 g per labelled serving, but the product “contain[s] primarily refined grain.”

So a product can bear a Whole Grain Stamp while being mostly white flour. 8 g of whole grain is roughly half a serving; the USDA/dietary-guidelines “serving” is 16 g.

FDA: the 2006 Draft Guidance for Industry and FDA Staff: Whole Grain Label Statements proposed 51% whole grain by weight for a whole-grain claim, but it was never finalised. There is no binding federal minimum for “made with whole grains.” The only federally enforceable whole-grain-adjacent claim is the health claim at 21 CFR 101.81 (soluble fibre / beta-glucan — see the barley/oats subsection), which is a fibre claim, not a whole-grain claim.

Verification note: I could not fetch the FDA guidance page or eCFR directly in this session (404 / redirect-block). The “51%, never finalised” status is consistent across secondary sources but I am flagging it as not primary-verified in this session.

Ingredient-list decoding, in order of trustworthiness:

whole wheat flour / whole grain wheat flour / whole durum wheat — genuinely whole grain (bran + germ + endosperm).
wheat flour, enriched wheat flour, enriched bleached flour, unbleached flour, semolina, durum flour — all refined. “Wheat flour” is not whole wheat. “Enriched” means iron + four B vitamins were added back after the bran and germ were removed; it does not restore fibre, magnesium, or the phytochemicals.
multigrain, stone-ground, made with whole grains, 7-grain, wheat bread, honey wheat — meaningless without checking the flour. Caramel colour is routinely used to make refined bread look whole.
whole grain oats, rolled oats, steel-cut oats — oats are almost never sold de-branned; these are whole grain by default.

Is fibre the active ingredient, or particle size, or the bran micronutrients?

Finely-milled whole grain flour has a glycaemic response close to white flour. This is verified, not folklore.

Zafar TA, Aldughpassi A, Al-Mussallam A, Al-Othman A. “Microstructure of Whole Wheat versus White Flour and Wheat-Chickpea Flour Blends and Dough: Impact on the Glycemic Response of Pan Bread.” Int J Food Sci 2020;2020:8834960. RCT/XO + MECH. Direct quote: “For nutritional consideration, whole wheat bread is recommended over the white bread. However, it has a similarly high effect on glycemic response (GR) as the white bread.” Mechanism measured in the same paper: whole wheat flour had a higher alpha-amylase activity (significantly lower falling number) than white flour, and a much wider particle-size distribution (z ≈ 1679.5 nm, PDI 1.0 vs 658.9 nm, PDI 0.74 for white). Substituting chickpea flour into either bread significantly lowered GR — i.e. the legume, not the whole-grain status, moved the needle.

Atkinson FS, Brand-Miller JC, Foster-Powell K, Buyken AE, Goletzke J. “International tables of glycemic index and glycemic load values 2021: a systematic review.” Am J Clin Nutr 2021;114(5):1625–1632. >4,000 entries. Their own summary: “Cereals and cereal products, however, including whole-grain or whole-meal versions, showed wide variation in GI values, presumably arising from variations in manufacturing methods. Breads, breakfast cereals, rice, savory snack products, and regional foods were available in high-, medium-, and low-GI versions.” Meanwhile “Dairy products, legumes, pasta, and fruits were usually low-GI foods (≤55)… and had consistent values around the world.” Whole-grain status does not predict GI. Food form does.

Jenkins’ own current position (Jia Y, Miller V, Jenkins DJ, J Nutr 2026;156(7):101560): “not all whole grains are equivalent, and the degree of processing matters. Intact whole barley, as consumed in barley stew, produces a lower postprandial glucose response (or glycemic index) than milled barley flour used in breads.”

Best current synthesis: the health signal attached to “whole grains” is probably carried by (a) total dietary fibre, especially cereal fibre — Reynolds found fibre and whole grains produce near-identical outcome estimates, and graded fibre evidence higher (moderate) than whole-grain evidence (low); (b) physical intactness / particle size, which governs glycaemic and satiety response independently of composition; (c) the bran micronutrient and phytochemical load (magnesium, folate, alkylresorcinols, phenolics), for which the human causal evidence is weakest; and (d) a large chunk of healthy-user confounding. Anyone who tells you it is definitively (a) or definitively (c) is over-claiming.

The intervention trials — where the story gets much thinner

This is where the wellness narrative and the evidence part company.

Brownlee IA, Moore C, Chatfield M, et al. “Markers of cardiovascular risk are not changed by increased whole-grain intake: the WHOLEheart study, a randomised, controlled dietary intervention.” Br J Nutr 2010;104(1):125–34. RCT, n=316, aged 18–65, BMI >25, habitual whole grain <30 g/d, randomised to control / 60 g WG per day for 16 wk / 60 g for 8 wk then 120 g for 8 wk. Compliance was good and confirmed. Result: “there were no significant differences in any markers of CVD risk between groups” — no change in BMI, body fat, waist, fasting lipids, glucose, insulin, or inflammatory/coagulation/endothelial markers. The authors conclude the message “may need to be clarified to consider the source of WG and/or other diet and lifestyle factors linked to the benefits… seen in observational studies.”
Rahmani S, Sadeghi O, Sadeghian M, Sadeghi N, Larijani B, Esmaillzadeh A. “The Effect of Whole-Grain Intake on Biomarkers of Subclinical Inflammation: A Comprehensive Meta-analysis of Randomized Controlled Trials.” Adv Nutr 2020;11(1):52–65. MA-RCT, 14 RCTs, 1,238 adults. Pooled results: CRP WMD −0.29 mg/L (95% CI −1.10 to 0.52) — null. IL-6 WMD −0.08 pg/mL (−0.27 to 0.11) — null. TNF-α WMD −0.06 pg/mL (−0.25 to 0.14) — null. PAI-1 — null. Benefits appeared only in subgroups with already-elevated CRP or in “unhealthy individuals.” Their own conclusion: “Unlike observational studies, we found no significant effect of whole-grain consumption on serum concentrations of inflammatory cytokines.” For a healthy adult with presumably normal CRP, this is the directly relevant subgroup and the answer is “no measurable anti-inflammatory effect.”
Vanegas SM, Meydani M, Barnett JB, et al. Am J Clin Nutr 2017;105(3):635–650. RCT, n=81 (49 men, 32 postmenopausal women, 40–65 y), 6-wk provided-food weight-maintenance diets, WG vs RG. Found: increased stool weight (p<0.0001) and frequency, increased Lachnospira, decreased Enterobacteriaceae, higher stool acetate and total SCFAs, higher terminal effector memory T cells, higher LPS-stimulated TNF-α ex vivo — but “no effect on other markers of cell-mediated immunity or systemic and gut inflammation.” Title word is “modest,” and that is accurate.
The one trial with a genuinely interesting result for a fat-loss goal: Karl JP, Meydani M, Barnett JB, et al. “Substituting whole grains for refined grains in a 6-wk randomized trial favorably affects energy-balance metrics in healthy men and postmenopausal women.” Am J Clin Nutr 2017;105(3):589–599. RCT, controlled feeding, n=81. WG diet supplied 207 ± 39 g whole grains and 40 ± 5 g fibre/day; RG diet 0 g whole grain and 21 ± 3 g fibre. Body weight held constant by design. Results: resting metabolic rate +43 ± 25 kcal/d (p=0.04), stool weight +76 ± 12 g/d, stool energy content +57 ± 17 kcal/d (p=0.003). Combined: a 92 kcal/d (95% CI 28–156) higher net daily energy loss on whole grains. Caveat the authors themselves add: excluding non-adherent participants made the RMR difference non-significant.
Reynolds’ own RCT pooling for whole grains: body weight −0.62 kg (95% CI −1.19 to −0.05) across 11 trials (n=498 vs 421), total cholesterol −0.09 mmol/L (−0.23 to +0.04, ns), systolic BP −1.01 mmHg (−2.46 to +0.44, ns), HbA1c SMD −0.54 (−1.28 to +0.20, ns).

Honest bottom line for this subject: the cohort signal for whole grains is large and consistent; the randomised signal is small (≈0.6 kg body weight, ≈90 kcal/day of energy-balance shift, no change in inflammation or lipids in healthy people). The gap between them is the size of the healthy-user effect plus the fact that a 4–16 week trial cannot reproduce a 20-year exposure. Whole grains are worth choosing — mainly for the fibre, the ~90 kcal/day faecal-energy + RMR effect, and the micronutrient density — not because they will lower the adult’s CRP or the adult’s blood pressure.

Detection strings: whole wheat, whole-wheat, whole grain, wholegrain, whole meal, wholemeal, whole durum, whole grain oats, rolled oats, steel cut oats, bran, wheat berries, wheat berry Refined-grain detection strings: enriched flour, enriched wheat flour, all-purpose flour, white flour, bleached flour, unbleached flour, semolina, durum flour, white bread, white bun, brioche, panko, breadcrumbs Ambiguity flags: wheat flour matches BOTH refined (“wheat flour”) and whole (“whole wheat flour”) — must anchor on the preceding word whole. multigrain and made with whole grain are marketing, not composition. bran also matches bran muffin (usually a refined-flour cake). oat flour in ultraprocessed bars ≠ intact oats.

White rice specifically

The T2D cohort signal

Hu EA, Pan A, Malik V, Sun Q. “White rice consumption and risk of type 2 diabetes: meta-analysis and systematic review.” BMJ 2012;344:e1454. MA-PC. 4 articles → 7 distinct cohort analyses; 13,284 incident T2D cases among 352,384 participants; follow-up 4–22 y.

Asian (Chinese and Japanese) populations, highest vs lowest: RR 1.55 (95% CI 1.20–2.01)
Western populations, highest vs lowest: RR 1.12 (0.94–1.33) — not statistically significant
P for interaction = 0.038
Whole-population dose-response: RR 1.11 (1.08–1.14) per serving/day
Context the paper itself supplies: Asian cohorts averaged 3–4 servings/day; Western cohorts averaged 1–2 servings/week.

Sun Q, Spiegelman D, van Dam RM, Holmes MD, Malik VS, Willett WC, Hu FB. “White rice, brown rice, and risk of type 2 diabetes in US men and women.” Arch Intern Med 2010;170(11):961–9. PC. 39,765 men + 157,463 women (HPFS, NHS, NHS II).

White rice ≥5 servings/wk vs <1/month: pooled RR 1.17 (1.02–1.36)
Brown rice ≥2 servings/wk vs <1/month: RR 0.89 (0.81–0.97)
Substituting 50 g/d cooked white rice with brown rice: 16% lower risk (9–21%)
Substituting the same white rice with whole grains as a group: 36% lower risk (30–42%) — i.e. brown rice is a mediocre whole grain by this cohort’s own estimate.

Bhavadharini B, Mohan V, Dehghan M, et al. “White Rice Intake and Incident Diabetes: A Study of 132,373 Participants in 21 Countries.” (PURE) Diabetes Care 2020;43(11):2643–2650. PC. 132,373 people, mean 9.5 y follow-up, 6,129 incident cases.

Overall ≥450 g/d cooked vs <150 g/d: HR 1.20 (1.02–1.40), p-trend 0.003
South Asia: HR 1.61 (1.13–2.30)
“Other regions” (SE Asia, Middle East, South America, North America, Europe, Africa): HR 1.41 (1.08–1.86)
China: HR 1.04 (0.77–1.40) — null, despite China having the highest rice intakes in the study.

That China null is the most informative single number in the rice literature. If white rice were straightforwardly diabetogenic by dose, the highest-dose population should show the strongest effect. It shows none. That pattern is much more consistent with rice being a marker for a dietary/socioeconomic pattern than with rice being a cause.

ABSOLUTE risk

Baseline: 3.85 T2D cases per 1,000 person-years (Muraki 2016, US health professionals, 15,362 cases / 3,988,007 PY).

Exposure	RR/HR	Cases/1,000 PY	Δ per 1,000 PY	Δ per 1,000 over 10 y	NNH (10 y)
White rice ≥5 svg/wk vs <1/mo (Sun 2010)	1.17	4.50	+0.65	+6.5	~154
Per +1 serving/day (Hu 2012)	1.11	4.27	+0.42	+4.2	~238
Swap 50 g/d white→brown (Sun 2010)	0.84	3.23	−0.62	−6.2	NNT ~161
Swap 50 g/d white→other whole grains	0.64	2.46	−1.39	−13.9	NNT ~72

Scaled to this subject. The adult is a young adult, lean, and trains at high volume — every one of those is a large protective factor and the cohort baseline is drawn from people averaging 50–65 y with BMI ~26. Dividing by a conservative factor of 3–5: eating white rice 4–5×/week costs the adult on the order of +1.3 to +2.2 extra T2D cases per 1,000 men like the adult per decade, i.e. roughly 1 in 450–750 over ten years, and that estimate assumes the association is causal, which the China null argues against.

Is the signal confounded?

Yes, substantially, and here is the specific case:

Total dietary pattern. In Western cohorts, high white-rice eaters are typically eating rice as part of a takeaway/convenience pattern; in Asian cohorts, high white-rice eaters are typically lower-income and less physically active in the modern urban samples. Neither of these is “rice.”
Rice as majority of energy. When rice supplies 50–70% of energy, “high rice intake” is functionally a proxy for low intake of everything else — low protein, low vegetables, low dairy, low fat. The exposure contrast is a whole-diet contrast dressed as a single-food contrast.
FFQ measurement error on a staple eaten 2–3×/day is severe; rice portion sizes are poorly captured.
Effect size vs noise floor. The Western RR of 1.12 (0.94–1.33) is not statistically significant and sits squarely in the band where residual confounding alone routinely produces spurious associations. Aune 2016 similarly found no clear association for white rice or total rice in the CVD/mortality analysis.
Contrary interventional evidence: Devlin BL, Parr EB, Radford BE, Hawley JA. Clin Nutr 2021;40(4):2200–2209. XO RCT, n=24 adults with T2D, four evening-meal conditions (boiled potato / roasted potato / cooled boiled potato / basmati rice as control), each meal 50% CHO. Result: all three potato meals produced lower nocturnal glucose AUC than the basmati rice control (p<0.001). Rice is not a metabolic free pass relative to potato.

Glycaemic index of actual rice types

The single most important fact: rice GI variance within “white rice” is larger than the difference between rice and most other staples. Sources:

Haxhari F, et al. “Endosperm structure and Glycemic Index of Japonica Italian rice varieties.” Front Plant Sci 2024;14:1303771. 25 Italian genotypes tested in vivo: GI range 49 to 92 vs glucose. Low-GI varieties (<55) had large starch granules and low endosperm porosity.
Rondanelli M, Ferrario RA, Barrile GC, et al. “The Glycemic Index of Indica and Japonica Subspecies Parboiled Rice Grown in Italy…” J Med Food 2023;26(6):422–427. XO, n=10 healthy adults, BMI 18.5–25, ISO method. All parboiled rices tested were LOW-GI: brown long-B parboiled 48.1 ± 6.4; ribe 52.0 ± 1.8; black 52.3 ± 7.6; long-B 52.4 ± 3.9 (flora) / 53.4 ± 5.1 (conventional); Roma 54.4 ± 7.9; arborio 54.4 ± 7.9; red 56.1 ± 7.0. Rice cakes by contrast: classic 83.3 ± 8.9, brown 102.2 ± 5.5.
Bayaga CLT, et al. Sci Rep 2026;16(1):20272. XO, n=10, ISO 26642:2010: black rice GI 49, red rice 69, white rice 71 (GL 12.25 / 17.75 / 17.75).
Chen Y, et al. Carbohydr Polym 2026;390:125776. MECH. Amylose content outperformed digestion kinetics as a GI predictor across 7 varieties spanning 8.1%–31.9% amylose. This is the real mechanism: high-amylose = low GI; waxy/sticky/low-amylose = high GI.

What that means for named supermarket varieties:

Basmati — high amylose, long slender grain, stays separate → consistently at the low end of white rices.
Parboiled / converted (Uncle Ben’s-style) — the parboiling step retrogrades starch in the husk before milling → verified low-GI (48–56), the single most reliable GI lever available in a white rice.
Jasmine — lower amylose, aromatic, sticky-ish → sits at the high end of white rices; the widely-quoted “jasmine GI ~109” figure is a real entry in the Sydney GI database but I could not primary-verify it in this session — treat as unverified recall, though the direction (jasmine high) is well supported by the amylose mechanism.
Sushi / short-grain / glutinous — lowest amylose → highest GI of the rices.
Brown rice — the bran slows things somewhat but brown rice is not reliably low-GI; the Rondanelli data show brown parboiled at 48 while ordinary brown rice generally lands in the 50s–60s.

Practical ranking for this subject (best→worst by glycaemic behaviour): parboiled > basmati > long-grain white ≈ brown > jasmine > short-grain/sushi. For a post-training meal the ranking arguably inverts — a high-GI rice is a legitimate glycogen-repletion tool.

Resistant starch from cooling and reheating — real, but heavily overstated

Sonia S, Witjaksono F, Ridwan R. “Effect of cooling of cooked white rice on resistant starch content and glycemic response.” Asia Pac J Clin Nutr 2015;24(4):620–5. XO RCT, n=15 healthy adults, randomised single-blind crossover. Measured RS:

Freshly cooked: 0.64 g RS/100 g
Cooled 10 h at room temp: 1.30 g/100 g
Cooled 24 h at 4 °C then reheated: 1.65 g/100 g
Glycaemic response, cooled+reheated vs fresh: 125 ± 50.1 vs 152 ± 48.3 mmol·min/L, p=0.047 → a ~18% lower iAUC.

Do the arithmetic on the “cooling cuts the calories” claim. The absolute RS gain is ~1.0 g per 100 g cooked rice. RS yields roughly 2 kcal/g via colonic SCFA rather than ~4 kcal/g, so the energy saving is ≈2 kcal per 100 g cooked rice — about 4 kcal on a 200 g serving. The viral claim of “50% fewer calories” (from the widely-circulated Sri Lankan coconut-oil-plus-cooling protocol) traces to a conference presentation, not a peer-reviewed paper; I could not locate a peer-reviewed publication of it and flag it as unverified/likely never published.

A related supporting MECH/XO finding: Krishnan V, et al. Int J Biol Macromol 2020;162:1668–1681 showed cooking fat type (rice bran oil > coconut oil) modestly increases RS via amylose–lipid (RS-V) complexing, maximum RS gain 2.26% — same order of magnitude, same conclusion: real chemistry, trivial calories.

Honest verdict: the ~18% glycaemic-response reduction is a real and reproducible finding worth using; the calorie-reduction claim is off by an order of magnitude.

Arsenic in rice — the numbers, and what they actually mean for the adult

Concentrations. Source: FDA, “Arsenic in Rice and Rice Products Risk Assessment Report” (May 2014, revised March 2016), Table 4.2. REG, HPLC-ICP-MS speciation. Inorganic arsenic (iAs), mean ppb (= µg/kg), uncooked:

Rice type	n	Mean iAs (ppb)	SEM	Range
Brown, long/medium/short grain, regular	98	160.5	4.1	34–249
Brown parboiled	1	191.3	—	—
Brown jasmine	2	132.5	18.5	114–151
Brown basmati	13	122.7	11.3	66–200
White parboiled	38	111.9	3.8	71–182
White long grain, regular	148	103.3	2.2	23–196
White medium grain	91	80.9	2.6	39–174
White short grain	23	78.9	3.5	52–102
White jasmine	11	78.4	6.6	34–110
White basmati	40	61.8	3.9	20–144
White instant/pre-cooked	14	57.6	7.5	31–134
Infant brown rice cereal	59	119.9	6.4	30–254
Infant white rice cereal	86	105.3	2.2	21–151

FDA’s own executive-summary rounding: 92 ppb in white rice, 154 ppb in brown rice.

Brown rice has ~55–67% more inorganic arsenic than white rice, because iAs concentrates in the bran that polishing removes. Independently corroborated by Scott CK & Wu F. “Arsenic content and exposure in brown rice compared to white rice in the United States.” Risk Anal 2025;45(8):2183–2196, which reports US means of iAs 0.093 (white) vs 0.138 (brown) and notes iAs is 33% of total As in US white rice vs 48% in US brown rice; rice bran itself is far higher again. (Their Table 4 unit label reads “µg/kg,” which cannot be right for values of 0.093–0.278 — those are plainly µg/g / mg/kg, i.e. 93 and 138 ppb, which matches FDA. I flag the unit typo rather than propagating it.)

By origin: US Southern-grown rice (Arkansas/Louisiana/Texas, on former cotton land treated with arsenical pesticides) runs higher than California rice, and imported basmati (India/Pakistan) and jasmine (Thailand) run lowest. The FDA table above does not break out by state, so I am reporting the origin ordering as directionally supported by the variety data (white basmati 61.8 vs US white long grain 103.3) but not state-level-verified in this session.

The benchmark. EFSA CONTAM Panel, “Update of the risk assessment of inorganic arsenic in food.” EFSA J 2024;22(1):e8488. REG.

Reference point: BMDL05 = 0.06 µg iAs/kg bw per day, derived from a skin-cancer study, BMR 5%.
iAs treated as a genotoxic carcinogen with epigenetic effects → margin-of-exposure approach.
EFSA’s finding for European adults: MOE 2 to 0.4 for mean consumers, 0.9 to 0.2 at the 95th percentile — “low… and as such raise a health concern despite the uncertainties.”
The older, much less stringent EFSA 2009 reference range was BMDL01 0.3–8 µg/kg bw/day; the 2021 exposure assessment (EFSA J 2021;19(1):e06380) found mean adult LB exposures generally below that older range.

FDA action level for inorganic arsenic in infant rice cereal: 100 ppb (Guidance for Industry, finalised August 2020).

Now the adult’s actual exposure. Arithmetic, 80 kg male:

Assume a realistic meal-delivery serving of 75 g dry rice (~200 g cooked, ~1 cup), eaten 4×/week.

Rice	iAs (µg/g)	µg per serving	µg/week	µg/day	µg/kg bw/day	MOE vs 0.06
White basmati	0.0618	4.6	18.5	2.65	0.033	1.81
White long grain	0.1033	7.7	31.0	4.43	0.055	1.08
White parboiled	0.1119	8.4	33.6	4.80	0.060	1.00
Brown regular	0.1605	12.0	48.2	6.88	0.086	0.70

At a heavier 150 g dry × 4/week, double all of the above: white long grain → 0.111 µg/kg bw/day, MOE 0.54; brown → 0.172, MOE 0.35.

Cooking method changes this materially. Menon M, Nicholls A, Smalley A, Rhodes E. “A comparison of the effects of two cooking methods on arsenic species and nutrient elements in rice.” Sci Total Environ 2024;914:169653. Excess-water (EW) and parboiled-and-absorbed (PBA) methods both removed 54–58% of iAs from white and brown rice; EW was better for already-parboiled rice (~50% vs ~39%). Applying 56% removal to white long grain at 75 g × 4/wk: 0.024 µg/kg bw/day, MOE ≈ 2.5. (Costs: measurable losses of K, Fe, Cu, Mo — Menon 2024 — and of enriched iron/folate/thiamin/niacin — FDA 2016.)

Two regulators, same data, opposite emotional register — and both are right.

EFSA’s MOE framing says the entire European adult population is already at MOE < 1–2 for inorganic arsenic across the whole diet. By that framework any rice intake is “of concern.” But note what MOE < 1 means: exposure is at or above a dose associated with a 5% relative increase over background skin-cancer incidence, in populations (Bangladesh, Taiwan, Chile) exposed via drinking water at concentrations orders of magnitude above anything in a US meal kit. It is a hazard-ranking tool, not a personal risk estimate.
FDA’s quantitative risk assessment puts a number on it: lifetime lung + bladder cancer risk attributable to all rice and rice products at current US average intake is 39 cases per million people (10 bladder + 29 lung) — against a background of 90,000 lung+bladder cancer cases per million over a lifetime. That is an absolute lifetime risk increase of 0.0039 percentage points, ~1 in 25,600. At a full one serving per day, FDA models 74–184 cases per million depending on rice type = 1 in 5,400 to 1 in 13,500 lifetime.

Is there evidence of measurable harm at realistic Western rice intakes? No. The FDA number is a linear-extrapolation model output, not an observed excess. I found no epidemiological study demonstrating harm in a Western population from dietary rice arsenic at these levels. The correct label is precautionary, not demonstrated.

Actionable, in priority order: (1) prefer white basmati — it is simultaneously the lowest-arsenic and among the lowest-GI rices, which is a rare free lunch; (2) cook in excess water and drain if convenient (~56% reduction); (3) do not switch to brown rice for health reasons and then eat it daily — you trade ~1.2 g extra fibre per 100 g cooked for ~55% more inorganic arsenic; (4) rice cakes and brown-rice syrup are the worst-value items in this class (rice cake GI 83–102, and brown rice syrup is concentrated rice solids).

Detection strings: white rice, jasmine rice, basmati, basmati rice, parboiled rice, converted rice, long grain rice, long-grain rice, sushi rice, short grain rice, arborio, calrose, brown rice, brown jasmine, rice pilaf, coconut rice, sticky rice, glutinous rice, rice cake, puffed rice, brown rice syrup Ambiguity flags — IMPORTANT: a bare rice substring matches rice vinegar (condiment, ~0 CHO impact), rice wine, rice flour (refined, high-GI thickener), rice noodles, rice paper, brown rice syrup (added sugar, not a grain), wild rice (not rice at all — it’s Zizania, a different genus, and nutritionally much better), and riced cauliflower (a vegetable). Match wild rice and rice vinegar and brown rice syrup and riced cauliflower FIRST and exclude them before matching rice.

Potatoes vs sweet potatoes

Is the potato reputation deserved? Mostly no — the reputation belongs to the fryer.

Mozaffarian D, Hao T, Rimm EB, Willett WC, Hu FB. “Changes in Diet and Lifestyle and Long-Term Weight Gain in Women and Men.” NEJM 2011;364(25):2392–2404. PC. 120,877 US women and men, free of chronic disease and not obese at baseline; three cohorts; 4-year intervals, follow-up 1986–2006 / 1991–2003 / 1986–2006.

This is the single most-cited “potatoes make you fat” paper, and the abstract materially undersells the split. The abstract lists “potato chips (1.69 lb)” and “potatoes (1.28 lb)”. Table 2 breaks it out (4-year weight change per one additional daily serving):

Item	4-y weight change	95% CI
French fries	+3.35 lb	2.29 to 4.42
Potato chips	+1.69 lb	1.30 to 2.09
Boiled, baked, or mashed potatoes	+0.57 lb	0.26 to 0.89
Refined grains	+0.39 lb	0.21 to 0.58
Sugar-sweetened beverages	+1.00 lb	0.83 to 1.17
100% fruit juice	+0.31 lb	0.14 to 0.47
Whole grains	−0.37 lb	−0.48 to −0.25

So ~85% of the “potato” weight-gain signal is fries and chips. Plain potatoes at +0.57 lb per 4 years per daily additional serving are barely distinguishable from refined grains at +0.39 lb, and this is an observational within-person change model on self-reported weight with all the residual confounding that implies.

Potatoes and T2D

Muraki I, Rimm EB, Willett WC, Manson JE, Hu FB, Sun Q. “Potato Consumption and Risk of Type 2 Diabetes: Results From Three Prospective Cohort Studies.” Diabetes Care 2016;39(3):376–384. PC. 70,773 NHS women + 87,739 NHS II women + 40,669 HPFS men; 3,988,007 person-years; 15,362 incident T2D cases.

Total potatoes ≥7 svg/wk vs <1 svg/wk: HR 1.33 (1.17–1.52); 2–4 svg/wk: HR 1.07 (0.97–1.18) — not significant
Baked/boiled/mashed, per 3 svg/wk: HR 1.04 (1.01–1.08)
French fries, per 3 svg/wk: HR 1.19 (1.13–1.25)
Replacing 3 svg/wk of total potatoes with whole grains: HR 0.88 (0.84–0.91)

Same pattern. The plain-potato hazard per 3 servings/week is 4% — inside the noise floor. The fries hazard is 19% and reproducible.

Absolute risk (baseline 3.85/1,000 PY, from this same paper):

Exposure	HR	Δ per 1,000 PY	Δ per 1,000 over 10 y
Baked/boiled/mashed, +3 svg/wk	1.04	+0.15	+1.5 (1 extra case per ~650 people/decade)
French fries, +3 svg/wk	1.19	+0.73	+7.3 (1 per ~137/decade)
Total potatoes ≥7/wk vs <1/wk	1.33	+1.27	+12.7
Swap 3 svg/wk potatoes → whole grains	0.88	−0.46	−4.6

Scaled to a lean adult (÷3 to ÷5): plain potatoes three times a week cost the adult roughly 0.3–0.5 extra cases per 1,000 men per decade — i.e. essentially nothing. Fries three times a week cost ~1.5–2.4 per 1,000 per decade.

Preparation-separating re-analyses (2023–2025). I searched for the 2023–2025 papers that reportedly nullify the potato–T2D association after adjusting for preparation and could not verify a specific one in this session. What I can verify is that (a) Muraki 2016 itself already separates preparation and finds the plain-potato HR to be 1.04, and (b) the strongest interventional evidence points the opposite way from the cohorts (Devlin 2021, below). I am explicitly flagging “the 2023–2025 potato re-analyses” as unverified.

Contrary RCT evidence. Devlin BL, Parr EB, Radford BE, Hawley JA. Clin Nutr 2021;40(4):2200–2209. XO RCT, n=24 adults with T2D (age 58.3 ± 9.3, BMI 31.7 ± 6.8), randomised crossover, standardised breakfast + lunch, dinner = 40% of daily energy as boiled potato / roasted potato / cooled boiled potato / basmati rice control, all meals 50% CHO / 30% fat / 20% protein, CGM overnight. Results: no difference in postprandial glucose iAUC between rice and any potato condition; cooled potato produced higher postprandial insulin than rice (p=0.003); and all three potato meals produced lower nocturnal glucose AUC than rice (p<0.001). Authors’ conclusion: potato meals “can be considered suitable for individuals with T2DM when consumed as part of a mixed-evening meal.” If potato is metabolically fine for people who already have T2D, it is fine for a adult.

Satiety — potatoes are the highest-scoring food ever measured

Holt SHA, Miller JC, Petocz P, Farmakalidis E. “A satiety index of common foods.” Eur J Clin Nutr 1995;49(9):675–690. Design verified: isoenergetic 1000 kJ (240 kcal) servings of 38 foods, fed to groups of 11–13 subjects, satiety rated every 15 min over 120 min, then ad-libitum intake; index = AUC relative to white bread = 100.

Boiled potatoes: 323 ± 51% — the highest score of any food tested, sevenfold the croissant (47 ± 17%).
SI correlated positively with serving weight (r=0.66), protein, fibre and water content (r=0.64) and negatively with palatability (r=−0.64).
Other carbohydrate values from the published table (widely reproduced; I verified the design and the potato/croissant figures from the abstract, and the remaining values from reproductions of the paper’s table rather than the paper itself — treat these as secondary-sourced): porridge/oatmeal 209, wholemeal pasta 188, baked beans 168, wholemeal bread 157, popcorn 154, white rice 138, lentils 133, brown rice 132, white pasta 119, cornflakes 118, white bread 100, croissant 47.

Caveats that matter: n = 11–13 per food, a single 2-hour window, never replicated at scale, and the negative palatability correlation is a warning that SI partly measures “how boring is this food,” not just physiology. It is nonetheless the best available comparative satiety data and it consistently ranks potato first — which for a man in an energy deficit trying to preserve lean mass is a genuinely useful property.

Potassium — directly relevant because the adult is cutting sodium

USDA FoodData Central, SR Legacy COMP (values per 100 g, verified from the SR Legacy dataset directly):

Food	kcal	Protein	Fibre	Potassium	Sodium	Beta-carotene	Vit A RAE
Potato, baked, flesh + skin, no salt (170093)	93	2.5 g	2.2 g	535 mg	10 mg	6 µg	1 µg
Sweet potato, baked in skin, flesh, no salt (168483)	~90	2.01 g	3.3 g	475 mg	36 mg	11,509 µg	961 µg
Sweet potato, boiled, no skin (168484)	~76	1.37 g	2.5 g	230 mg	27 mg	9,444 µg	787 µg
White rice, long grain, cooked (168878)	130	2.69 g	0.4 g	35 mg	1 mg	0	0

The white potato beats the sweet potato on potassium (535 vs 475 mg/100 g) and beats white rice by 15×. A 250 g baked potato delivers ~1,340 mg potassium — about 38% of the WHO 3,510 mg/day potassium guideline in one item.

How much is that worth to the adult? Less than the wellness framing suggests, and I want to be honest about it. Aburto NJ, Hanson S, Gutierrez H, Hooper L, Elliott P, Cappuccio FP. “Effect of increased potassium intake on cardiovascular risk factors and disease: systematic review and meta-analyses.” BMJ 2013;346:f1378. MA-RCT (22 RCTs, 1,606 participants) + MA-PC (11 cohorts, 127,038). Increased potassium reduced systolic BP by 3.49 mmHg (1.82–5.15) and diastolic by 1.96 (0.86–3.06) — but the abstract states this was “an effect seen in people with hypertension but not in those without.” The adult is a normotensive adult. The potassium is still worth having (sodium–potassium ratio, muscle function, cramp prevention under high sweat losses) but the blood-pressure benefit specifically is unlikely to materialise for the adult. Do not score potato as a blood-pressure intervention.

Glycaemic index and the cooling effect

Atkinson 2021 MA states plainly: “Most varieties of potato were high-GI foods, but specific low-GI varieties have now been identified.” GI varies enormously by variety and preparation — roughly: instant mash and baked russet at the top (often 80–95), boiled floury russet high, boiled waxy/red/new potatoes distinctly lower, and cooled potato lower again. I did not obtain per-variety GI numbers from a primary table in this session; treat any specific per-variety GI number as unverified.

The cooling/RS effect, quantified. Nolte Fong JV, Miketinas D, Moore LW, et al. Nutrients 2022;14(2):268. XO RCT, n=30 overweight women, 250 g potato hot vs cold: hot = 9.2 g RS, cold = 13.7 g RS — i.e. cooling adds ~4.5 g RS per 250 g serving, worth roughly 9 kcal of reduced glycaemic energy plus a colonic SCFA yield. 70% of participants had a favourable postprandial glucose response to the cold potato. This is a bigger absolute RS gain than in rice (~1 g/100 g), but still a side effect, not a transformation. And note Devlin 2021 found cooled potato produced the highest insulin response of the potato conditions — so “cool your potatoes” is not an unambiguous win.

Glycoalkaloids — a real toxin, not a real-world problem at normal intake

EFSA CONTAM Panel. “Risk assessment of glycoalkaloids in feed and food, in particular in potatoes and potato-derived products.” EFSA J 2020;18(8):e06222. REG.

Acute reference point: LOAEL 1 mg total potato glycoalkaloids/kg bw per day for GI symptoms (nausea, vomiting, diarrhoea).
“In humans, no evidence of health problems associated with repeated or long-term intake of GAs via potatoes has been identified.” No chronic reference point could be derived.
Occurrence: whole tubers 10–150 mg/kg fresh weight; peel 300–640 mg/kg vs flesh 12–100 mg/kg (3–10× concentrated in peel); sprouts 2,000–7,300 mg/kg.
Guidance levels: the historical “200 mg TGA/kg unpeeled uncooked potato is safe” figure; Germany’s BfR now recommends ≤100 mg/kg.
Adult exposure: mean UB 23.3 µg/kg bw/day, P95 78.3 µg/kg bw/day → MOE ≈ 43 at the mean and ≈ 13 at P95 against the 1 mg/kg LOAEL. EFSA concludes a health concern for adults only in the surveys with the highest P95 exposures; the concern is concentrated in young children.

Verdict: for an 80 kg adult, the acute LOAEL is ~80 mg total glycoalkaloids — roughly 0.8–8 kg of normal potato. Not a practical concern. The genuine rule is behavioural, not dietary: discard green, sprouted, or bitter potatoes, and peel where there is greening.

Sweet potato — better, but not for the reasons people say

Vitamin A is the real advantage and it is enormous: 961 µg RAE per 100 g baked (>100% of the 900 µg/day RDA for adult men) vs 1 µg for white potato. 11,509 µg beta-carotene.
Fibre: 3.3 vs 2.2 g/100 g — a 50% relative advantage but only 1.1 g in absolute terms.
Potassium: sweet potato loses (475 vs 535 mg/100 g), and boiled without skin loses badly (230 mg).
GI: the wellness claim that sweet potato is “low GI” is preparation-dependent and frequently wrong. Baked sweet potato is high-GI; boiled sweet potato is markedly lower (the mechanism is the same starch-gelatinisation story as white potato). Note also that the USDA data show boiling-without-skin costs you half the potassium and 25% of the beta-carotene.
Satiety: no equivalent Holt-style data for sweet potato; unverified.

Verdict: sweet potato is modestly better than white potato, driven almost entirely by vitamin A. White potato is the better potassium source and the better-documented satiety food. For a man eating a varied [calorie target removed] diet who is not vitamin-A deficient, the practical difference between them is small and both are good.

French fries and potato chips as their own class

Weight: +3.35 lb per 4 y per daily serving (fries), +1.69 lb (chips) — Mozaffarian 2011 PC.
T2D: HR 1.19 per 3 servings/week — Muraki 2016 PC.
Mortality: Veronese N, Stubbs B, Noale M, et al. “Fried potato consumption is associated with elevated mortality: an 8-y longitudinal cohort study.” Am J Clin Nutr 2017;106(1):162–167. PC, n=4,440 (Osteoarthritis Initiative, age 45–79, 8 y follow-up, only 236 deaths). Total potato consumption: HR 1.11 (0.65–1.91) — null. Fried potatoes 2–3×/wk: HR 1.95 (1.11–3.41); ≥3×/wk: HR 2.26 (1.15–4.47); unfried potatoes: no association. Heavily caveat this one: 236 events, subgroup analysis, wide CIs, and it drew three published critical letters (AJCN 2018;107(5):847–850). A doubling of mortality from fries three times a week is not a credible effect size; the direction is probably right, the magnitude is not.
Sodium: the reason fries matter most for this subject. Restaurant/meal-kit fries routinely carry 200–400 mg sodium per serving on top of whatever else is in the meal.
Acrylamide — and here the popular belief is wrong. EFSA 2015 opinion (EFSA J 2015;13(6):4104) set a BMDL10 of 0.17 mg/kg bw/day for neoplastic effects and regards MOE < 10,000 as of concern for genotoxic carcinogens; typical dietary exposures give MOEs of roughly 90–425, i.e. “of concern” by that rule. (I could not fetch the EFSA 2015 opinion directly — 403 — so the 0.17 figure is taken from peer-reviewed papers citing it: Zhu B, et al. *Food Chem 2021;352:129438 cites “BMDL10 of carcinogenicity at 0.17 mg/kg”; Lee S & Kim HJ, Int J Environ Res Public Health 2020;17(20):7619 cites BMDL10 0.18 and 0.31 mg/kg bw/day. Flagged as secondary-verified.)* Measured levels (Lee & Kim 2020): potato crisps 546 µg/kg, French fries 372 µg/kg, coffee 353 µg/kg. BUT — Filippini T, Halldorsson TI, Capitão C, et al. “Dietary Acrylamide Exposure and Risk of Site-Specific Cancer: A Systematic Review and Dose-Response Meta-Analysis of Epidemiological Studies.” Front Nutr 2022;9:875607. MA-PC. 16 studies, 31 papers, 1,151,189 participants, 48,175 incident cancers, median follow-up 14.9 y. Result: “Pooled analysis showed no association between the highest vs. lowest dietary acrylamide exposure and each site-specific cancer investigated, with no evidence of thresholds in the dose-response meta-analysis.” Only lung cancer in smokers showed an association. Conclusion: acrylamide is not the reason to avoid fries. Energy density, sodium, and the eating context are.

Detection strings — plain potato: potato, potatoes, russet, yukon gold, red potato, baby potato, fingerling, new potatoes, mashed potato, roasted potato, boiled potato, smashed potato, potato wedges Detection strings — sweet potato: sweet potato, sweetpotato, sweet potatoes, yam, garnet yam, japanese sweet potato, kumara Detection strings — fried potato class: french fries, fries, steak fries, curly fries, tater tots, hash browns, potato chips, crisps, home fries, potato skins Ambiguity flags: potato matches sweet potato and potato starch (a refined thickener) — match sweet potato and potato starch/potato flour first and exclude. yam in US usage almost always means orange sweet potato, but true yam (Dioscorea) is a different food with far less beta-carotene. hash browns and home fries should score with the fried class, not the plain class. mashed potato is plain unless the ingredient list also shows butter/cream in quantity — consider a separate modifier.

Quinoa, farro, barley, oats, bulgur, freekeh, buckwheat, millet, wild rice

USDA composition, per 100 g cooked (SR Legacy, verified directly from the dataset)

Grain	kcal	Protein	Fibre	Potassium	Magnesium	Lysine	Lysine mg/g protein
White rice, long grain	130	2.69 g	0.4 g	35 mg	12 mg	0.097 g	36
Brown rice, long grain	123	2.74 g	1.6 g	86 mg	39 mg	0.099 g	36
Quinoa	120	4.40 g	2.8 g	172 mg	64 mg	0.239 g	54
Barley, pearled	~123	2.26 g	3.8 g	93 mg	22 mg	0.084 g	37
Bulgur	83	3.08 g	4.5 g	68 mg	32 mg	0.085 g	28
Oats, cooked (porridge)	71	2.54 g	1.7 g	70 mg	27 mg	0.135 g	53
Buckwheat groats, roasted	~92	3.38 g	2.7 g	88 mg	51 mg	0.172 g	51
Millet	119	3.51 g	1.3 g	62 mg	44 mg	0.067 g	19
Wild rice	101	3.99 g	1.8 g	101 mg	32 mg	0.170 g	43
Spelt	127	5.50 g	3.9 g	143 mg	49 mg	—	—
(Lentils, for scale)	116	9.02 g	7.9 g	369 mg	36 mg	0.630 g	70

Two things jump out. First: bulgur and pearled barley beat every other grain here on fibre per 100 g cooked — bulgur has 11× the fibre of white rice. Second: millet is nutritionally unimpressive (1.3 g fibre, lysine 19 mg/g protein) and does not deserve its health-food positioning; it is a fine gluten-free starch, nothing more.

Beta-glucan: barley and oats are the only grains in this list with drug-like RCT evidence

Whitehead A, Beck EJ, Tosh S, Wolever TM. “Cholesterol-lowering effects of oat β-glucan: a meta-analysis of randomized controlled trials.” Am J Clin Nutr 2014;100(6):1413–21. MA-RCT, 28 RCTs.

≥3 g/day oat beta-glucan (OBG) vs control: LDL −0.25 mmol/L (95% CI 0.20–0.30) = −9.7 mg/dL; total cholesterol −0.30 mmol/L = −11.6 mg/dL. Both p<0.0001.
No effect on HDL or triglycerides. No dose effect across 3.0–12.4 g/d, no duration effect across 2–12 wk.
Crucial for this subject: “LDL cholesterol lowering was significantly greater with higher baseline LDL cholesterol” and greater in people with diabetes.

Ho HV, Sievenpiper JL, Zurbau A, et al. “The effect of oat β-glucan on LDL-cholesterol, non-HDL-cholesterol and apoB for CVD risk reduction: a systematic review and meta-analysis of randomised-controlled trials.” Br J Nutr 2016;116(8):1369–1382. MA-RCT, 58 trials, n=3,974. Median dose 3.5 g/d: LDL −0.19 mmol/L (−0.23, −0.14), non-HDL −0.20, apoB −0.03 g/L. Heterogeneity I² 79–99%.

Ho HV, Sievenpiper JL, Zurbau A, et al. “…the effect of barley β-glucan on LDL-C, non-HDL-C and apoB…” Eur J Clin Nutr 2016;70(11):1239–1245 (erratum 70(11):1340). MA-RCT, 14 trials, n=615. Median 6.5–6.9 g/d barley beta-glucan: LDL −0.25 mmol/L (−0.30, −0.20), non-HDL −0.31 mmol/L; no significant apoB change.

The authorised claim thresholds — both are 3 g/day:

EFSA NDA Panel, EFSA J 2010;8(12):1885. REG. Verified verbatim: “The Panel concludes that a cause and effect relationship has been established between the consumption of oat beta-glucan and lowering of blood LDL-cholesterol concentrations… in order to bear the claim, foods should provide at least 3 g of oat beta-glucan per day.”
FDA: 21 CFR 101.81 authorises the soluble-fibre/CHD claim on the same ≥3 g/day beta-glucan basis (confirmed by Whitehead 2014’s own framing: “Health claims… approved by food standards agencies worldwide, are based on a diet containing ≥3 g/d of oat β-glucan”). I could not fetch eCFR directly (redirect-blocked); the per-RACC sub-thresholds are unverified in this session.

How much food is 3 g of beta-glucan? Oats run roughly 3–5 g beta-glucan per 100 g dry, so ~60–100 g dry oats (about 1 to 1¼ cups dry / 2 cups cooked). Barley is similar or higher. I could not verify a primary compositional figure for oat beta-glucan percentage in this session — the USDA SR Legacy dataset has a beta-glucan nutrient code but no populated values for oats or barley. Treat the 3–5% figure as unverified recall. The practical upshot is unchanged: a normal single portion of oatmeal or a barley side dish delivers meaningfully less than 3 g; you need a real, daily, oat-or-barley-centric habit to hit the claim threshold.

Does it matter for someone with normal lipids? Partly. Two honest points:

Whitehead 2014 found the effect is baseline-dependent — larger when LDL is high. A normolipidaemic adult will capture less than 9.7 mg/dL.
But LDL-lowering benefit is roughly proportional and cumulative over exposure time. Gencer B, Marston NA, Im K, et al. Lancet 2020;396(10263):1637–1643. MA-RCT, 244,090 patients, 29 trials found major vascular events RR 0.85 per 1 mmol/L LDL reduction in patients under 75 (0.74 in those ≥75). Applying that exchange rate to a sustained −0.25 mmol/L gives roughly 0.85^0.25 ≈ 0.96, i.e. a ~4% relative reduction on a ~5-year trial horizon — and lifelong exposure from early adulthood plausibly delivers more than a 5-year trial does. Small, real, and cheap to obtain.

Quinoa: is the “complete protein” claim real?

Yes, more or less — and the USDA data let us check it precisely rather than repeat the marketing.

“Complete protein” is not a technical standard. The testable version is: does quinoa meet the FAO/WHO 2007 adult indispensable amino acid scoring pattern, in which lysine — the amino acid all cereals are short of — must be ≥45 mg per g protein?

From the USDA SR Legacy table above:

Quinoa, cooked: 0.239 g lysine / 4.40 g protein = 54 mg/g protein — clears 45.
White rice: 36 mg/g — fails.
Brown rice: 36 mg/g — fails.
Bulgur (wheat): 28 mg/g — fails badly (this is the classic wheat lysine deficit).
Millet: 19 mg/g — fails badly.
Buckwheat: 51 mg/g — clears. Oats: 53 mg/g — clears. Wild rice: 43 mg/g — borderline.
Lentils: 70 mg/g — clears easily.

So the claim survives contact with the data: quinoa (and buckwheat, and oats) are genuinely not lysine-limited, which is unusual among grains. Three honest qualifications:

Quinoa is a pseudocereal, not a grass — so “the only complete grain” is trivially true by taxonomy-shopping.
Protein quantity is the binding constraint for an active adult, not quality: 4.4 g protein per 100 g cooked quinoa. A 200 g serving is ~8.8 g protein. That is a rounding error against the adult’s likely 160–200 g/day target. Quinoa is not a protein source for the adult; it is a better-than-average starch.
Digestibility-corrected scores are moderate, not animal-protein-tier. Quevedo-Olaya JL, Schmiele M, Correa MJ. Foods 2025;14(17):2987 (review) reports Andean grains at 13–18 g protein per 100 g dry providing “adequate levels of lysine, methionine, and threonine, meeting FAO requirements for adult nutrition,” while Mukunzi Y & Aryee ANA. Foods 2025;14(18):3237 found for tri-colour quinoa flour that “lysine remained the limiting amino acid” (relative to its own profile). I could not verify a single authoritative DIAAS value for cooked quinoa in this session — flagged as unverified.

Other quinoa notes: highest magnesium of the common grains (64 mg/100 g cooked); contains saponins on the seed coat (bitter, mildly irritant, removed by the rinsing/pre-washing that virtually all retail quinoa now receives) — not a health concern.

Farro / spelt / emmer

Farro is a marketing term covering einkorn, emmer, and spelt — all hulled ancient wheats. They are wheat. They contain gluten, their fibre and mineral content is modestly better than modern refined wheat, and USDA cooked spelt is genuinely good: 5.5 g protein and 3.9 g fibre per 100 g cooked — the highest protein of any grain in the table above. There is essentially no RCT evidence specific to farro/spelt for any clinical outcome. Score it on composition and on the fact that it is eaten as an intact grain, which is the processing property that actually predicts glycaemic response (Jia/Jenkins 2026).

Bulgur and freekeh

Bulgur = parboiled, dried, cracked wheat. Parboiling before cracking gives it the same retrogradation advantage parboiled rice has, plus 4.5 g fibre/100 g cooked (11× white rice) at only 83 kcal/100 g — the lowest energy density of any grain in the table. Genuinely one of the best starch swaps available. Its weakness is protein quality (lysine 28 mg/g).
Freekeh = durum wheat harvested green and roasted. Compositionally similar to bulgur with somewhat higher protein. Essentially zero human clinical trial evidence — score on composition and intactness, with low confidence.

Buckwheat, millet, wild rice

Buckwheat (Fagopyrum, not wheat, gluten-free): 51 mg lysine/g protein, highest magnesium after quinoa (51 mg/100 g), contains rutin. Good.
Millet: the weakest grain in this section — 1.3 g fibre/100 g cooked and lysine 19 mg/g. Frequently high-GI. Fine, unremarkable.
Wild rice (Zizania, an aquatic grass, not Oryza): highest protein of the true “rices” (3.99 g/100 g cooked), lysine 43 mg/g, and — importantly for the arsenic section — it is not rice and does not carry rice’s arsenic burden. A strictly better choice than brown rice on every axis except cost and cook time.

Which of these are meaningfully better than white rice, and by how much?

Per 100 g cooked, versus white rice’s 0.4 g fibre / 35 mg K / 12 mg Mg / lysine 36:

Grain	Fibre gain	K gain	Mg gain	Verdict
Bulgur	+4.1 g (11×)	+33 mg	+20 mg	Biggest fibre upgrade available
Pearled barley	+3.4 g (9.5×)	+58 mg	+10 mg	Plus beta-glucan LDL effect
Spelt/farro	+3.5 g	+108 mg	+37 mg	Plus +2.8 g protein
Quinoa	+2.4 g (7×)	+137 mg	+52 mg	Plus complete AA profile
Buckwheat	+2.3 g	+53 mg	+39 mg	Gluten-free, good AA profile
Wild rice	+1.4 g	+66 mg	+20 mg	Plus +1.3 g protein, no arsenic
Brown rice	+1.2 g (4×)	+51 mg	+27 mg	and +55% inorganic arsenic
Millet	+0.9 g	+27 mg	+32 mg	Marginal

The headline: brown rice is the weakest of the whole-grain upgrades and the only one with a countervailing contaminant cost. Bulgur, barley, quinoa and farro are all better swaps than brown rice, and Sun 2010’s own substitution analysis agrees (brown rice −16% T2D risk vs whole grains generally −36%).

Detection strings: quinoa, red quinoa, tricolor quinoa, farro, spelt, emmer, einkorn, barley, pearl barley, pearled barley, hulled barley, oats, rolled oats, steel cut oats, steel-cut oats, oatmeal, oat groats, bulgur, bulghur, cracked wheat, freekeh, frikeh, buckwheat, kasha, soba, millet, wild rice, amaranth, teff, sorghum, kamut, khorasan Ambiguity flags: oat matches oat milk (a beverage, mostly maltose-rich hydrolysed starch — score separately, NOT as a whole grain) and oat flour in ultraprocessed bars. soba noodles are frequently 20–80% wheat flour despite the buckwheat name — do not assume. barley matches barley malt/malted barley (an added sugar) and barley malt syrup. wild rice blends are often 90% white rice by weight. quinoa flour and puffed quinoa are not intact quinoa.

Legumes and beans

Is this the strongest positive-evidence food class in the whole diet? Close to it — and uniquely, the RCT evidence is stronger than the observational evidence, which is the opposite of every other class in this section.

That inversion is the argument. For whole grains, the cohorts scream and the trials whisper. For pulses, the trials deliver reproducible effects on hard biomarkers (LDL, apoB, blood pressure, fasting glucose, fat mass) while the cohorts are graded “low to very low.” Effects that survive randomisation are the ones worth acting on.

The RCT evidence

Lipids — Back S, Yang S, Rossi A, … Sievenpiper JL, Chiavaroli L. “Effect of Different Types of Whole Dietary Pulses on Established Therapeutic Lipid Targets for Cardiovascular Risk Reduction: An Updated Systematic Review and Dose-Response Meta-Analysis of Randomized Controlled Trials.” J Am Heart Assoc 2026;15(10):e046659. MA-RCT. 38 trials, 52 comparisons, n=2,095, median 6 wk, median dose 130 g/d (0.5–0.67 cup/d).

LDL-C −0.14 mmol/L (95% CI −0.19 to −0.08) = −5.4 mg/dL
non-HDL-C −0.22 mmol/L (−0.30 to −0.14) = −8.5 mg/dL
apoB −0.08 g/L (−0.13 to −0.03)
HDL −0.03 mmol/L (trivial reduction)
Beans specifically show a linear dose-response up to 1 cup/day: −0.25 mmol/L LDL per 0.5 cup (−0.48 to −0.02) and −0.45 mmol/L non-HDL per 0.5 cup.
GRADE moderate-to-high for all outcomes except apoB. That “moderate-to-high” grade is rare in nutrition and is the single most important sentence in this subsection.

Consistent predecessors: Ha V, Sievenpiper JL, de Souza RJ, et al. CMAJ 2014;186(8):E252–62. MA-RCT, 26 RCTs, n=1,037, median 130 g/d: LDL −0.17 mmol/L (−0.25 to −0.09) = −6.6 mg/dL (apoB and non-HDL not significant in that older, smaller dataset). And Shaygantabar M, et al. “Effect of Non-Soy Legumes on Lipid Profile…” Nutr Rev 2026;84(10):1999–2013. MA-RCT, 31 RCTs: TC −5.16 mg/dL (−7.19, −3.13); LDL −3.48 mg/dL (−5.32, −1.53); TG null overall but significant in trials <8 wk and in participants with BMI <30 (−2.91 mg/dL) — i.e. the effect is present in non-obese people like this subject.

Translating to outcomes: at Gencer 2020’s exchange rate (RR 0.85 per 1 mmol/L LDL in under-75s), a sustained −0.14 to −0.17 mmol/L is a ~2–3% relative reduction in major vascular events over a trial-length horizon. Sustained from early adulthood the cumulative-exposure effect is larger. Small per year; large over a lifetime; and it comes attached to fibre, potassium, protein and satiety rather than to a pill.

Blood pressure — Jayalath VH, de Souza RJ, Sievenpiper JL, et al. “Effect of dietary pulses on blood pressure: a systematic review and meta-analysis of controlled feeding trials.” Am J Hypertens 2014;27(1):56–64. MA-RCT, 8 isocaloric trials, n=554 (with and without hypertension): systolic −2.25 mmHg (−4.22 to −0.28), p=0.03; mean arterial −0.75 mmHg; diastolic −0.71 (ns). Heterogeneity significant for all outcomes. Note this effect was seen in people with and without hypertension, unlike potassium supplementation — relevant to a normotensive subject who is cutting sodium.

Glycaemic control — Sievenpiper JL, Kendall CW, Esfahani A, et al. “Effect of non-oil-seed pulses on glycaemic control: a systematic review and meta-analysis of randomised controlled experimental trials in people with and without diabetes.” Diabetologia 2009;52(8):1479–95. MA-RCT, 41 trials.

Pulses alone (11 trials): fasting blood glucose SMD −0.82 (−1.36 to −0.27), insulin SMD −0.49 (−0.93 to −0.04)
Pulses in low-GI diets (19 trials): glycated proteins (HbA1c/fructosamine) SMD −0.28 (−0.42 to −0.14)
Pulses in high-fibre diets (11 trials): FBG −0.32, glycated proteins −0.27
Heterogeneity high and unexplained for most outcomes — flagged by the authors.

Body composition — Rahmanian R, Shaygantabar M, Hekmatdoost A, et al. “Effects of Non-Soy Legumes on Body Weight and Body Composition: A Systematic Review and Meta-Analysis of Randomized Controlled Trials.” Food Sci Nutr 2026;14(1):e71365. MA-RCT, 36 trials.

Waist circumference −1.61 cm (−2.06 to −1.16)
Fat mass −2.00 kg (−2.24 to −1.78)
Body weight −0.98 kg (−1.63 to −0.33)
BMI −0.24 kg/m² (ns)

That waist-circumference and fat-mass result is the most directly on-target finding in this entire section for a man whose stated goal is visceral fat loss with lean mass preserved. Caveat: a −2.00 kg fat mass effect with a CI that tight across 36 heterogeneous trials is suspiciously precise and this is a lower-profile journal; I would want to see it replicated by the Toronto 3D group before treating the magnitude as solid. The direction is corroborated by an independent well-controlled trial: Bäck S, Päivärinta E, Pellinen T, et al. Eur J Nutr 2025;64(6):259. RCT, n=102 healthy working-age men (mean 38 y), 6 wk, partial substitution of red/processed meat (760→200 g/wk) with non-soy legumes: LEGUME group had lower total and LDL cholesterol (both p<0.001) and lower BMI and weight (both p=0.009), higher fibre, PUFA and iron intakes, at the cost of lower vitamin B12 and iodine status.

Satiety, the second-meal effect, and the SCFA mechanism

Li SS, Kendall CW, de Souza RJ, et al. “Dietary pulses, satiety and food intake: a systematic review and meta-analysis of acute feeding trials.” Obesity 2014;22(8):1773–80. MA-RCT, 9 acute trials. Pulses produced a 31% greater satiety incremental AUC (RoM 1.31, 95% CI 1.09–1.58, p=0.004, I²=0%) vs isocaloric controls — but no significant effect on second-meal food intake (MD −19.94 kcal, ns). Honest reading: pulses make you feel fuller; the evidence that this translates into eating less at the next meal is not there in the acute data.
Second-meal glucose effect is real, though — Mollard RC, Wong CL, Luhovyy BL, Cho F, Anderson GH. “Second-meal effects of pulses on blood glucose and subjective appetite following a standardized meal 2 h later.” Appl Physiol Nutr Metab 2014;39(7):849–51. XO. All pulses (chickpeas, yellow peas, navy beans, lentils) lowered blood glucose AUC over the following 2 h vs white bread; after the standardised second meal, glucose was lower following lentils and chickpeas at 150 and 165 min, and AUC was lower after lentils.
SCFA/microbiome mechanism is plausible and partially demonstrated but weaker than usually claimed. Marinangeli CPF, Harding SV, Zafron M, Rideout TC. “A systematic review of the effect of dietary pulses on microbial populations inhabiting the human gut.” Benef Microbes 2020;11(5):457–468. From 2,444 screened articles, only 5 studies met inclusion, results were “inconsistent,” and the authors conclude more human studies are needed. Anyone asserting a confident pulse→microbiome→health causal chain is ahead of the data. The related whole-grain SCFA data (Vanegas 2017: higher stool acetate and total SCFAs, higher Lachnospira, lower Enterobacteriaceae) is the better-evidenced version of the same mechanism.
Processing destroys the advantage. Ramdath DD, Wolever TMS, Siow YC, et al. Foods 2018;7(5):76. XO. Mean GI of processed lentils ranged from 25 ± 3 (boiled) to 66 ± 6 (spray-dried); lentil-based food items elicited a relative glycaemic response of 40 ± 3% vs 73 ± 3% for matched potato-based items (p<0.001). Boiled whole lentils at GI 25 are among the lowest-GI foods that exist. Spray-dried lentil powder in a protein bar is not the same food.

The observational evidence is much weaker — and the Blue Zones claim is worse than weak

Viguiliouk E, Glenn AJ, Nishi SK, et al. “Associations between Dietary Pulses Alone or with Other Legumes and Cardiometabolic Disease Outcomes: An Umbrella Review and Updated Systematic Review and Meta-analysis of Prospective Cohort Studies.” Adv Nutr 2019;10(Suppl_4):S308–S319. MA-PC, 28 unique cohorts. Highest vs lowest intake:

CVD incidence RR 0.92 (0.85–0.99) — GRADE low
CHD incidence RR 0.90 (0.83–0.99) — very low
Hypertension RR 0.91 (0.86–0.97) — very low
Obesity RR 0.87 (0.81–0.94) — very low
No association with MI, stroke, or diabetes incidence, or with CVD/CHD/stroke mortality.

Read that carefully: the pulse cohort data show no association with incident diabetes at all, despite pulses having the best interventional glycaemic data of any food class. That is a good illustration of how little single-food cohort estimates should move you in either direction.

On the Blue Zones “beans are the cornerstone of longevity” claim — be skeptical. The methodological problems are now documented in the peer-reviewed literature:

Echeverry-Raad J, Sturmberg JP. “Red zones: the true color behind the myth of blue zones geographic longevity.” Rev Salud Publica (Bogota) 2025;27(3):119673. Peer-reviewed critical essay: “the lack of comprehensive global epidemiological studies, biased population selection, and uncontrolled confounding variables compromise these claims. Moreover, recent investigations indicate that many longevity records in BZs may result from clerical errors or fraudulent documentation, especially in regions with unreliable vital records.”
Poulain M, Herm A. “How to become a centenarian in four weeks? Myths and limits of longevity recipes: a critical review.” Minerva Med 2026;117(3):147–160. Notable because Poulain is a co-originator of the Blue Zones concept, and the adult’s own review now cautions that “studies of individual centenarians provide only limited evidence linking specific behaviors to survival beyond age 100.”
Saul Newman’s widely-discussed analysis arguing that remarkable-age records concentrate in regions with poor birth registration, high poverty and pension-fraud incentives — I could not fetch the preprint in this session (bioRxiv returned 429/1015) and flag it as unverified.

The Blue Zones bean claim is ecological correlation on top of possibly-corrupt age data. It is not evidence. Fortunately, pulses do not need it — the RCT stack above is far better than what almost any other food class can offer.

The downsides, settled honestly

Lectins. Real toxin, real illness, trivially preventable — and nothing to do with the popular “lectin-free diet” narrative. FDA Bad Bug Book, 2nd ed., Phytohaemagglutinin (kidney bean lectin) chapter. REG:

Raw kidney beans: 20,000–70,000 haemagglutinating units (hau). Fully cooked: 200–400 hau — a ~100× reduction.
Toxic dose: “as few as four or five raw beans.” Onset 1–3 h, recovery usually 3–4 h, no reported mortality.
White kidney beans ≈ one-third the toxin of red; broad beans 5–10%.
“Bender and Readi found that boiling the beans for 10 minutes (100 °C) completely destroyed the toxin.” FDA advises 30 min of boiling for margin. Soak ≥5 h, discard the water, boil in fresh water ≥30 min.
The actual hazard is slow cookers: “Studies of casseroles cooked in slow cookers revealed that the food often reached internal temperatures of only 75 °C or less, which is inadequate for destruction of the toxin.” Seven UK outbreaks 1976–79; further incidents 1988; US reports anecdotal. Case reports continue (e.g. Haile AM, et al. Int J Emerg Med 2026;19(1):49 — an 8-year-old with hypovolaemic shock and prerenal AKI after a home-cooked red kidney bean dish).
Verdict for meal-delivery food: a non-issue. Canned beans are retort-sterilised at >100 °C; any properly boiled bean has 0.5–2% of the raw lectin. The “lectins cause chronic disease/leaky gut” claim has no human evidence. Note the irony that the same PHA is used clinically to test immune competence.

Phytates. Genuinely reduce non-haem iron and zinc absorption in the same meal — this is well-established MECH/human-absorption science. Two reasons it does not matter for this subject: (1) the adult eats ~[calorie target removed] of a mixed omnivorous diet with animal protein present, which supplies haem iron unaffected by phytate and supplies the meat factor that enhances non-haem absorption; (2) soaking, sprouting, fermentation and cooking all lower phytate substantially (verified in the quinoa context by Goussi R, et al. J Sci Food Agric 2026 — “both boiling and steaming significantly reduced phytic acid content”; and in sourdough by Ogaji AO, et al. Sustain Microbiol 2025;2(4):qvaf030 — LAB fermentation caused “a significant decrease in phytic acid content”). Phytate is a real consideration for populations with borderline iron/zinc status on near-exclusively plant diets — Reynolds 2019 flags exactly this — and is not a consideration here.

FODMAPs / GI distress. The one genuinely legitimate downside and the one most likely to affect adherence. Galacto-oligosaccharides (raffinose, stachyose) are fermented in the colon; gas, bloating and urgency are common on a step-change increase. Practical, evidence-adjacent management: ramp intake gradually over 2–4 weeks (adaptation is real), rinse canned beans (removes some oligosaccharides along with sodium), and note that canned/pressure-cooked beans are generally better tolerated than home-boiled because more oligosaccharide leaches into the discarded liquid. For a recreational sport active adult, timing matters more than dose — do not load beans in the 3–4 hours before training.

Sodium in canned beans — quantified. USDA SR Legacy COMP, mg sodium per 100 g:

Product	Sodium	Potassium	Fibre
Black beans, boiled from dry, no salt	1 mg	355 mg	8.7 g
Black beans, boiled from dry, with salt	237 mg	355 mg	8.7 g
Black beans, canned, low sodium	138 mg	308 mg	6.9 g
Red kidney beans, canned, solids + liquids	256 mg	260 mg	4.3 g
Red kidney beans, canned, drained	231 mg	277 mg	5.5 g
Red kidney beans, canned, drained + rinsed	208 mg	250 mg	6.0 g
Red kidney beans, canned, low sodium	117 mg	260 mg	5.3 g
Chickpeas, canned, solids + liquids	278 mg	144 mg	4.4 g
Chickpeas, canned, drained	246 mg	126 mg	6.4 g
Chickpeas, canned, drained + rinsed	212 mg	109 mg	6.3 g
Baked beans, canned, plain/vegetarian	343 mg	224 mg	4.1 g
Hummus, commercial	426 mg	312 mg	5.5 g

Two findings here contradict popular advice.

Draining and rinsing removes far less sodium than the commonly-quoted “~40%.” By USDA’s own paired entries, kidney beans go 256 → 208 mg/100 g (−19%) and chickpeas 278 → 212 mg/100 g (−24%). Rinse anyway — it is free — but do not treat it as a solution.
The effective levers are upstream: buy no-salt-added / low-sodium cans (−54%, to ~117–138 mg/100 g) or cook from dry (1 mg/100 g — a 250-fold reduction). For a man actively reducing sodium, that is the single highest-leverage swap in this entire section.
Hummus is the sodium trap in this class at 426 mg/100 g — a 60 g portion is ~256 mg. It is still a good food (312 mg potassium, 5.5 g fibre, 7.8 g protein per 100 g); it just needs to be counted as a sodium item, not a free vegetable dip.

Detection strings: black beans, black bean, pinto beans, kidney beans, red beans, cannellini, white beans, navy beans, great northern, butter beans, lima beans, fava beans, broad beans, chickpea, chickpeas, garbanzo, garbanzo beans, lentil, lentils, red lentil, green lentil, french lentil, du puy, beluga lentil, split pea, split peas, black eyed pea, black-eyed peas, adzuki, mung bean, edamame, soybeans, refried beans, hummus, falafel, dal, daal, dahl, channa, chana Ambiguity flags: bean matches green beans / string beans / haricots verts (a non-starchy vegetable, not a pulse — score separately), vanilla bean, cocoa beans, coffee beans, bean sprouts (much lower fibre). pea matches green peas (a starchy vegetable, intermediate) and snow peas/snap peas (non-starchy). refried beans are frequently made with lard and carry ~450–600 mg sodium per serving — score them lower than plain beans. baked beans are a sugar-bearing item (brown sugar/molasses sauce) as well as high-sodium. falafel is deep-fried — score as fried, not as a pulse.

Pasta, bread, corn/masa, and added-sugar-bearing whole foods

Pasta — the GI is genuinely lower than its flour implies

Why: pasta is made from coarsely-milled durum semolina extruded under pressure into a dense, low-porosity matrix in which starch granules are physically entrapped in a continuous gluten-protein network. Amylase access is limited by that matrix, not by the flour’s composition. This is the single clearest demonstration in food science that food form beats ingredient list for glycaemic response — and it is why “whole wheat” bread behaves like white bread (Zafar 2020) while white pasta behaves like a low-GI food.

Atkinson 2021 confirms empirically: “Dairy products, legumes, pasta, and fruits were usually low-GI foods (≤55…) and had consistent values around the world,” in explicit contrast to breads and cereals which spanned the full range.

Practical corollaries: al dente lowers GI further (less starch gelatinisation); overcooking raises it; fresh egg pasta and thin shapes are higher than dried thick shapes.

Chiavaroli L, Kendall CWC, Braunstein CR, Blanco Mejia S, Leiter LA, Jenkins DJA, Sievenpiper JL. “Effect of pasta in the context of low-glycaemic index dietary patterns on body weight and markers of adiposity: a systematic review and meta-analysis of randomised controlled trials in adults.” BMJ Open 2018;8(3):e019438. MA-RCT, 32 trial comparisons, n=2,448. Pasta within low-GI dietary patterns vs higher-GI patterns: body weight −0.63 kg (−0.84 to −0.42), BMI −0.26 kg/m² (−0.36 to −0.16); no effect on waist, WHR, body fat or SAD. GRADE moderate for weight/BMI. Two caveats: (1) there was not a single trial of pasta alone — the comparison is low-GI-pattern vs high-GI-pattern, which is not the same question; (2) the author group discloses extensive funding from Barilla among others. Take the direction, discount the magnitude.

Whole wheat pasta: more fibre than semolina; the GI advantage is smaller than people expect because semolina pasta is already low-GI.
Legume pasta (chickpea, red lentil, edamame, black bean): the genuine upgrade. It roughly doubles the protein and triples the fibre versus semolina and imports the pulse glycaemic profile. Supported mechanistically by Kanata MC, et al. Food Funct 2025;16(11):4548–4561 (XO RCT, n=15) — substituting large-particle chickpea flour into bread lowered glucose iAUC, raised GLP-1 iAUC, and lowered hunger/raised fullness vs wheat bread, whereas finely-milled chickpea flour did not. Particle size again.

Bread — sourdough vs commercial

Özer YE, Cengiz H, Demirci T, et al. “Glycemic responses to whole grain sourdough bread versus refined white bread in patients with gestational diabetes.” Wien Klin Wochenschr 2023;135(13-14):349–357. XO, 43 women with GDM + 38 healthy pregnant controls, identical breakfasts differing only in bread type. Result: white wheat bread caused 45.5% more insulin secretion and 9.6% higher first-hour postprandial glucose than sourdough whole grain wheat bread, in both GDM and healthy groups; second-hour glucose was the same.

Read that honestly: a ~10% blunting of the first-hour glucose peak. Sourdough is better than commercial bread, and the mechanism is credible (organic acids from lactic acid bacteria slow gastric emptying and inhibit amylase; fermentation also degrades phytate — Ogaji 2025). But it is a modest effect and, critically, “sourdough” on a supermarket label frequently means a fast-fermented dough with added acid or a sourdough “flavour,” not a genuinely long-fermented one. The reliable signals are a short ingredient list (flour, water, salt, culture — no commercial yeast, no vinegar, no dough conditioners) and a real fermentation time.

Ranking for this subject: genuine long-ferment whole-grain sourdough > whole-grain commercial bread > white sourdough > commercial white bread ≈ brioche/soft rolls (worst; typically also carry added sugar and 350–500 mg sodium per 100 g).

Corn, masa and tortillas

USDA SR Legacy COMP, per 100 g:

Corn tortilla, no added salt (173241): 222 kcal, 5.7 g protein, 5.2 g fibre, 154 mg K, 65 mg Mg, 11 mg sodium
Sweet corn, boiled, no salt (169999): ~96 kcal, 3.41 g protein, 2.4 g fibre, 218 mg K

Nixtamalization — soaking and cooking maize in an alkaline calcium hydroxide (lime) solution — is one of the genuinely great pieces of traditional food technology. It (a) liberates bound niacin into a bioavailable form, which is why maize-eating populations that adopted nixtamalization did not get pellagra and those that adopted maize without it did; (b) adds substantial calcium from the lime; (c) improves protein quality and gelatinises the starch. I verified the general nixtamalization literature exists but did not obtain a primary quantitative source for the niacin liberation figures in this session — the mechanism is textbook-settled but the specific numbers are unverified here.

Practical ranking: corn tortilla (nixtamalized masa, ~5.2 g fibre/100 g, 11 mg sodium) > whole wheat tortilla > flour tortilla (refined flour, usually with added fat — often partially hydrogenated or palm — and 400–600 mg sodium per 100 g). Two corn tortillas are one of the best-value starch vehicles in a meal-delivery context: fibre comparable to whole grain bread, essentially no sodium, naturally portion-limited.

Added-sugar-bearing “whole” foods — against the adult’s 36 g/day added-sugar target

Muraki I, Imamura F, Manson JE, Hu FB, Willett WC, van Dam RM, Sun Q. “Fruit consumption and risk of type 2 diabetes: results from three prospective longitudinal cohort studies.” BMJ 2013;347:f5001. PC. 66,105 + 85,104 women + 36,173 men; 3,464,641 person-years; 12,198 incident T2D cases.

Pooled HR per 3 servings/week:

Total whole fruit: 0.98 (0.97–0.99)
Blueberries 0.74 (0.66–0.83); grapes and raisins 0.88 (0.83–0.93); prunes 0.89 (ns); apples/pears 0.93; bananas 0.95; grapefruit 0.95; peaches/plums/apricots 0.97 (ns); oranges 0.99 (ns); strawberries 1.03 (ns); cantaloupe 1.10 (1.02–1.18)
Fruit juice: 1.08 (1.05–1.11)
P for heterogeneity between fruits <0.001 in all cohorts

Two things to take from this. First, whole fruit and its own juice point in opposite directions in the same cohort, on the same questionnaire, in the same people — that is about as clean a within-study contrast as nutritional epidemiology offers, and it is the strongest argument that the matrix (fibre, chewing, volume, rate of delivery), not the fructose, is what matters. Second, “grapes and raisins” — i.e. including dried fruit — came out at HR 0.88, among the most protective items. Dried fruit is not a hidden villain.

Dried fruit (raisins, dates, dried apricots, dried figs, prunes): sugar-dense by weight (~60–65 g sugar/100 g) but that sugar is intrinsic, not added, so it does not count against the adult’s 36 g/day added-sugar target. It carries fibre, potassium and polyphenols. The exceptions that DO count: sweetened dried cranberries, sweetened dried mango, candied/glacé fruit, and most “fruit-juice-infused” dried fruit — these have cane sugar or apple-juice concentrate added and are essentially candy. A supportive but weak MR signal: Gong L, et al. Br J Nutr 2024;132(8):988–995 found a protective genetic association between dried fruit intake and T2D without complications — Mendelian randomisation on a UK Biobank self-report exposure, so low confidence.
Fruit juice: HR 1.08 per 3 servings/week; +0.31 lb per 4 years per daily serving (Mozaffarian 2011). 100% fruit juice is technically not “added sugar” under the US Nutrition Facts rule, which is a labelling artefact rather than a physiological fact — for this subject’s purposes, count juice sugar as if it were added. A 250 mL glass of orange juice is ~22 g sugar with ~0.5 g fibre.
Honey, maple syrup, agave, brown rice syrup, date syrup, molasses in sauces and glazes: these are added sugars by regulatory definition and by physiology, regardless of “natural” framing. This is where meal-delivery meals hide sugar — teriyaki, barbecue, sweet chilli, hoisin, honey-mustard, orange sauce, glazes and marinades routinely carry 8–18 g added sugar per serving, i.e. 22–50% of the adult’s 36 g/day target in a single sauce, plus 400–900 mg sodium. The sauce, not the starch, is usually the sugar problem in these meals.
Note also brown rice syrup: a double hit — it is an added sugar and it is a concentrated rice product, so it is the one rice-derived ingredient where the arsenic question is genuinely worth flagging.

Detection strings — pasta: pasta, spaghetti, penne, rigatoni, fusilli, linguine, fettuccine, orzo, macaroni, elbow, farfalle, rotini, bucatini, cavatappi, lasagna, gnocchi, couscous, israeli couscous, whole wheat pasta, chickpea pasta, lentil pasta, red lentil pasta, edamame pasta, banza Detection strings — bread: sourdough, whole wheat bread, multigrain bread, white bread, ciabatta, focaccia, baguette, brioche, naan, pita, whole wheat pita, english muffin, bagel, croutons, breadcrumbs, panko Detection strings — corn/masa: corn tortilla, masa, masa harina, nixtamal, hominy, polenta, grits, arepa, tortilla chips, sweet corn, corn kernels, corn on the cob, elote Detection strings — added sugar in savoury items: honey, maple syrup, agave, brown rice syrup, cane sugar, brown sugar, molasses, corn syrup, high fructose corn syrup, date syrup, hoisin, teriyaki, bbq sauce, barbecue sauce, sweet chili, sweet and sour, honey mustard, glaze, orange sauce, fruit juice concentrate, apple juice concentrate Ambiguity flags: corn matches corn syrup, high fructose corn syrup, cornstarch, corn flour, cornmeal, popcorn, baby corn (a non-starchy vegetable) and corn oil — you must match corn syrup, cornstarch, corn oil and baby corn first and exclude them before matching corn. couscous is pasta (semolina), NOT a whole grain, despite its health-food positioning — flag it explicitly, since misclassifying couscous as a whole grain is a very common error. gnocchi is potato + refined flour, not pasta. tortilla alone is ambiguous between corn and flour. honey matches honeydew and honey mustard. pita/naan are usually refined. polenta/grits are usually de-germed cornmeal, i.e. refined.

Cross-cutting notes for the scoring pipeline

Preparation dominates identity in this section. Potato→fries moves the score by ~7 points; lentil→spray-dried lentil powder moves GI from 25 to 66; rice→rice cake moves GI from ~50 to ~100. Wherever the ingredient string carries a preparation cue (fried, crispy, breaded, tempura, glazed, candied, instant, puffed, syrup, flour, powder), it should override the base food score.
Do not double-count sodium. Canned beans, hummus, baked beans, refried beans and flour tortillas all carry a sodium penalty that presumably also flows through a separate sodium term in the meal score. Score them here on their carbohydrate/fibre merits and let the sodium term do its own work, or the subject’s sodium reduction goal will be penalised twice.
Displacement is the right frame for near-neutral items. White rice is not harmful; it is an opportunity cost. A meal that uses white rice instead of bulgur loses ~4 g of fibre and ~50 mg of magnesium. That is what the −1 encodes, not a toxicological claim.
The adult is an active adult at [calorie target removed]. No starch in this section should be scored below −1 on glycaemic grounds alone. The negative scores in the table below are driven by added fat + energy density + sodium (fries, chips) or by added sugar (juice, sweetened dried fruit, glazes) — never by “it’s a carbohydrate.”
The three highest-value swaps identified in this section, in order: (a) canned beans → no-salt-added or dry-cooked beans (250-fold sodium reduction, no downside); (b) white/brown rice → bulgur, barley, or quinoa (up to 11× the fibre, no arsenic, better amino acid profile); (c) sweetened sauces → herb/acid/spice-based sauces (recovers 20–50% of the daily added-sugar budget and several hundred mg of sodium per meal).

Machine-readable scoring table

```

Section 3 — Fats and Oils

Scope note: scores are for a healthy adult at ~[calorie target removed]/day, cutting visceral fat, reducing sodium. Fats are energy-dense; in a fixed-calorie context the question is almost never “is this fat good” but “what does this fat displace.” That framing is applied throughout.

Evidence-quality caveat that applies to this entire section: outside of lipid-panel endpoints, essentially all long-term fat-and-disease evidence is (a) prospective cohorts with FFQ exposure measurement, or (b) a small number of old, methodologically compromised RCTs. Effect sizes in the HR 0.8–1.2 band from cohorts are inside the noise floor of nutritional epidemiology. Lipid endpoints (LDL-C in mg/dL from controlled feeding) are the one place we have genuinely reliable, replicated, dose-responsive human causal data.

3.1 Extra-virgin olive oil (EVOO)

The PREDIMED problem — stated precisely

Original: Estruch R et al., “Primary Prevention of Cardiovascular Disease with a Mediterranean Diet,” NEJM 2013;368:1279–1290. Design: multicentre PRCT, n=7,447, Spain, high-CV-risk primary prevention, 3 arms (MedDiet+EVOO ~1 L/week supplied; MedDiet+mixed nuts 30 g/d; control = advice to reduce dietary fat). Median follow-up 4.8 y. Primary composite (MI, stroke, CV death) HR ≈ 0.70 for both intervention arms.
Retraction and republication: 13 June 2018. NEJM retracted the 2013 paper and simultaneously published a corrected version (Estruch R et al., NEJM 2018;378:e34). Verified.
What actually went wrong (verified): randomisation departed from protocol for 1,588 of 7,447 participants (21%), via three distinct failures — (i) household members of enrolled participants were enrolled and assigned to the same arm without their own randomisation; (ii) at one of 11 sites, entire clinics were allocated rather than individual patients; (iii) at another site, apparent inconsistent use of randomisation tables.
What the reanalysis did: re-ran with statistical correction for within-family and within-clinic correlation, and separately re-ran excluding the 1,588 improperly-assigned participants. Effect estimates were essentially unchanged (~30% relative reduction in the composite endpoint persisted).

Honest reading. This is the single most important nuance in the whole “olive oil is proven” story and it is usually stated wrong in both directions.

It is NOT true that “PREDIMED was debunked.” The effect survived two different sensitivity analyses.
It is ALSO not true that PREDIMED is a clean individually-randomised RCT demonstrating that olive oil causes CV risk reduction. For 21% of the sample it is functionally a cluster/quasi-randomised trial, and the paper is best described as a randomised trial with a partially compromised allocation.
Further limitations that predate the retraction and are independent of it: (a) the intervention was a whole dietary pattern, not olive oil in isolation — you cannot attribute the effect to EVOO specifically, since the EVOO arm also got MedDiet counselling; (b) the control arm was actively told to reduce fat, i.e. a not-obviously-neutral comparator, and control-arm counselling intensity was lower for the first ~3 years (a known asymmetry criticised in the literature); (c) unblinded, food supplied free to intervention arms only; (d) population was high-CV-risk Spanish adults aged 55–80 — externally quite distant from a adult.
Design: RCT (partially compromised allocation). Strength for our purposes: moderate, not high.

The cohort evidence

Guasch-Ferré M, Li Y, Willett WC et al., “Consumption of Olive Oil and Risk of Total and Cause-Specific Mortality Among U.S. Adults,” J Am Coll Cardiol 2022;79(2):101–112. Design: two prospective cohorts, n=92,383 (60,582 NHS women + 31,801 HPFS men), 1990–2018, 28 y follow-up, 36,856 deaths. Highest category >7 g/day (>~½ tbsp/day) vs lowest (never/<4.5 g/mo). Verified.
- Total mortality HR 0.81 (0.78–0.84); CVD mortality 0.81 (0.75–0.87); cancer 0.83 (0.78–0.89); neurodegenerative 0.71 (0.64–0.78); respiratory 0.82 (0.72–0.93).
- Absolute translation: US male all-cause mortality in this age-and-health band is not the right baseline; use the cohort’s own. Over 28 y in a cohort of middle-aged health professionals, roughly 40% died (36,856/92,383). A 19% relative reduction on a ~40% 28-year absolute risk is roughly 40% → ~34%, i.e. ~6 absolute percentage points over 28 years, or crudely ~0.2 pp/year. That is a large number for a single food — which is itself the reason to be suspicious of it.
- Why to discount it heavily: in 1990s America, using olive oil at all was a marker of an entire education/income/health-behaviour cluster. The comparison group is people who essentially never ate olive oil. Residual confounding by socioeconomic status and overall diet quality is the dominant plausible explanation for an HR of 0.71 for neurodegenerative mortality — there is no credible mechanism by which half a tablespoon of olive oil daily cuts dementia death by 29%. Treat the direction as probably real and the magnitude as substantially inflated.
- Notably, substituting 10 g/d of olive oil for margarine/butter/mayo/dairy fat was associated with 8–34% lower mortality — the substitution analyses are more believable than the absolute-intake analyses.

Polyphenols — EVOO vs refined, quantified

Total phenolics in olive oil: broadly ~100–600 mg/kg, with typical commercial EVOO at ~100–250 mg/kg and “high-polyphenol” oils >300 mg/kg (verified in review literature; exact figures vary by cultivar, harvest timing, and assay, so treat as an order-of-magnitude range, not a precise constant).
Key compounds: hydroxytyrosol, tyrosol, oleuropein aglycone, oleocanthal (the peppery throat-catch compound; the “ibuprofen-like COX inhibition” finding is Beauchamp et al., Nature 2005 — mechanistic/in-vitro plus a pharmacological analogy, not a clinical outcome study; the doses implied are far below therapeutic NSAID doses).
EFSA health claim (verified): permitted for oils containing ≥5 mg hydroxytyrosol and derivatives per 20 g of oil, i.e. ≥250 mg/kg. Claim wording relates to protection of blood lipids from oxidative damage. Critically, a large share of supermarket “extra virgin” oils do not meet this threshold.
Refined / “pure” / “light” olive oil: refining (deodorising, neutralising, bleaching) strips essentially all polar phenolics. Refined olive oil retains the fatty acid profile (~70–75% oleic acid) but is nutritionally close to “a monounsaturated fat with vitamin E.”
Does the polyphenol difference matter clinically? The best-designed evidence is the EUROLIVE study (Covas MI et al., Ann Intern Med 2006, crossover feeding trial in ~200 European men across 5 countries, comparing low- / medium- / high-polyphenol olive oils at 25 mL/d). It found a dose-dependent increase in HDL-C and decrease in oxidised LDL with higher phenolic content — real, replicated, but the effect sizes are small and the endpoint is a biomarker, not events. Note: I recalled the EUROLIVE design and approximate n from memory and did not re-verify n in this session — treat n≈200 as unverified recall; the existence, design and direction of the trial are well-established.
Verdict: EVOO > refined olive oil, but the increment attributable to polyphenols is modest and biomarker-level. Most of olive oil’s defensible benefit is (a) it’s ~73% MUFA and low in SAT, so it lowers LDL relative to butter/coconut/palm, and (b) it displaces worse fats. The polyphenol story is a real but second-order bonus. Scoring should reflect that: EVOO gets a good score mostly for being olive oil, and a small bonus for being extra-virgin.
Heat: phenolics degrade with heating and with storage (light/oxygen). Frying at 180 °C for extended periods substantially depletes them; a normal sauté depletes them less. EVOO’s smoke point (~190–210 °C for good-quality low-FFA oil) is higher than the “don’t cook with EVOO” folklore claims, and its high oxidative stability (high MUFA, high antioxidant load) makes it more stable in cooking tests than many high-PUFA oils. The “never heat EVOO” advice is not well supported.
Fraud/adulteration: a persistent, real problem — repeated national surveys (notably UC Davis 2010–11 and various Italian/EU enforcement actions) have found substantial fractions of imported “extra virgin” oil failing IOC/EVOO sensory or chemical standards. I did not verify a specific failure percentage in this session — treat any specific % as unverified. Practical consequence for scoring: a meal listing “extra virgin olive oil” cannot be assumed to actually contain a high-polyphenol oil, which is another reason to keep the EVOO-over-plain-olive-oil bonus small.

Tier: beneficial. Score: EVOO +6, plain/refined olive oil +4. Confidence: moderate-high for direction, low for magnitude.

3.2 Olive POMACE oil — the one that actually deserves scrutiny

This appears in prepared/meal-delivery foods because it is cheap, has a neutral flavour, high smoke point, and can legally carry the word “olive.”

What it is

After mechanical pressing yields virgin oil, the remaining solid+moist residue (pomace, ~5–8% residual oil) is dried and extracted with hexane. The crude pomace oil is then chemically refined at high temperature (neutralisation, bleaching, deodorisation at ~200–260 °C). “Olive pomace oil” sold at retail is refined pomace oil blended with some virgin oil for flavour. Under EU/IOC standards it is a distinct category and may not be called “olive oil” unqualified.

Composition

Fatty acid profile is broadly similar to olive oil (high oleic). From a pure LDL-cholesterol standpoint, it behaves like a monounsaturated oil — this is the honest good news.
Polyphenols: essentially stripped by refining. No EFSA hydroxytyrosol claim eligibility.
Contains elevated squalene, triterpenic alcohols/acids (oleanolic, maslinic acid), and aliphatic alcohols relative to virgin oil — some literature argues these are mildly beneficial. Weak, mechanistic-level evidence.

The contaminant issues — quantified

Mineral oil hydrocarbons (MOSH/MOAH). This is the genuinely distinguishing problem, and it is well documented. Commercial olive pomace oil products have been measured at ~33–205 mg/kg MOSH and ~2–55 mg/kg MOAH (verified via peer-reviewed olive-oil MOH literature). For comparison, virgin olive oils occasionally exceed a 2.0 mg/kg MOAH reference level, and pomace oils typically run ~10× higher. Mechanism: solvent extraction reconcentrates hydrocarbons that remain on the solid residue and that accumulate during open-air pomace storage.
- Important mitigating finding, and it is stronger than I initially assumed. In the peer-reviewed 2D-GC characterisation of commercial pomace oils, the MOAH fraction contained only highly alkylated 1–2 aromatic-ring compounds, with the absence of 3-or-more-ring species confirmed in all samples tested. The ≥3-ring aromatics are precisely the genotoxic species that drive EFSA’s MOAH concern; the 1–2-ring alkylated compounds are not genotoxic. The same authors found evidence that the hydrocarbons are largely endogenous — taken up by the plant from soil and atmosphere during growth — rather than introduced by processing contamination, and that atmospheric deposition during pomace storage was insignificant. Solvent extraction yields 2–6× higher concentrations than physical centrifugation alone, which explains the pomace-vs-virgin gap.
- The authors explicitly argue against applying a 2 mg/kg MOAH withdrawal threshold to these oils, on two grounds: 2 mg/kg is essentially the analytical limit of detection rather than a toxicologically derived value, and the aromatic species actually present are non-genotoxic.
- Correction to the intuitive read: the raw “10× more MOAH than virgin olive oil” number looks alarming and is the number that circulates, but the speciation data substantially defuse it. I initially framed this as a live precautionary risk; the better-supported position is that it is a regulatory-analytical artefact more than a toxicological one.
- Honest status: there is no harmonised EU maximum limit for MOAH in olive pomace oil as a legally binding number in the same way there is for, say, benzo[a]pyrene; MOH is governed by recommendations, monitoring (Commission Recommendation on MOH monitoring) and a de facto ~2 mg/kg MOAH action threshold in some contexts. This is a regulatory-margin / precautionary issue, not a demonstrated human harm at dietary exposure.
PAHs / benzo[a]pyrene. Historically the pomace-drying step (direct-flame hot-air drying) produced high PAH loads; this is why PAH limits in pomace oil exist and why producers moved to indirect drying and activated-carbon treatment. EU maximum levels are set in Regulation (EC) 1881/2006 (and successor Reg. (EU) 2023/915), with BaP used as the marker PAH alongside a PAH4 sum. Well-run modern pomace oil complies. Occasional enforcement failures happen.
3-MCPD esters and glycidyl esters (GE). Formed during high-temperature deodorisation of any refined oil — this is not pomace-specific (refined palm is the worst offender). EU maximum levels (verified in the search literature): 2.5 mg/kg for 3-MCPD/3-MCPD esters combined and 1.0 mg/kg for glycidyl esters in vegetable oils; the IOC advised 1.25 mg/kg for 3-MCPD esters in refined olive oil. Virgin/cold-pressed oils are essentially free of both because they are never heated in refining.
Residual hexane. Legally limited (EU solvent residue limits, ~1 mg/kg order of magnitude for extraction solvents in the final fat). Hexane is highly volatile and is removed in deodorisation. This is, in my judgement, the least substantiated of the concerns — hexane residue in refined oils is not a credible dietary risk, despite being the thing wellness content focuses on most.

Verdict on pomace oil

Two-part answer, because the honest answer is not a single direction:

Nutritionally (fatty acids, LDL effect): pomace oil is roughly as good as refined olive oil. It is far better than coconut/palm/butter for LDL. If the alternative fat in that meal was butter, pomace oil is an improvement.
Contaminant-wise: pomace oil carries by far the highest MOSH/MOAH load of common culinary oils (roughly an order of magnitude above virgin), but the aromatic fraction is speciated as non-genotoxic 1–2-ring alkylated compounds with 3+-ring species absent. It also carries the refining-derived 3-MCPD/GE load that virgin oils don’t — that one is real but is shared with all refined oils and is worse in refined palm.
So: it is closer to “refined oil with a scary name” than I expected going in. The honest downgrade versus EVOO is not driven by contaminants; it is driven by the complete loss of polyphenols in refining. A meal using pomace oil instead of EVOO has swapped a ~+6 ingredient for a ~+2 ingredient, and the reason is missing hydroxytyrosol, not mineral oil.
Residual honest uncertainty: the speciation evidence comes from a limited number of published sample sets; MOSH (the saturated fraction) does accumulate in human tissue and EFSA has flagged it as a data gap, so “no demonstrated harm” is not the same as “demonstrated safe.”

Tier: neutral (nutritionally fine, contaminant concern largely defused on speciation; polyphenol-stripped). Score: +2. Confidence: moderate.

3.3 Nuts

Cohort and trial evidence

Bao Y, Han J, Hu FB et al., “Association of Nut Consumption with Total and Cause-Specific Mortality,” NEJM 2013;369:2001–2011. Prospective cohorts (NHS + HPFS), ~76,464 women and ~42,498 men (this n is unverified recall — the paper and journal/year are correct and well-established, but I did not re-verify the exact per-cohort n in this session). Inverse dose-response: eating nuts ≥7×/week vs never associated with ~20% lower all-cause mortality. Same healthy-user confounding caveat as olive oil.
PREDIMED nut arm (30 g/d mixed nuts: 15 g walnuts, 7.5 g almonds, 7.5 g hazelnuts) achieved the same ~30% composite-event reduction as the EVOO arm — subject to all the PREDIMED caveats above.
Lipid effects are well established and RCT-grade: nut feeding trials consistently lower LDL-C modestly (order of ~5 mg/dL at ~50–100 g/d), with the largest effects in people with higher baseline LDL.

The calorie paradox — resolved, with numbers

Nuts are ~5.5–7 kcal/g by Atwater factors yet consistently fail to cause the predicted weight gain in trials. Two mechanisms, both measurable:

Metabolisable energy is lower than Atwater predicts, because intact cell walls trap lipid. Novotny JA, Gebauer SK, Baer DJ (USDA), Am J Clin Nutr 2012 — controlled feeding, measured ME of almonds was ~32% lower than the Atwater prediction. Verified.
Processing destroys the effect. In the follow-up work (Gebauer/Novotny, USDA), measured ME was whole natural almonds 4.42 kcal/g, whole roasted 4.86, chopped 5.04, and almond butter 6.53 kcal/g — with almond butter statistically indistinguishable from the Atwater prediction (P=0.08). Verified. This is one of the cleanest, most under-appreciated findings in the section: whole almonds “lose” roughly a quarter to a third of their calories; almond butter loses none.
Satiety/compensation: nut calories are substantially compensated for at subsequent meals in feeding studies.

Practical scoring consequence: whole nuts and nut butters are not the same food for a person in a calorie deficit trying to lose visceral fat. Whole nuts: score positive. Nut butters/nut flours/nut pastes: score positive but lower, and count their calories at face value.

By type

Walnuts — highest ALA (~2.5 g/oz), the only common nut that is a meaningful plant omega-3 source; most of the “nuts and lipids/endothelial function” RCT literature uses walnuts. Best-evidenced single nut.
Almonds — best-evidenced for the ME/calorie effect; good vitamin E, magnesium.
Pistachios — favourable protein-per-calorie, and the in-shell format slows consumption (a real, if trivial-sounding, effect measured in feeding studies).
Peanuts — botanically a legume; cohort mortality data track tree nuts closely and peanuts are far cheaper, so they carry most of the population-level benefit. Note aflatoxin (see contaminants section) — controlled in regulated supply.
Cashews — lower ME reduction than almonds, higher starch, still fine.
Pecans/macadamia — highest fat, lowest protein; fine but least distinctive.
Candied/honey-roasted/salted nuts — added sugar and a real sodium load; downgrade.

Tier: beneficial. Score: whole nuts +6; nut butter/flour +3; candied or heavily salted nuts +1. Confidence: high for lipids and ME, moderate for mortality.

3.4 Seeds

Flax — highest ALA and by far the richest source of lignans. Ground vs whole matters enormously: whole flaxseed passes through largely intact and is poorly absorbed; ground flax is bioavailable. Flax has the most credible blood-pressure signal of any seed (meta-analyses of RCTs report reductions on the order of several mmHg systolic with ~30 g/d ground flax) — relevant to this subject, who is reducing sodium for presumably BP/health reasons. I did not re-verify the exact pooled mmHg estimate in this session — treat the specific number as unverified; the direction and existence of the RCT meta-analytic literature is solid.
Chia — high fibre (~10 g/oz, much of it mucilaginous soluble fibre), ALA. Weight-loss claims from chia have repeatedly failed in RCTs — this is a case where the evidence contradicts the marketing. Score for fibre, not for magic.
Pumpkin seeds (pepitas) — excellent magnesium and zinc density; good protein-per-calorie.
Sunflower seeds — high linoleic acid, vitamin E; often salted (sodium flag).
Sesame / tahini — sesamin/sesamol lignans, high calcium if unhulled. Tahini is a nut-butter-equivalent: full calorie absorption, no matrix effect.
Hemp hearts — unusually good protein content (~10 g/30 g) and a favourable 3:1 LA:ALA ratio.

Tier: beneficial. Score: whole/ground seeds +4; tahini/seed butters +3. Confidence: moderate.

3.5 Avocado — and the finding that matters most for THIS subject

Lichtenstein AH et al. (HAT — Habitual Diet and Avocado Trial), “Effect of Incorporating 1 Avocado Per Day Versus Habitual Diet on Visceral Adiposity: A Randomized Trial,” J Am Heart Assoc 2022;11:e025657. Design: multicentre, randomised, parallel-arm, non-blinded; n=1,008 adults with elevated waist circumference; 6 months; 923 (92%) with complete data; primary outcome VAT volume by MRI. Verified.
Result — the primary outcome was null. VAT increased in both arms: +0.074 L (avocado) vs +0.057 L (habitual); estimated mean difference +0.017 L (95% CI −0.024 to +0.058), P=0.405. Verified.
Secondary: small but nominally significant reductions favouring avocado — total cholesterol −2.94 mg/dL (−5.54 to −0.35), P=0.026; LDL-C −2.47 mg/dL (−4.80 to −0.13), P=0.038. Verified. No significant change in hepatic fat, inflammatory markers, or metabolic syndrome components.

This is the single most directly relevant trial in the entire fats section for this subject, and it is negative on the adult’s exact goal. An avocado a day for six months, in ~1,000 people with abdominal obesity, did not reduce visceral fat. It produced an LDL reduction of ~2.5 mg/dL, which is real but clinically trivial in isolation (for scale, LDL-lowering interventions are generally considered meaningful at ≥10 mg/dL).

Also worth noting for calorie accounting: a large avocado is ~250–320 kcal. In a [calorie target removed] budget with a deficit target, avocado is a good fat but an expensive one. It is not a free win.

Verdict: avocado is a genuinely good food (fibre ~7 g per half, potassium ~500 mg per half — directly useful given the sodium-reduction goal, since potassium is the counter-ion in the sodium-potassium ratio that actually predicts BP), but the popular framing of avocado as a belly-fat food is contradicted by the best available RCT. Score positive, but modestly, and not as a visceral-fat intervention.

Tier: beneficial. Score: +4. Confidence: high (we have a large null RCT plus solid nutrient data — unusually good evidence for a single whole food).

3.6 Coconut oil — settled

Neelakantan N, Seah JYH, van Dam RM, “The Effect of Coconut Oil Consumption on Cardiovascular Risk Factors: A Systematic Review and Meta-Analysis of Clinical Trials,” Circulation 2020;141(10):803–814. Design: meta-analysis of 16 RCTs, ≥2 weeks, coconut oil vs non-tropical vegetable oils or palm oil. Verified.
Result: LDL-C +10.47 mg/dL (95% CI 3.01–17.94) vs non-tropical vegetable oils. Also raised total cholesterol and HDL-C. No effect on triglycerides, body weight, body fat %, waist circumference, fasting glucose, or CRP. Verified.

Absolute translation. Meta-analyses of statin and Mendelian-randomisation data give roughly a ~22% relative risk reduction in major vascular events per 1 mmol/L (≈38.7 mg/dL) LDL reduction, with lifelong genetic exposure showing a substantially larger effect per unit. A +10.5 mg/dL LDL rise ≈ +0.27 mmol/L ≈ roughly a 5–6% relative increase in long-term vascular event risk if sustained. For a adult male with, say, a ~10% lifetime-to-age-75 baseline event risk on the low side, that is a fraction of a percentage point in absolute terms per year of exposure — small individually, but strictly in the wrong direction, and with zero demonstrated compensating benefit (the meta-analysis found no fat-loss, glycaemic, or inflammatory benefit at all).

The MCT defence fails on chemistry. Coconut oil is ~45–50% lauric acid (C12:0). Lauric acid is nominally a medium-chain fatty acid by carbon count, but metabolically it behaves largely like a long-chain fat: a majority of absorbed lauric acid is packaged into chylomicrons and travels via lymph, not directly to the liver via the portal vein the way C8/C10 do. Commercial “MCT oil” is C8/C10 precisely because C12 doesn’t behave the way the MCT literature requires. Studies on MCT oil do not transfer to coconut oil.

The Pacific-islander defence fails on confounding. The Kitavan and Tokelauan observations describe populations eating whole coconut (fibre, water, whole-food matrix) within a physically active, low-obesity, essentially no-processed-food dietary pattern with high fish intake. They are cross-sectional observations of a whole way of life; they say nothing about adding refined coconut oil to a Western diet.

Verdict: coconut oil raises LDL, provides no measured compensating benefit, and the two popular defences (MCT, Pacific islanders) both fail on inspection. It is not poison — at the doses in a single meal the absolute effect is small — but there is no evidence-based reason to prefer it over olive/canola. If a meal uses coconut oil for flavour authenticity (Thai curry, etc.), that’s a legitimate culinary reason, not a health one. Coconut milk in a curry is a somewhat different and gentler case (diluted, whole-food-adjacent, contributes fewer grams of SFA per serving than straight oil) and should be scored less harshly than coconut oil.

Tier: harmful (mildly). Score: coconut oil −4; coconut milk/cream −2. Confidence: high for the LDL effect, moderate for the clinical translation.

3.7 Butter, ghee, animal fats

What we know

Saturated fat raises LDL-C. This is not seriously disputed and is the most replicated finding in nutrition — controlled feeding studies, dose-responsive, mechanism understood (SFA downregulates hepatic LDL receptor expression).
What IS disputed is whether reducing SFA reduces events, and what you replace it with.
Hooper L et al., “Reduction in saturated fat intake for cardiovascular disease,” Cochrane Database Syst Rev 2020, CD011737.pub3. Design: meta-analysis of RCTs. Result: reducing saturated fat reduced combined cardiovascular events by 17%, RR 0.83 (95% CI 0.70–0.98), from 12 trials, n=53,758. Verified. But: no significant effect on all-cause mortality or cardiovascular mortality. Effect was driven by trials that replaced SFA with polyunsaturated fat.
Mozaffarian D, Micha R, Wallace S, PLoS Med 2010 — meta-analysis of 8 RCTs of SFA→PUFA replacement; reported ~19% reduction in CHD events, ~10% per 5% energy replaced. I did not re-verify the exact effect estimate or trial count in this session — treat the specific numbers as unverified recall; the paper and its direction are well-established.

The strongest case against — presented fairly

Ramsden CE et al., “Re-evaluation of the traditional diet-heart hypothesis: analysis of recovered data from Minnesota Coronary Experiment (1968–73),” BMJ 2016;353:i1246. Design: recovered raw data from a double-blind RCT, n=9,423 institutionalised adults aged 20–97, randomised to SFA replaced by corn oil / corn-oil margarine vs control. Data were recovered 2013–2015 from 9-track magnetic tapes and paper records; the full survival analysis had never been published. Verified.
- Result: the intervention lowered serum cholesterol effectively but did NOT reduce CHD death or all-cause mortality. In the accompanying meta-analysis of similar cholesterol-lowering interventions: CHD mortality RR 1.13 (0.83–1.54), all-cause mortality RR 1.07 (0.90–1.27) — i.e. point estimates in the wrong direction with CIs spanning null. Verified.
- There was a signal that greater cholesterol reduction associated with higher mortality in participants over 65.
Ramsden’s earlier Sydney Diet Heart Study recovered-data re-analysis (BMJ 2013) similarly found higher all-cause and CV mortality in the safflower-oil (high-LA, no omega-3) intervention group.
Legitimate criticisms of these: MCE had very high participant turnover (median follow-up ~1 year for many, because it ran in mental hospitals and nursing homes with high discharge/death rates), limiting power to detect atherosclerosis-timescale effects; the intervention margarine of that era contained meaningful trans fat, which confounds the “PUFA” exposure; Sydney used safflower oil providing LA with essentially no omega-3, an unusual exposure. These criticisms are real. But so is the fact that these were randomised, blinded, and pre-specified, and their existence means anyone claiming “RCTs prove replacing butter with corn oil saves lives” is overstating.

Honest synthesis for butter/ghee

Butter raises LDL relative to olive/canola. The event-level evidence that this translates into harm is suggestive but not clean: the best RCT meta-analysis (Cochrane 2020) finds an events benefit but no mortality benefit, and the two best recovered-data RCTs find no benefit at all. Butter is also, unlike coconut oil, a food with a dairy matrix (see the dairy discussion in the proteins section — cheese behaves differently from its SFA content, and butter is closer to pure fat than cheese is).

Practical position: butter as a flavour component (a pat on vegetables, a small amount in a sauce) is a minor issue and does not deserve a big penalty. Butter as a primary cooking medium in quantity is a modest negative. Ghee is butter with water and milk solids removed — slightly more concentrated SFA per gram, no meaningful health advantage over butter despite its wellness reputation; its real advantages are a higher smoke point and no lactose/casein.

Tier: contested (leaning mildly harmful). Score: butter −2, ghee −2, lard/tallow/duck fat −2. Confidence: moderate for LDL, low for events.

3.8 Seed oils — an even-handed verdict

The seed-oil discourse deserves a careful answer because both sides are overclaiming.

The case that seed oils are harmful, steelmanned

Linoleic acid intake in the US rose enormously over the 20th century, tracking the obesity and metabolic disease epidemic. (Ecological correlation — the weakest form of evidence, but it is the origin of the hypothesis.)
LA is the precursor to arachidonic acid and thence to pro-inflammatory eicosanoids. (Mechanistic, real biochemistry.)
The omega-6:omega-3 ratio in the modern diet is ~15–20:1 vs an ancestral estimate near 1–4:1. (Real, though the ancestral estimate is soft.)
The recovered-data RCTs (Ramsden, above) found no benefit and possible harm from swapping SFA for high-LA oils — the strongest actual RCT evidence the seed-oil skeptics have, and it is genuinely their best card.
Repeatedly-heated frying oil generates aldehydes (4-HNE, acrolein), oxidised triglyceride polymers, and trans isomers — this is real and measurable.

The case against, and why it wins on the current evidence

The inflammation mechanism does not show up in humans. Johnson GH & Fritsche K, “Effect of dietary linoleic acid on markers of inflammation in healthy persons: a systematic review of randomized controlled trials,” J Acad Nutr Diet 2012. 15 RCTs (8 parallel, 7 crossover). Result: NONE of the studies found significant increases in CRP, fibrinogen, PAI-1, cytokines, soluble adhesion molecules, or TNF-α. Authors’ conclusion: “virtually no evidence is available from randomized, controlled intervention studies among healthy, noninfant human beings to show that addition of LA to the diet increases the concentration of inflammatory markers.” Verified. This is the crux: the entire “seed oils are inflammatory” claim is a mechanistic extrapolation that fails when directly tested in humans.
Conversion of LA to arachidonic acid is tightly regulated and saturable. Increasing dietary LA does not proportionally raise tissue AA in humans — this has been repeatedly measured. The eicosanoid-cascade argument assumes a substrate-driven pipeline that does not behave that way in vivo.
Biomarker cohort evidence points the other way. Marklund M, Wu JHY, Imamura F et al., “Biomarkers of Dietary Omega-6 Fatty Acids and Incident Cardiovascular Disease and Mortality: An Individual-Level Pooled Analysis of 30 Cohort Studies,” Circulation 2019;139(21):2422–2436. ~69,000 participants across 30 cohorts, 13 countries, follow-up 2.5–~32 years, >15,000 CV events. Higher circulating/tissue linoleic acid was significantly associated with LOWER risk of total CVD, CV mortality, and ischaemic stroke. Verified. This design is important: it uses objective fatty-acid biomarkers rather than FFQ, which removes the single biggest measurement-error problem in nutritional epidemiology. It is still observational and still subject to confounding, but it is much better observational evidence than the usual.
Arachidonic acid in the same pooled analysis was not associated with higher CVD risk.

The part of the skeptic case that survives

Deep-fried food — repeatedly reheated commercial frying oil is a legitimately different exposure from a tablespoon of canola in a pan. Aldehyde and polar-compound formation in restaurant fryers is real and measurable, and this is a defensible reason to penalise deep-fried items — but the penalty belongs to “deep-fried,” not to “seed oil.” The same fryer with beef tallow produces its own oxidation products.
Displacement matters more than the oil. Foods high in seed oils are overwhelmingly ultra-processed foods. The correlation between high seed-oil intake and poor health outcomes at the population level is very plausibly a marker of eating a lot of packaged food, not a property of the linoleic acid molecule. This is a confounding argument that cuts against the seed-oil hypothesis while explaining why it feels true.
Genuine uncertainty: we do not have a large, long, modern RCT of high-LA vs low-LA diets with hard endpoints in a contemporary population, and we probably never will. Anyone claiming certainty in either direction is overreaching.

Verdict. On the current evidence, there is no good human evidence that seed oils used as a normal cooking fat cause inflammation or cardiovascular harm, and reasonable biomarker evidence that linoleic acid is neutral-to-protective. They are not health foods and there is no reason to seek them out; they are also not the villain. In a meal, “canola oil” or “sunflower oil” should be scored approximately neutral — slightly positive versus butter/coconut on the LDL axis, slightly negative versus EVOO on the polyphenol/displacement axis. A scoring system that heavily penalises “soybean oil” appearing in an ingredient list is encoding a wellness narrative that the human RCT data does not support, and would systematically mis-rank the 800 meals.

Tier: contested (evidence leans neutral/benign). Score: canola/sunflower/safflower/soybean/grapeseed 0; “vegetable oil” (unspecified) 0; high-oleic variants +1. Confidence: moderate.

3.9 Other fats, briefly

Sesame oil — high in sesamin/sesamolin lignans, good oxidative stability. Toasted sesame oil is used in small flavour quantities. Mildly positive: +1.
Peanut oil — high MUFA, high smoke point, neutral. 0. Refined peanut oil is not an allergen risk; cold-pressed can be.
Palm oil — ~50% SFA (mostly palmitic acid, the most LDL-raising common SFA). Refined palm is the worst common oil for 3-MCPD and glycidyl esters because of the high deodorisation temperatures required. Red/unrefined palm oil is a different, carotenoid-rich product. Score refined palm: −3. Also note “palm kernel oil” is different again and even more saturated (~80%).
Margarine / spreads — post-2018 US trans-fat ban (PHO no longer GRAS), modern soft tub spreads are interesterified or blended and contain negligible trans fat. Plant-sterol-fortified spreads genuinely lower LDL (~8–10% at 2 g sterols/day — this is one of the better-evidenced functional foods). Modern margarine: 0 to +1; stick margarine or anything listing “partially hydrogenated”: −8 (trans fat is the one fat with unambiguous, large, consistent harm evidence — though it should be nearly absent from current products).
Fat as cooking medium vs dressing — for LDL purposes it doesn’t matter. For two things it does: (i) uncooked EVOO retains more polyphenols; (ii) carotenoid absorption from vegetables requires co-ingested fat, so a fat-containing dressing on a salad meaningfully increases lutein/lycopene/beta-carotene uptake (this is well demonstrated in feeding studies — fat-free dressing on a salad substantially reduces carotenoid absorption). So a small amount of oil with vegetables is genuinely functional, not just calories.
MCT oil (C8/C10) — distinct from coconut oil; ketogenic, thermogenic effects are real but small; no relevance to visceral fat loss beyond calorie substitution. 0.

Cross-cutting: how to weight fats when scoring 800 meals

Total fat quantity dominates fat quality for a visceral-fat goal. The subject’s outcome depends on energy balance. A meal with 45 g of “good” fat is worse for the adult’s stated goal than a meal with 15 g of “mediocre” fat, holding protein and vegetables constant. The score adjustments here should be applied as modifiers on top of a macro/energy assessment, never as a substitute for it.
The fat-quality spread is narrower than popular discourse suggests. The honest ordering by LDL effect is: trans fat ≪ coconut/palm/butter < neutral seed oils ≈ pomace/refined olive < EVOO ≈ nuts ≈ avocado ≈ fish oil. The gap between “seed oil” and “olive oil” is much smaller than the gap between “any of those” and “trans fat,” and smaller than the gap between “deep-fried” and “not deep-fried.”
Preparation beats ingredient. Deep-fried > any oil-identity question in importance. Score the cooking method separately and more heavily.
Watch nut-butter vs whole-nut — genuine 25–35% calorie difference, verified, and directly relevant to a calorie-deficit goal.

Machine-readable table — Fats and Oils

Section 4 — Vegetables, Fungi, Fermented Foods, Herbs & Spices

4.0 The honest epistemic status of this entire section — read first

This needs saying plainly before any individual food, because it governs how much weight the rest of the section can bear:

There is no large randomised controlled trial showing that eating more vegetables reduces hard clinical endpoints in humans. Not one. The entire case for vegetables rests on three legs:

Prospective cohorts — consistent, directionally uniform, but with hazard ratios almost entirely in the 0.80–0.95 band, which is inside the range that residual confounding can generate. Vegetable intake is one of the strongest single markers of the entire healthy-user cluster: not smoking, exercising, higher income, higher education, more preventive care, lower alcohol, less processed food. Statistical adjustment for these is done with error-laden covariates and cannot be assumed to work.
Intermediate-marker RCTs — real, but the endpoints are blood pressure, LDL, endothelial function, microbiome composition, and inflammatory markers, not events.
Mechanism — micronutrients, fibre, polyphenols, nitrate. Mechanistically rich, and mechanism has an extremely poor track record of predicting clinical outcomes in nutrition (β-carotene, vitamin E, folate, and antioxidant supplement trials all failed, and several caused harm).

The counter-argument, which I think is decisive for practical purposes: vegetables are the highest-nutrient-density, lowest-energy-density food class available. For a person whose explicit goal is losing visceral fat at a fixed protein intake, vegetables’ primary mechanism of benefit is not phytochemical at all — it is volume and satiety per calorie, plus fibre, plus potassium (which directly matters given the sodium-reduction goal, since the sodium-to-potassium ratio predicts blood pressure better than sodium alone). Those mechanisms are physically robust and do not depend on any epidemiology being correct.

So: score vegetables positively and confidently, but for energy density, fibre, potassium and micronutrient density, not for sulforaphane. And do not let phytochemical stories produce large score differentials between vegetables — the between-vegetable evidence is much weaker than the vegetables-versus-no-vegetables evidence.

A representative effect size for the cohort literature: Pollock RL, “The effect of green leafy and cruciferous vegetable intake on the incidence of cardiovascular disease: A meta-analysis,” JRSM Cardiovasc Dis 2016;5:2048004016661435. 8 cohort studies, pooled RR 0.842 (95% CI 0.753–0.941), p=0.002 — a 15.8% relative reduction in CVD incidence. Verified. Note this is a small meta-analysis in a minor journal; treat the magnitude loosely. Absolute translation: for a adult male, 10-year ASCVD risk is typically well under 1%. A 16% relative reduction on <1% absolute risk is a rounding error over ten years. The case for vegetables in this subject is about the lifetime trajectory and the body-composition goal, not about near-term event prevention — and it is important to be honest that a scoring system cannot promise the adult event reduction.

4.1 Cruciferous vegetables

Members: broccoli, broccolini, cauliflower, cabbage, napa cabbage, brussels sprouts, kale, collards, bok choy, arugula/rocket, watercress, radish, daikon, kohlrabi, turnip, mustard greens, horseradish, wasabi.

The glucosinolate/sulforaphane story, and where it stops

Crucifers store glucosinolates (glucoraphanin in broccoli). When tissue is damaged, the plant enzyme myrosinase hydrolyses these to isothiocyanates, principally sulforaphane. Sulforaphane is a potent Nrf2 activator, inducing phase-II detoxification enzymes (GST, NQO1). This is well-characterised, real biochemistry.
The critical practical fact: myrosinase is a protein and is denatured by heat. Boiling destroys it and also leaches glucosinolates into the water. Steaming lightly (~3–4 min) retains substantially more. Frozen broccoli is typically blanched before freezing, which inactivates myrosinase — so frozen crucifers deliver glucosinolates but little conversion capacity. Human gut bacteria (particularly Bacteroides spp.) have myrosinase-like activity and partially rescue conversion, but the yield is substantially lower and highly variable between individuals. Published comparisons consistently find raw > lightly steamed ≫ boiled/frozen for sulforaphane yield, with order-of-magnitude differences at the extremes. I did not verify a specific percentage yield table in this session — treat the ordering as well-established and any specific % as unverified. Adding a small amount of raw crucifer (mustard seed, radish, arugula) to cooked broccoli restores myrosinase and measurably increases sulforaphane yield — a real, cheap trick.
What has actually been tested in humans: the Qidong, China broccoli-sprout beverage trials (Johns Hopkins/Kensler group) in a population with high aflatoxin and airborne-benzene exposure. These demonstrated increased urinary excretion of detoxification conjugates (aflatoxin-N7-guanine, benzene mercapturic acid) — i.e. biomarker endpoints of enhanced conjugation and excretion, not cancer incidence. They are excellent proof-of-mechanism in humans. They are not evidence that broccoli prevents cancer.
Cohort evidence for crucifers and cancer is directionally favourable but modest and heterogeneous, and shares all the confounding problems in §4.0.

Goitrogens

Glucosinolate breakdown products (thiocyanates) competitively inhibit iodide uptake by the thyroid. Realistic dietary risk: essentially zero in an iodine-replete person. The documented cases involve extreme intakes (there is a well-known case report of hypothyroid myxoedema coma in a woman consuming roughly 1–1.5 kg of raw bok choy daily for months) or iodine-deficient populations. A person eating normal amounts of cooked crucifers, especially in an iodised-salt country, has no realistic concern. Cooking further reduces goitrogen activity.

Tier: beneficial. Score: +6. Confidence: moderate (high for nutrient density and low energy density; low for the phytochemical-specific claims).

4.2 Alliums

Members: garlic, onion, red onion, shallot, leek, scallion/green onion, chive, ramp.

Allicin is not present in intact garlic. Crushing/chopping brings alliin into contact with alliinase, generating allicin, which is itself unstable and degrades within minutes to hours into diallyl sulfides and other organosulfur compounds. Heating garlic immediately after chopping inactivates alliinase and largely prevents allicin formation; letting crushed garlic stand ~10 minutes before heating preserves more downstream sulfur compounds. This is real chemistry, and it means “garlic” in a cooked meal is a chemically different exposure from raw crushed garlic, and “garlic powder” is different again.
Blood pressure: Ma X, Zhang H, Jia J, “The effect of garlic on the lowering of blood pressure in the patients with hypertension: an updated meta-analysis and trial sequential analysis,” Asian Biomed 2025;19(3):131–140. 12 reports, 405 hypertensive patients treated with garlic derivatives. Diastolic BP mean difference −4.256 mmHg (95% CI −5.99 to −2.5x), with a significant systolic reduction as well. Verified.
The caveat that matters most: these trials use aged garlic extract or standardised garlic supplements at doses equivalent to multiple cloves per day, in hypertensive patients. The subject here is normotensive. Antihypertensive interventions produce much smaller effects in normotensive people, and a clove or two of garlic distributed through a meal is not the tested exposure.
Quality flag on this literature: one recent garlic-BP meta-analysis in Prostaglandins Other Lipid Mediat (2024) was retracted in 2026 (verified). The garlic supplement literature has a meaningful paper-mill and low-quality-journal problem; treat pooled estimates with caution.
Onions are the major dietary source of quercetin in Western diets. Quercetin supplement trials have been broadly unimpressive.

Verdict: alliums are a genuinely good flavour-per-calorie and flavour-per-milligram-of-sodium tool, which is their most defensible benefit for this subject. The pharmacological claims are supplement-dose claims. Tier: beneficial. Score: +2 (fresh garlic/onion/shallot/leek), +1 (garlic powder/onion powder). Confidence: moderate.

4.3 Leafy greens, and the nitrate/nitrite paradox

Members: spinach, kale, romaine, mixed greens, spring mix, chard, collards, mustard greens, watercress, arugula, butter lettuce, iceberg (much lower density).

Dietary nitrate — the active adult-relevant part, told honestly

Pathway: dietary nitrate (NO₃⁻) → reduced to nitrite (NO₂⁻) by oral commensal bacteria → further reduced to nitric oxide (NO) in the acidic/hypoxic stomach and tissues. This is the entero-salivary nitrate-nitrite-NO pathway and it is well established. Antibacterial mouthwash abolishes it — a striking and replicated finding.
Effects: modest blood pressure reduction (on the order of a few mmHg systolic at 5–8 mmol nitrate), reduced O₂ cost of submaximal exercise, and improvements in time-trial performance in endurance events. Typical ergogenic doses in the literature are ~6–8 mmol nitrate, which is roughly 500 mL of beetroot juice or a large serving (~200 g) of spinach/rocket/beetroot.
Sceptical counterweight, verified: Eroglu MN et al., “The Effects of Beetroot Juice Supplementation on Performance and Fatigue During Single and Repeated Sprints: A Systematic Review and Meta-Analysis,” Nutrients 2026;18(15):2513. 23 randomised trials, 401 participants. Improved time to peak power in 4 Wingate studies (SMD −0.92) and handgrip strength (SMD 0.40), but no significant effect on countermovement jump, peak or mean power, 10 m or 20 m sprint, RPE, heart rate, or blood lactate. Certainty low to very low. Authors conclude the evidence “does not support its routine use to enhance short-distance or repeated-sprint performance.” Verified.
Practical translation for a recreational sport player: recreational sport is a repeated-sprint/jump sport, which is exactly the domain where the beetroot meta-analysis is null. The nitrate literature’s real strength is in endurance time-trial and O₂-economy contexts. So the ergogenic case here is weak, and the BP case is modest. Score greens well for other reasons.

The nitrate/nitrite paradox — the chemistry that resolves it

This is the single most confusing apparent contradiction in the whole food-scoring problem, and the resolution is clean:

Vegetables supply the overwhelming majority of dietary nitrate — typically ~80%+ of total intake, far more than cured meats — yet vegetable nitrate is associated with benefit while cured-meat nitrite is classed as a carcinogenic mechanism. Three reasons, all chemical rather than rhetorical:

Nitrate is not the hazard; N-nitroso compounds are. Nitrite under acidic conditions can nitrosate secondary amines and amides to form N-nitrosamines and N-nitrosamides, which are the actual genotoxins. Nitrate itself is inert until reduced.
Vegetables co-deliver potent nitrosation inhibitors. Ascorbate (vitamin C) and polyphenols out-compete amines for nitrosating species, reducing nitrite to NO instead of forming nitrosamines. A leafy green delivers nitrate packaged with the specific chemistry that blocks the harmful pathway. This is also precisely why cured meat manufacturers add sodium ascorbate/erythorbate to cured products — it is a legally mandated or standard nitrosamine-suppression measure in many jurisdictions.
Meat supplies the reaction partners that vegetables lack: abundant secondary amines from protein, plus heme iron, which catalyses endogenous nitrosation and generates nitrosyl-heme. High-temperature cooking of nitrite-cured meat additionally forms nitrosamines directly (this is the specific reason bacon has been a regulatory focus).

So the paradox is not a paradox: the same nitrogen atom in a different chemical neighbourhood has opposite consequences. Any scoring system must therefore treat vegetable nitrate as positive/neutral and cured-meat nitrite as part of the processed-meat penalty — and, importantly, must not treat “celery powder cured” as an improvement (celery powder is simply a nitrate source that gets bacterially reduced to nitrite; see the proteins section).

Oxalate

Spinach, chard, and beet greens are high-oxalate. Relevance: calcium-oxalate kidney stones. Realistic risk for a healthy male with no stone history: low. The primary determinants of stone risk are fluid intake, urine volume, and calcium intake (counterintuitively, higher dietary calcium reduces stone risk by binding oxalate in the gut). The realistic scenario that causes trouble is very large daily spinach smoothies plus low fluid intake. Cooking and draining reduces oxalate substantially. Not worth a score penalty; worth knowing.

Lutein/zeaxanthin

Concentrated in kale, spinach, and other dark greens. The AREDS2 trial supports lutein/zeaxanthin for age-related macular degeneration progression in people who already have it — irrelevant to a adult. Absorption requires co-ingested fat, which is a genuine argument for oil in a salad.

Tier: beneficial. Score: dark leafy greens +7; iceberg/light lettuce +2. Confidence: moderate-high (driven by nutrient density and energy density, not phytochemistry).

Beets

High nitrate, high folate, and containing betalains. Same nitrate caveats. Beets are relatively sugar-dense for a vegetable (~9 g carbohydrate/100 g). Score +4.

4.4 Tomatoes and lycopene

Bioavailability: lycopene bioavailability is substantially higher from cooked/processed tomato (paste, sauce, canned) than raw, because heat disrupts the cell matrix and isomerises trans-lycopene to more absorbable cis forms. Co-ingested fat further increases absorption several-fold. This is one of the few cases where processing genuinely improves a food’s nutritional delivery, and it means canned tomatoes and tomato paste in a prepared meal are a feature, not a compromise.
Prostate cancer — the honest arc: early cohort work (Giovannucci and colleagues, HPFS, 1990s) reported inverse associations between tomato-product intake and prostate cancer, generating enormous enthusiasm. Subsequent larger analyses and pooled reviews weakened the signal considerably. The FDA reviewed the evidence for a qualified health claim and concluded there was very limited or no credible evidence for lycopene and prostate (or other) cancers, rejecting the strong claims. The broader collapse of the antioxidant-supplement hypothesis (SELECT: selenium and vitamin E showed no benefit, with vitamin E showing a statistically significant increase in prostate cancer) is directly relevant context: single-antioxidant mechanisms have repeatedly failed when tested.
Honest verdict: tomatoes are a good, low-calorie, potassium-rich, vitamin-C-rich vegetable. The lycopene-prostate story is much weaker than its cultural footprint and should not drive scoring. Tomato products also carry a sodium flag: canned sauces and pastes are frequently heavily salted.

Tier: beneficial. Score: +4 (fresh or plain canned/paste); flag sodium separately for prepared sauces. Confidence: moderate.

4.5 Mushrooms

Members: white/button, cremini, portobello, shiitake, oyster, maitake, enoki, king trumpet, porcini.

The sodium angle is the most actionable finding here and it is directly relevant to this subject. Mushrooms are the richest common culinary source of free glutamate and 5’-ribonucleotides (guanylate), which act synergistically to produce umami. Substituting finely chopped mushrooms for a portion of the meat in a dish (“the blend”) has been repeatedly shown in sensory and product-development research to allow substantial sodium reduction (on the order of 25–30% in tested formulations) with no loss of perceived flavour intensity or acceptability. For a person actively reducing sodium, mushrooms are a functional ingredient, not just a vegetable. Specific % sodium-reduction figures are from the mushroom-blend product literature and are unverified in this session; the qualitative finding is well established.
Ergothioneine: mushrooms are by far the dominant dietary source of this sulfur amino acid derivative. Humans have a dedicated transporter (OCTN1/SLC22A4) that concentrates it in tissues with high oxidative stress — the existence of a specific transporter is a genuinely interesting argument that it is physiologically important rather than incidental. But there are no outcome RCTs; it remains a hypothesis with unusually good circumstantial support.
Vitamin D: mushrooms contain ergosterol, which converts to vitamin D2 on UV exposure. UV-treated mushrooms can contain very large amounts. D2 is less potent than D3 at raising and maintaining 25(OH)D. Untreated supermarket mushrooms grown in the dark contain negligible vitamin D — so “mushrooms are a vitamin D source” is only true for explicitly UV-exposed product.
Cohort evidence: mushroom consumption has been examined in pooled analyses for gastric cancer (Ba DM et al., Eur J Cancer Prev 2023;32(3):222–228, Stomach Cancer Pooling Project plus meta-analysis — verified as existing; I did not verify its effect estimate in this session) and in various mortality analyses. Directionally favourable, small effect sizes, standard confounding caveats.
Agaritine in Agaricus bisporus (button/cremini/portobello) is a hydrazine derivative that is mutagenic in vitro and carcinogenic in some rodent assays at high doses. Cooking substantially degrades it. There is no human epidemiological signal of harm from mushroom consumption, and the cohort evidence points the other way. Not a real concern; do not penalise.

Tier: beneficial. Score: +5. Confidence: moderate (high for the sodium-substitution and energy-density case, low for ergothioneine).

4.6 Nightshades and inflammation

The claim: nightshade vegetables (tomato, potato, eggplant, all peppers including paprika and chili, goji) contain glycoalkaloids (solanine, tomatine, capsaicin is separate) that provoke inflammation and worsen arthritis.

The evidence: essentially none. There is:

No randomised controlled trial demonstrating that nightshade elimination improves arthritis, inflammatory markers, or any other endpoint. The claim traces primarily to Norman Childers, a horticulturist, who in the 1980s promoted the hypothesis based on self-reported questionnaire responses from a self-selected membership organisation — an uncontrolled, self-selected, unblinded survey, which is the weakest possible design and is essentially an anecdote-collection exercise.
Glycoalkaloids are genuinely toxic at high doses (green/sprouted potatoes — see the carbohydrate section), but the doses in normal nightshade vegetables are far below any toxic threshold, and toxicity when it occurs is acute gastrointestinal/neurological, not inflammatory-arthritic.
Meanwhile tomatoes and peppers are in the same epidemiological bucket as other vegetables, i.e. associated with better outcomes.

Verdict: this is a wellness claim with no supporting evidence, and I am comfortable saying so plainly. The one legitimate carve-out is that a small number of individuals have genuine IgE-mediated allergy or non-allergic intolerance to specific nightshades — a real but individual phenomenon that says nothing about population-level scoring. Do not apply any nightshade penalty. Score: 0 adjustment for “being a nightshade.”

Capsaicin, separately

Real pharmacology (TRPV1 agonist), real acute thermogenic and appetite effects in trials — and trivial in magnitude. Meta-analyses of capsaicin/capsinoid supplementation report increases in energy expenditure on the order of tens of kcal/day, which is inside the measurement noise of free-living energy balance and is not a weight-loss strategy. Capsaicin’s genuine value for this subject is the same as herbs and alliums: flavour intensity without sodium. Score: +1.

4.7 Fermented foods and the microbiome

Wastyk 2021 — get the details right, because it is usually oversold

Wastyk HC, Fragiadakis GK, Perelman D, … Gardner CD, Sonnenburg JL, “Gut-microbiota-targeted diets modulate human immune status,” Cell 2021;184(16):4137–4153.e14. Verified.

Design: 17-week randomised prospective trial, n=18 per arm (high-fibre vs high-fermented-food), healthy adults, with deep multi-omic microbiome and immune profiling.
The primary outcome — cytokine response score — was UNCHANGED. Verified. This is routinely omitted when the study is cited.
High-fibre arm: increased microbiome-encoded glycan-degrading CAZymes, stable community diversity, and three distinct immunological trajectories that corresponded to baseline microbiota diversity.
High-fermented-food arm: steadily increased microbiota diversity and decreased inflammatory markers. Verified.
Limitations that must travel with this citation: n=18 per arm is very small; 17 weeks is short; all endpoints beyond the (null) primary are secondary/exploratory across a very large number of measured analytes; unblindable. It is a genuinely important and well-executed hypothesis-generating study, and it is the best evidence we have that fermented food does something distinct from fibre. It is not a demonstration of clinical benefit.
The counterintuitive and most interesting result is that fibre did not do what everyone expected and fermented food did — the opposite of the prior.

The practical problem that dominates everything: are the cultures alive?

For a shipped, cooked, prepared meal this is decisive and usually ignored:

Live cultures require the food to be raw/refrigerated and unpasteurised. Heat kills them at ordinary cooking temperatures.
Cooked into a meal → dead: miso in a hot soup, kimchi in fried rice, sauerkraut on a hot dish, tempeh that has been cooked (tempeh is always cooked), sourdough (baked — the culture is dead in every loaf of bread ever made).
Shelf-stable → almost always pasteurised: canned/jarred sauerkraut and pickles on an unrefrigerated shelf, most commercial kombucha at retail (variable), soy sauce, vinegar, most miso that is shelf-stable.
Possibly live: refrigerated raw kimchi, refrigerated unpasteurised sauerkraut, yogurt/kefir with “live and active cultures,” refrigerated miso added off the heat.
In a meal-delivery product that is cooked and reheated by the consumer, the probability that any fermented component still contains viable organisms is low.

However — and this matters — fermentation produces benefits that survive the death of the organism: the microbes have already pre-digested the substrate, reducing phytate and FODMAPs, liberating amino acids and vitamins (natto’s vitamin K2, tempeh’s improved protein digestibility), and generating bioactive peptides and organic acids. Sourdough’s lower glycaemic response and improved mineral bioavailability are properties of the dough, not of live yeast in the finished bread. So a pasteurised fermented food is not worthless; it is just not a probiotic.

The sodium tension — a genuine conflict for this subject

This is where the fermented-food recommendation collides head-on with the sodium-reduction goal. Approximate sodium loads:

Soy sauce: ~900–1,000 mg sodium per tablespoon (~15 mL). Reduced-sodium versions ~500–600 mg.
Miso paste: ~600–900 mg per tablespoon.
Kimchi: ~500–700 mg per 100 g serving.
Sauerkraut: ~400–700 mg per 100 g.
Olives and pickles: 300–900 mg per serving depending on cure. (These are standard composition-table values, cited from general nutrient-database knowledge; I did not re-verify each figure in this session — treat as approximate but reliable to within ~20%.)

Against a target of reducing sodium (the US DGA limit is 2,300 mg/day; AHA’s ideal is 1,500 mg), two tablespoons of soy sauce is roughly 80% of the AHA ideal daily intake in a single condiment. For this subject, the sodium cost of most fermented savoury foods exceeds their plausible microbiome benefit, especially since the cultures are probably dead by the time the adult eats them.

Practical resolution: the fermented foods worth eating for this subject are the low-sodium, reliably-live ones — yogurt and kefir (covered in the proteins section) — plus vinegar (essentially sodium-free, and there is modest RCT evidence for acetic acid blunting postprandial glycaemia). Kimchi and sauerkraut should be scored net-neutral: real fermentation value, offset by a real sodium load. Soy sauce and miso used as primary seasoning should be scored negative on sodium grounds, which is a conclusion about salt, not about fermentation.

Tier: contested (benefit real but small and probably absent in cooked products; sodium cost real). Scores: vinegar +2; refrigerated raw kimchi/sauerkraut 0; cooked kimchi/sauerkraut/tempeh in a dish +1 (for the food itself); miso as seasoning −1; soy sauce −2; kombucha 0. Confidence: moderate.

4.8 Herbs and spices at culinary doses — do the arithmetic

This is the section where a scoring system is most likely to go wrong, because the popular literature is enormous and the doses are absurd.

Turmeric — the arithmetic, spelled out

Turmeric root is up to ~5% curcuminoids by weight (verified: Nelson et al. state “up to ∼5%”; commonly cited working figure is ~3%).
A meal containing a generous 1 teaspoon of ground turmeric ≈ 3 g delivers ~90–150 mg of curcuminoids. A realistic meal-delivery portion is more like ¼–½ teaspoon, i.e. ~25–75 mg curcuminoids.
Clinical trials use 500–2,000 mg of curcumin per day, and almost always with piperine (black pepper extract) or a lipid/nanoparticle formulation, because unformulated curcumin’s oral bioavailability is famously near-zero — it is poorly absorbed, rapidly glucuronidated and sulfated, and rapidly excreted.
The gap is therefore roughly 10- to 80-fold in dose, before accounting for the formulation difference — and the formulation difference alone is often cited as a further 10–20× in systemic exposure. Combined, a meal’s turmeric plausibly delivers two to three orders of magnitude less systemic curcumin than a trial dose.

And the trial doses themselves don’t work

Nelson KM, Dahlin JL, Bisson J, Graham J, Pauli GF, Walters MA, “The Essential Medicinal Chemistry of Curcumin,” J Med Chem 2017;60(5):1620–1637. Verified. Key verified findings:

Curcumin is classified as both a PAINS (pan-assay interference compound) and an IMPS (invalid metabolic panacea) candidate — it produces false positives across assay types via aggregation, membrane disruption, redox cycling, metal chelation, and fluorescence interference.
“The likely false activity of curcumin in vitro and in vivo has resulted in >120 clinical trials of curcuminoids against several diseases.”
“No double-blinded, placebo controlled clinical trial of curcumin has been successful.”
Conclusion: curcumin is “an unstable, reactive, nonbioavailable compound and, therefore, a highly improbable lead.”

Verdict on turmeric: overhyped by a very large margin, and this is not a close call. Even the supplement doses lack a single successful properly-controlled trial per a review in a top medicinal chemistry journal. The dose in a meal is orders of magnitude below even that. Score turmeric at 0.

The others, briefly

Ginger — the best-evidenced of the culinary spices, and the evidence is specifically for nausea (pregnancy, post-operative, chemotherapy-induced), typically at ~1 g/day of ginger powder, which is achievable-ish in a heavily gingered dish. DOMS/muscle-soreness trials are small and inconsistent. Anti-inflammatory claims at culinary dose are not established. Score +1 (achievable dose, but a modest and narrow effect).
Cinnamon — glycaemic trials are mixed and mostly in people with T2D; effects on fasting glucose are small and heterogeneous, and largely absent in normoglycaemic people. Important safety detail: cassia cinnamon (the common supermarket cinnamon) contains coumarin, which is hepatotoxic at sustained high intake; EFSA set a coumarin TDI of 0.1 mg/kg bw/day, which a person taking large cassia doses daily can exceed. Ceylon (“true”) cinnamon is very low in coumarin. At the fraction-of-a-teaspoon dose in a meal, neither the benefit nor the risk is meaningful. Score 0.
Rosemary, oregano, thyme, basil, parsley, cilantro — very high antioxidant capacity per gram in vitro, negligible grams consumed. Rosemary extract genuinely retards lipid oxidation in cooked meat, which is a real, mechanistically-sound food-chemistry benefit (it reduces HCA and lipid-oxidation-product formation during cooking), but it is a small effect on an outcome that itself isn’t well linked to human harm. Score 0 to +1.
Black pepper — piperine’s role is real but it is a bioavailability enhancer (CYP3A4 and glucuronidation inhibitor), not a therapeutic agent. Notably, the same mechanism means piperine can affect drug metabolism. Score 0.
Cumin, coriander, paprika, chili powder, garlic powder, onion powder — flavour, negligible pharmacology at these doses.

The actual reason herbs and spices deserve a positive score

They enable sodium reduction. This is not a hand-wave: it is the single most evidence-supported role of culinary herbs and spices, it has been demonstrated in behavioural trials (spice-focused cooking education reduced measured sodium intake in free-living adults), and it maps directly onto this subject’s explicit goal. A heavily herbed and spiced dish can achieve high flavour intensity at low sodium; a blandly seasoned dish reaches palatability through salt. Score the presence of a diverse spice/herb profile as a small positive (+1 to +2) for this reason and this reason only, and do not attribute it to phytochemistry.

4.9 Other vegetables, briefly

Food	Note	Score
Carrots	Beta-carotene (bioavailability improves with cooking + fat), fibre, very low energy density. Note the CARET/ATBC trials found β-carotene supplements increased lung cancer in smokers — a permanent lesson about isolating phytochemicals. Whole carrots are fine.	+4
Bell peppers	Exceptionally high vitamin C (higher than citrus by weight), low calorie	+4
Butternut/winter squash	Carotenoids, potassium, more starch than most vegetables but still low energy density	+4
Zucchini/summer squash	Very low energy density, low nutrient density. Good volume filler	+3
Asparagus	Folate, inulin-type fructans (prebiotic)	+4
Green beans	Solid, unremarkable, good volume	+3
Peas	Higher protein and starch than most vegetables (~5 g protein/100 g); closer to a legume	+4
Artichoke	Among the highest fibre of any vegetable (~5 g/100 g), inulin	+5
Eggplant	Low calorie, but acts as an oil sponge — the fat added in cooking often dominates	+2
Cucumber, celery	Near-zero energy density, minimal nutrients; celery contributes some nitrate and potassium	+2
Corn (as vegetable)	Starchy; fine, but should be scored as a starch not a vegetable	+1
Seaweed/nori	Iodine. Nori is modest (~16–43 µg per sheet) but KELP/kombu is extreme — a single gram can exceed the 1,100 µg/day UL by many-fold, and chronic excess iodine causes thyroid dysfunction. Nori in sushi: fine. Kelp as a featured ingredient or supplement: genuine caution. Specific µg figures are standard composition values, unverified in this session.	nori +2; kelp-heavy −1
Olives	Good fat, polyphenols — but cured, so meaningfully high sodium (~250–450 mg per 10 olives)	+1
Pickles	Sodium load with little nutritional return	−1
Berries	Anthocyanins; the Cassidy/Rimm cohort work (Nurses’ Health Study II) reported inverse associations between anthocyanin-rich berry intake and myocardial infarction in young/middle-aged women. Design: prospective cohort; I did not verify the effect estimate or n in this session — mark unverified. Berries are the lowest-sugar, highest-fibre fruit and the best fruit choice for this subject	+5
Other whole fruit	Fibre + water + micronutrients; whole fruit intake is consistently associated with lower T2D risk while fruit juice is associated with higher — an unusually clean natural contrast that argues the matrix matters	+4
Fruit juice	Sugar without the matrix; counts toward the added-sugar-equivalent burden even when “no added sugar”	−2
Dried fruit	Concentrated sugar, easy to over-consume, but retains fibre and potassium; often has added sugar and sulfites	0

Machine-readable table — Vegetables, Fungi, Fermented, Herbs & Spices

Section 5 — Contaminants, Cooking Chemistry, and Packaging

Framing note that governs this entire section: almost every item here is a hazard, not a quantified risk. A hazard is “this molecule can cause harm at some dose.” A risk is “at the dose you actually get, your probability of harm changes by X.” Wellness content systematically conflates the two, and regulatory agencies systematically communicate hazard (because that is their job) in language that reads as risk. The discipline applied below is: find the exposure number, find the reference number, compute the ratio, and state the absolute risk.

Second framing note: for a adult cutting visceral fat, the total contribution of everything in this section to the adult’s health outcomes is small compared to energy balance, protein adequacy, sodium, and fibre. This section should generate small score adjustments, with two exceptions flagged at the end.

5.1 Acrylamide (roasted/fried/baked starches)

What it is

Formed by the Maillard reaction between asparagine and reducing sugars above ~120 °C in low-moisture, starchy foods. Highest in: French fries, potato chips, roasted potatoes, toast/crisp bread, breakfast cereals, roasted coffee, biscuits. Not formed in boiling or steaming (water caps the temperature at 100 °C). Discovered in food in 2002 (Swedish NFA), which is why it feels “new.”

Hazard basis

IARC Group 2A (“probably carcinogenic to humans”). Genotoxic and carcinogenic in rodents — the glycidamide metabolite forms DNA adducts. EFSA’s CONTAM Panel (2015) concluded acrylamide in food “potentially increases the risk of developing cancer for consumers in all age groups.” Verified via EFSA.

What the human data actually show — this is the part that gets omitted

EFSA’s own 2015 conclusion states that human studies “have provided limited and inconsistent evidence of increased risk of developing cancer.” Verified. EFSA’s position is explicitly precautionary and animal-derived.
Filippini T et al., “Dietary Acrylamide Exposure and Risk of Site-Specific Cancer: A Systematic Review and Dose-Response Meta-Analysis of Epidemiological Studies,” Front Nutr 2022;9:875607. Design: systematic review + dose-response meta-analysis of prospective/epidemiological studies. 16 studies, 1,151,189 participants. Mean estimated dietary acrylamide dose across studies: 23 µg/day. Result: no association between highest vs lowest dietary acrylamide exposure and ANY site-specific non-gynaecological cancer examined. Verified.
Pelucchi C et al., “Dietary acrylamide and cancer risk: an updated meta-analysis,” Int J Cancer 2015;136(12):2912–2922. Result: no meaningful associations for most cancers. Continuous per-10-µg/day estimates all between 0.95 and 1.03, none significant. The only borderline signal was kidney cancer, RR 1.20 (95% CI 1.00–1.45) — i.e. the lower bound sits exactly on the null. Among never-smokers, borderline signals for endometrial (RR 1.23, 1.00–1.51) and ovarian (RR 1.39, 0.97–2.00) — both gynaecological, irrelevant to this subject. Verified.
Indirect corroboration: Gaesser GA et al., “Bread Consumption and Cancer Risk: Systematic Review and Meta-Analysis of Prospective Cohort Studies,” Curr Dev Nutr 2024;8(12):104501. 24 publications, 1,887,074 adults. Of 108 reported hazard ratios, 97 (79%) were either non-significant (86) or indicated LOWER cancer risk (11). Pooled: site-specific cancer HR 1.01 (0.89–1.14); total cancer mortality HR 0.90 (0.73–1.11). Whole-grain bread associated with lower colorectal cancer risk. Verified. Bread is a major acrylamide vector; if dietary acrylamide were an important human carcinogen at dietary doses, this is where it should have shown up in 1.9 million people. It did not.

Absolute risk translation

Typical adult dietary acrylamide exposure is ~0.4–0.5 µg/kg bw/day (~30–40 µg/day for an 80 kg adult; the meta-analytic mean was 23 µg/day). EFSA’s BMDL10 for neoplastic effects in rodents is on the order of 0.17 mg/kg bw/day (this specific BMDL figure is unverified recall from EFSA’s 2015 opinion — the EFSA topic page I fetched did not state it; treat the number as approximate, the margin-of-exposure conclusion below is what matters and is robust). That yields a margin of exposure in the low hundreds, which EFSA flags as “a concern” by its own MOE convention (<10,000 for a genotoxic carcinogen triggers concern). But the MOE convention is a precautionary bookkeeping device, not a risk estimate, and when the actual human outcome data are examined across >3 million person-observations, there is no detectable signal.

Verdict

Overstated at dietary exposure. The honest statement: acrylamide is a genuine rodent carcinogen with a plausible genotoxic mechanism, regulators are right to push industry to reduce it (the EU’s Regulation 2017/2158 benchmark levels are sensible), and there is no epidemiological evidence of harm to adults at real dietary doses. For scoring 800 meals: do not apply a meaningful acrylamide penalty. The reason to downgrade “French fries” is calories, fat, sodium, and displacement — not acrylamide. A meal with roasted potatoes should not be penalised for acrylamide at all.

Practical mitigations if one cares anyway: don’t over-brown (golden not dark), soak cut potatoes before roasting, store potatoes above 8 °C rather than refrigerated (cold storage raises reducing sugars → more acrylamide, a genuinely counterintuitive fact).

Score adjustment: 0 for acrylamide as such. Confidence: high that it is not a meaningful dietary risk for this subject.

5.2 Heterocyclic amines (HCAs) and polycyclic aromatic hydrocarbons (PAHs) — grilled/charred meat

Formation chemistry

HCAs: form when amino acids + sugars + creatine/creatinine react at high temperature. Requires muscle tissue (creatine is the limiting reagent) — this is why HCAs form in meat/fish and essentially not in plant foods, tofu, or eggs. Highest in well-done pan-fried, grilled, and barbecued meat. Formation rises steeply above ~150 °C and with time.
PAHs (incl. benzo[a]pyrene): form when fat drips onto flame/hot coals, producing smoke that deposits on the meat surface. This is a surface deposition phenomenon, so it’s about the cooking geometry, not the meat. Also present in smoked foods.

What the human evidence shows

NCI’s official position (verified): “population studies have not established a definitive link between HCA and PAH exposure from cooked meats and cancer in humans.” The limitation cited is exposure measurement — FFQs cannot capture doneness reliably, and individual NAT2/CYP1A2 acetylator genotype substantially modifies how HCAs are metabolically activated, so any true effect is heterogeneous across people.
The dose gap is the headline fact, and NCI states it explicitly (verified): rodent studies that produced tumours used doses “equivalent to thousands of times the doses that a person would consume in a normal diet.” This is a 3–4 order-of-magnitude extrapolation.
Some case-control studies (which are more prone to recall bias) report associations between well-done meat and colorectal/pancreatic/prostate cancer; prospective cohorts are weaker and more inconsistent. There are no federal guidelines for HCA/PAH intake, precisely because the human evidence doesn’t support setting one.

Where this interacts with processed/red meat

This matters for interpreting the IARC processed-meat classification (covered in the proteins section): HCAs/PAHs are one of four competing mechanistic candidates (alongside N-nitroso compounds, heme iron, and sodium), and the fact that none of them has been isolated as the operative mechanism in humans is itself informative about how strong the underlying signal is.

Verdict

Overstated, but not zero, and the mitigation is nearly free. The mechanistic case is real; the human dose-response case is not established. In practical terms, the difference between “grilled chicken” and “charred, blackened grilled chicken” is a real chemical difference and the second is worth a small penalty. The difference between “grilled” and “baked” is not worth penalising.

Score adjustment: grilled/barbecued meat 0; explicitly charred/blackened/burnt −1. Confidence: moderate. Free mitigations (NCI-endorsed, verified): avoid direct flame, flip frequently, microwave-precook before grilling, remove charred portions, don’t make gravy from drippings. A marinade (especially acidic/herb-containing) measurably reduces HCA formation.

5.3 Advanced glycation end products (AGEs)

The claim

Dietary AGEs — formed by browning/dry-heat cooking, high in roasted, grilled, fried, and broiled foods, especially high-fat animal foods — drive inflammation, insulin resistance, and vascular ageing via the RAGE receptor.

What the RCTs show — and this is better evidence than most of this section

Sohouli MH et al., “The impact of low advanced glycation end products diet on obesity and related hormones,” Sci Rep 2020;10:22194. Meta-analysis of 13 RCTs. Low-AGE vs high-AGE diets: BMI −0.30 kg/m² (95% CI −0.52 to −0.09, p=0.005); weight −0.83 kg (−1.55 to −0.10, p=0.026); leptin −19.85 ng/mL (−29.88 to −9.82, p<0.001); adiponectin +5.50 µg/mL (1.33 to 9.67). Heterogeneity high (I² 56–82%). Verified. Crucially, the authors’ own conclusion is that although the effects were statistically significant, “no clinical significance was observed.”
Sohouli MH et al., “The Impact of Low Advanced Glycation End Products Diet on Metabolic Risk Factors: A Systematic Review and Meta-Analysis of Randomized Controlled Trials,” Adv Nutr 2021;12(3):766–776. 13 RCTs. Low-AGE diets reduced insulin resistance (HOMA-IR) −1.204 (−2.057 to −0.358, p=0.006); fasting insulin −5.47 µU/mL (−9.72 to −1.23); total cholesterol −5.49 mg/dL (−10.22 to −0.75); LDL −6.26 mg/dL (−11.66 to −0.87). Verified.

Honest interpretation

This is the strongest interventional evidence in the whole contaminants section — actual RCTs with actual metabolic endpoints, not rodent extrapolation. The HOMA-IR and LDL effects are not trivial (an LDL reduction of ~6 mg/dL is larger than what an avocado a day produced in the HAT trial).

But there is a serious confounding problem baked into the design: you cannot construct a low-AGE diet without also changing the food. Low-AGE arms are systematically steamed/boiled/poached, lower in fried and roasted fatty animal food, and higher in vegetables and moisture-cooked dishes. The trials cannot separate “fewer AGEs” from “less fried fatty meat and more boiled vegetables.” Most trials are small, short, and unblindable. The observed effects are entirely consistent with “the low-AGE arm ate a better diet.”

Additionally, the dominant source of circulating AGEs in humans is endogenous formation from glycaemia, not diet; the fraction of dietary AGEs actually absorbed is modest (~10% order of magnitude, and the estimate is contested).

Verdict

Real but probably not independent. The right way to encode this is not an “AGE penalty” but the cooking-method signal it proxies: heavily fried and dry-roasted fatty animal foods score worse; steamed, poached, braised, and stewed preparations score better. That happens to be correct for several independent reasons (fat added, calorie density, sodium), so it should be scored once, as a preparation term, and not double-counted as a separate AGE penalty.

Score adjustment: heavy dry-heat/fried fatty animal preparations −1 (as part of a preparation term, not additive). Confidence: moderate.

5.4 Arsenic in rice

The facts, quantified

Rice hyperaccumulates arsenic because paddy flooding mobilises arsenite in soil. Inorganic arsenic (iAs) is the toxic, IARC Group 1 form; organic arsenic species (mostly in seafood) are essentially non-toxic. Only the inorganic fraction matters.
Brown rice has more than white rice, because arsenic concentrates in the bran. Verified quantification (Food Chem Toxicol 2020;141:111420): brown varieties 189 µg/kg total As vs white rice 132 µg/kg — roughly 1.4× higher. This is the single most counterintuitive item in this section: the “healthier” rice carries more arsenic.
Geographic ordering (well-replicated in the literature, though I did not verify specific per-origin means in this session — treat the ordering as solid and the exact numbers as unverified): US south-central grown rice (Arkansas, Louisiana, Texas, on former cotton land treated with arsenical pesticides) is highest; California rice is lower; basmati from India/Pakistan and jasmine from Thailand are generally among the lowest. Rice syrup/rice cakes/rice cereal concentrate it.
FDA has set an action level for inorganic arsenic in infant rice cereal (guidance issued August 2020) and completed an Arsenic in Rice and Rice Products Risk Assessment (2016). FDA’s consumer guidance is generic: eat a variety of foods. Verified that these exist; the specific ppb value of the action level was not stated on the page I fetched — it is 100 ppb by recall, marked unverified.

Absolute risk — the number that settles it

A published risk assessment (Int J Environ Res Public Health 2022;19:16460, Brazilian population) computed incremental lifetime cancer risk (ILCR) of 6.0 × 10⁻⁶ for brown rice consumption (and lower for white). Verified.

Translate that: 6 in 1,000,000 additional lifetime cancer risk. For context, an individual’s baseline lifetime risk of developing cancer is roughly 40% (400,000 per million). The arsenic-from-rice increment is therefore on the order of 0.0015% of the baseline — i.e. it changes lifetime cancer risk from ~40.0000% to ~40.0006%. Regulators treat 1×10⁻⁶ as a de minimis threshold, which is why the number is flagged as “high” in the paper’s own framing — but a regulatory de minimis threshold for involuntary population exposure is a completely different standard from personal decision-relevance.

Verdict

A real contaminant, correctly regulated, and irrelevant to this subject’s decisions at realistic intake. The one group for whom it genuinely matters is infants and young children (small body weight, rice cereal as a large share of total diet) — which is exactly where FDA set its action level. A adult male eating rice several times a week is not in that group.

Mitigations if desired, in order of effectiveness: cook rice in excess water (6:1) and drain it — removes roughly 40–60% of iAs (well replicated); rinse before cooking (~10–25%); choose basmati/jasmine/California over US-south-grown; vary the grain. Note the trade-off: excess-water cooking also leaches water-soluble B vitamins from enriched rice.

Score adjustment: 0 for arsenic. White vs brown rice should be scored on fibre/micronutrients, and the arsenic direction (brown worse) roughly cancels the small fibre advantage in significance terms — neither is worth much. Confidence: high.

5.5 Cadmium

Sources in a meal-delivery context, in rough order: cocoa/dark chocolate, sunflower seeds, spinach and other leafy greens, shellfish (especially crab/oyster hepatopancreas), organ meats, root vegetables from contaminated soil, rice.
Cadmium is a genuine cumulative nephrotoxin with a biological half-life in the human kidney measured in decades (~10–30 years), which is what makes it different from most dietary contaminants — there is no clearance you can rely on. It is IARC Group 1 (occupational inhalation).
The dominant driver of body cadmium burden in non-occupationally exposed people is smoking, not food. Smokers have roughly double the blood cadmium of non-smokers. This subject is an active adult; assuming non-smoking, dietary cadmium is the whole exposure and it sits below tolerable intake for typical diets.
The genuinely actionable case is dark chocolate and cacao, where Consumer Reports testing (2022) found several products exceeding California Prop 65 maximum allowable dose levels for cadmium and lead. Prop 65 MADLs are famously conservative (lead MADL 0.5 µg/day, which is far below any level associated with measurable harm in adults), so “exceeds Prop 65” is not the same as “unsafe” — but repeated daily large servings of dark chocolate is the one realistic dietary pattern where cadmium is worth a thought.
Score adjustment: 0 in general; −1 only for meals featuring large dark-chocolate/cacao quantities as a primary component. Confidence: moderate.

5.6 Lead — spices and protein powders

Spices

This is the one heavy-metal issue in this section with a documented, large, real-world exposure event, and it deserves to be separated from the general noise.

Turmeric adulteration with lead chromate is a documented practice — lead chromate is a bright yellow pigment added to enhance colour, primarily documented in turmeric from certain regions of Bangladesh and India. This has been linked to elevated blood lead in consumers and traced through the supply chain in published environmental-health investigations. Unlike most items in this section, this is not a trace-contaminant margin issue; it is deliberate adulteration producing exposures that matter.
FDA and state health departments (notably NYC DOHMH) have issued multiple recalls of specific imported spice brands for lead. Elevated blood lead in children has been traced to imported spices in several US investigations.
However: this is a supply-chain provenance issue, not a property of turmeric. Spices sold by regulated US retailers and used by commercial meal-delivery kitchens buying from major distributors are a low-risk channel. And the quantity of turmeric in a meal is a fraction of a gram.
Score adjustment: 0. The risk attaches to bulk-imported unregulated spice, not to a meal containing turmeric. It is worth knowing, not worth scoring.

Protein powders

The widely-cited findings come from the Clean Label Project (2018 and 2024 reports) and a Consumer Reports 2010 investigation. Both found detectable lead, cadmium, arsenic, and BPA across many protein powder products, with plant-based and chocolate-flavoured products testing higher.
Critical methodological caveat, and it is not a minor one: the Clean Label Project is an advocacy nonprofit, its reports are not peer-reviewed, its methodology and thresholds have been substantively criticised by toxicologists, and it benchmarks against Prop 65 MADLs (lead 0.5 µg/day) rather than against any health-based tolerable intake. Most “failures” are of the Prop 65 threshold, not of a toxicological threshold. Plant proteins test higher for the mundane reason that plants take up soil minerals; cocoa tests higher because cacao accumulates cadmium.
Honest position: detectable ≠ harmful, the analytical chemistry is probably right and the risk framing is probably inflated. A daily protein-shake habit using a plant/chocolate powder is the one pattern where it’s reasonable to prefer a third-party-tested (NSF Certified for Sport, Informed Sport) product — which is worth doing anyway for banned-substance reasons in an active adult.
Score adjustment: 0 to −1 for meals built around unspecified protein powder. Confidence: low (evidence base is advocacy-grade).

5.7 Microplastics and phthalates from plastic meal trays — the one the subject specifically asked about

Every meal in this dataset ships in a plastic tray, and many are microwaved in that tray. So this is the exposure with the highest contact frequency in the whole analysis. Here is what is actually known versus speculated.

KNOWN (measured)

Hussain KA et al., “Assessing the Release of Microplastics and Nanoplastics from Plastic Containers and Reusable Food Pouches: Implications for Human Health,” Environ Sci Technol 2023;57(26):9782–9792. Design: laboratory migration study using DI water and 3% acetic acid as food simulants; polypropylene containers and polyethylene pouches. Verified.
- Microwave heating produced the highest release of any scenario tested (vs refrigeration or room-temperature storage).
- Some containers released as many as 4.22 million microplastic and 2.11 billion nanoplastic particles from just 1 cm² of plastic within 3 minutes of microwaving. Verified.
- Refrigeration and room-temperature storage over six months also released millions to billions of particles.
- Polyethylene pouches released more than polypropylene containers.
- Exposure modelling: highest estimated daily intake was 20.3 ng/kg·day (infants, microwaved water) and 22.1 ng/kg·day (toddlers, microwaved dairy from PP containers). Verified.
- In vitro: extracted particles killed 76.7% and 77.2% of HEK293T human embryonic kidney cells at 1,000 µg/mL after 48 and 72 h. Verified.

Now do the arithmetic the headlines never do. “4.22 million particles” is a terrifying number and “20.3 ng/kg/day” is the same finding expressed as mass. For an 80 kg adult at the modelled worst case, that is ~1.6 µg/day ≈ 0.6 mg/year ≈ about 1/50,000th of a gram per year. The in vitro cytotoxicity was demonstrated at 1,000 µg/mL — a concentration roughly six orders of magnitude above anything implied by the exposure model. Particle counts are large because nanoparticles have negligible individual mass; the count is the right metric only if the mechanism of harm is count-dependent (surface area, cellular uptake), which is precisely what is unknown.

Also KNOWN (measured, human)

Zota AR, Phillips CA, Mitro SD, “Recent Fast Food Consumption and Bisphenol A and Phthalates Exposures among the U.S. Population in NHANES, 2003–2010,” Environ Health Perspect 2016;124(10):1521–1528. Design: cross-sectional, n=8,877, NHANES, 24-h dietary recall vs urinary metabolites. Verified.
- Dose-response relationship between fast food intake and phthalate exposure (p-trend <0.0001), but NOT with BPA.
- Highest consumers (≥34.9% of total energy from fast food) had 23.8% (11.9–36.9%) higher ΣDEHP metabolites and 39.0% (21.9–58.5%) higher DiNP metabolites than non-consumers. Verified.
- Notably, the association was driven by grain and meat items, and by fat content — consistent with lipophilic plasticiser migration during processing/handling, and pointing at industrial food handling equipment (PVC tubing, conveyor belts, gloves) as much as the final package.
So: packaged/processed food demonstrably raises measured phthalate body burden in humans. That is a real, replicated, human-biomarker finding, and it is the strongest fact in this subsection. A meal-delivery diet is, from a phthalate standpoint, structurally similar to the “high fast food” exposure category — the food is industrially processed and handled, then packaged.

SPECULATED (not established)

That these exposures cause disease in adults at current levels. Microplastics have been detected in human blood, placenta, lung, liver, and (in a widely-reported 2024 NEJM study) carotid atheroma with an association to cardiovascular events — but that finding is observational, small, and has substantial reverse-causation and detection-bias concerns (I did not verify the NEJM carotid plaque study’s design or n in this session — treat it as unverified recall). There is no human RCT and there never will be.
That nanoplastic particle counts translate to biological dose. Unknown.
Phthalates are anti-androgenic in animals and there is human epidemiology on reproductive endpoints, but the exposures in question are far below the doses producing effects in animals, and the human literature is inconsistent.

Verdict, and the practical recommendation that actually follows

This is the item in this entire section where I would say the concern is legitimate, the magnitude is unknown, and the mitigation is essentially free — which is an unusual combination and is what makes it worth acting on despite weak evidence.

The measured fact is unambiguous: microwave heating is the single highest-release scenario for particle migration, and heat + fat + acid is the worst combination for plasticiser migration.
The mitigation is: transfer the meal to a ceramic or glass dish before microwaving. Cost: one dish. Expected benefit: unquantified but strictly positive, eliminating the highest-release step in the chain.
The mitigation is not: avoiding meal delivery. The phthalate signal in NHANES was driven substantially by industrial processing upstream of the package, so the tray is only part of it, and the nutritional quality of the meal matters vastly more to this subject’s outcomes.

Score adjustment: this is a CONSTANT across all ~800 meals (they all ship in plastic trays), so it should NOT be a per-meal score term — it cancels out and adds only noise. It belongs as a single standing recommendation attached to the analysis, not a per-meal penalty. The only defensible per-meal differentiation would be if some meals are designed to be microwaved in a film-sealed tray with high fat and acid content (e.g. a tomato-and-cheese dish) versus a dry low-fat one — a distinction the ingredient lists probably cannot support reliably.

5.8 BPA and BPS in can linings

Where it appears in this dataset

Canned tomatoes, canned beans, canned coconut milk, canned tuna, canned corn — all common in prepared meals.

The regulatory split, stated accurately, because the agencies genuinely disagree

FDA (verified): “BPA is safe at the current levels occurring in foods”; “the available information continues to support the safety of BPA for the currently approved uses in food containers and packaging.” FDA completed a 4-year review of 300+ studies through 2013 and found nothing prompting revision. FDA rescinded authorisation for BPA in baby bottles/sippy cups (2012) and infant formula packaging (2013) — and FDA is explicit that these were abandonment-of-use regulatory clean-ups, not safety actions. Verified.
EFSA (verified): in April 2023 EFSA published a re-evaluation that substantially lowered the tolerable daily intake. The prior 2015 value was a temporary TDI of 4 µg/kg bw/day (itself reduced from an earlier 50 µg/kg bw/day). Verified. The specific new 2023 figure — 0.2 ng/kg bw/day, based on immune-system effects (increased Th17 cells), representing roughly a 20,000-fold reduction, with EFSA concluding that dietary exposure exceeds the new TDI for all age groups — is unverified recall; the EFSA pages I could reach in this session confirmed the direction and the 2015 comparator but not the new number. Treat the specific value as unverified.
The German BfR and the EMA both raised formal scientific objections to EFSA’s derivation. (Verified that joint EFSA/EMA and EFSA/BfR reports exist; the substance of the disagreement is unverified recall.) This is genuinely unusual — a 20,000-fold TDI reduction that a national risk-assessment body publicly disputes.

How to hold this honestly

You have two competent agencies reaching opposite conclusions from the same literature. That is not a situation where anyone should be confident. The reason for the divergence is methodological: EFSA weighted low-dose immunological endpoints from studies that FDA’s and BfR’s evidence-appraisal frameworks score as unreliable. The CLARITY-BPA core study (a large NTP/FDA/academic collaboration specifically designed to resolve the low-dose question) found no consistent low-dose effects in its core guideline-compliant arm, while academic arms reported various effects — and the two camps read that result in opposite directions.

BPS and other substitutes: “BPA-free” cans typically use BPS, BPF, polyester, or acrylic linings. BPS is structurally similar and has similar in-vitro estrogenic activity; EFSA’s own 2020 technical report called for more data on BPS occurrence and migration, and EFSA is only preparing guidance for non-BPA bisphenols with publication expected by 2027 (verified). So “BPA-free” is a marketing claim with an unevaluated substitute behind it — a regrettable-substitution pattern. It should not be treated as a safety upgrade.

Verdict

Genuinely unresolved, low priority for this subject. BPA exposure from canned food is real and measurable; whether it matters at these levels is disputed between regulators. Meanwhile, canned beans and canned tomatoes are two of the highest-value foods in the entire dataset (see the carbohydrate and vegetable sections). Downgrading a meal for containing canned beans would be a serious net-negative error.

Score adjustment: 0. Confidence: low on the underlying science, high that the correct action is to ignore it relative to the nutritional benefit of the foods involved.

5.9 A few others worth a line each

Aflatoxin (peanuts, corn, tree nuts, spices): IARC Group 1, potent hepatocarcinogen, and the one dietary contaminant with unambiguous human cancer causation — but that causation is established in regions with unregulated maize/groundnut storage and high hepatitis B prevalence. In the regulated US supply chain with 20 ppb action levels, this is a non-issue. 0.
Mercury: covered in the proteins/fish section; the only contaminant in this analysis with a genuinely actionable per-species differentiation.
Nitrate/nitrite: covered in the proteins section. The key asymmetry — vegetable nitrate vs cured-meat nitrite — is explained in the vegetables section.
PFAS in food packaging: molded-fibre/compostable bowls were a major PFAS vector; major US chains and manufacturers have largely phased out intentionally-added PFAS, and FDA announced in 2024 that grease-proofing PFAS food-contact uses had been phased out of the US market. Plastic trays are not a PFAS vector. 0.
Titanium dioxide (E171): banned as a food additive in the EU (2022) on genotoxicity-uncertainty grounds, still permitted in the US. An additive, not a whole food — out of scope here but worth flagging if it appears in ingredient lists.

5.10 Summary judgement for the scoring system

Score essentially nothing in this section per-meal. Ranked by how much each item should influence an 800-meal ranking:

Item	Should it affect the meal score?	Why
Acrylamide	No	1.15M-participant meta-analysis: no association at dietary doses
HCAs/PAHs	Barely (−1 for explicitly charred only)	Human link never established; rodent doses 1000× dietary
AGEs	Only via the preparation term	RCT effects real but confounded with “the low-AGE arm ate better food”; do not double-count
Arsenic in rice	No	ILCR ~6×10⁻⁶ vs ~40% baseline lifetime cancer risk; brown rice is the higher-arsenic option
Cadmium	No (−1 for dark-chocolate-heavy meals)	Smoking dominates; food contribution below TDI
Lead in spices	No	Real hazard, but a provenance issue in unregulated bulk supply, not in commercial kitchens
Heavy metals in protein powder	−1 at most	Evidence base is advocacy-grade, benchmarked to Prop 65 not to health thresholds
Microplastics/phthalates from trays	No — it is a constant	Same for all 800 meals; belongs as a standing “decant before microwaving” recommendation
BPA/BPS in cans	No	Regulators disagree; the foods involved (beans, tomatoes) are among the best in the dataset

The one behavioural recommendation that comes out of this entire section: transfer meals out of the plastic tray into ceramic/glass before microwaving. It is the only item where a measured exposure, a plausible mechanism, an unknown-but-nonzero risk, and a zero-cost mitigation all coincide.

The one framing correction: this section is where the largest gap sits between popular concern and evidence. A person optimising meal choice by avoiding acrylamide, arsenic, and BPA while ignoring sodium, energy density, and protein would be optimising the wrong variables by roughly two orders of magnitude.

Machine-readable table — Contaminants and Preparation

MACHINE-READABLE MASTER TABLE

All food classes across all sections, deduplicated. score_adjustment is an integer -10..+10 for a healthy adult at ~[calorie target removed]/day cutting visceral fat and reducing sodium. detection_strings are lowercase substrings, semicolon-separated; match longest-first and use word boundaries. See section F of the framework for the substring-collision failure modes.

food_class	detection_strings	tier	score_adjustment	confidence	key_citation
processed_red_meat	bacon;pancetta;guanciale;lardon;prosciutto;speck;capicola;capocollo;coppa;mortadella;salami;soppressata;genoa salami;hard salami;pepperoni;chorizo;andouille;kielbasa;linguica;longaniza;chourico;nduja;bratwurst;knockwurst;summer sausage;smoked sausage;italian sausage;breakfast sausage;sausage crumbles;bologna;hot dog;frankfurter;wiener;corned beef;pastrami;deli ham;black forest ham;honey ham;smoked ham;serrano ham;ham steak;prosciutto cotto;spam;luncheon meat;cold cuts;salt pork;ham hock;bresaola;jerky;beef stick;bacon bits	harmful	-6	high	Li C et al., Lancet Diabetes Endocrinol 2024;12(9):619-630 (n=1,966,444; processed meat HR 1.15/50 g/d for T2D); Micha R, Circulation 2010 (n=1,218,380; CHD RR 1.42/50 g/d); IARC/Bouvard, Lancet Oncol 2015 (Group 1, +18% CRC per 50 g/d = +0.7 pp lifetime absolute)
uncured_celery_cured_processed_meat	uncured;no nitrate;no nitrates or nitrites added;nitrate-free;nitrite-free;celery powder;celery juice powder;cultured celery;cultured celery powder;naturally cured;uncured bacon;uncured ham;uncured pepperoni	harmful	-6	high	McKeith AG, AMSA “Alternative Curing” fact sheet 6/2014 (celery juice powder ~27,462 ppm nitrate); Nuñez De González MT et al., J Agric Food Chem 2012;60(15):3981-3990 (470 retail products; nitrite comparable to conventional)
processed_poultry	turkey bacon;deli turkey;sliced turkey;smoked turkey breast;turkey ham;turkey salami;turkey pepperoni;chicken sausage;chicken apple sausage;turkey sausage;deli chicken;oven-roasted chicken breast (deli);chicken bologna;turkey franks	harmful	-4	moderate	USDA FoodData Central (turkey salami 1,110 mg Na/100 g; extra-lean deli turkey ham 1,040 mg Na/100 g; fat-free deli chicken breast 1,090 mg Na/100 g); IARC processing definition is species-independent
red_meat_lean_unprocessed	sirloin;top round;eye of round;flank steak;flat iron;filet mignon;beef tenderloin;pork loin;pork chop;pork tenderloin;lean ground beef;90/10 ground beef;93/7 ground beef;ground sirloin;bison;buffalo;venison;elk;goat;lean steak;london broil;tri-tip	contested	0	moderate	Johnston BC et al., Ann Intern Med 2019;171(10):756-764 (weak rec, low-certainty); Zeraatkar D, ibid. 703-710 (55 cohorts, >4M); Hoang T, BMC Cancer 2024 (MR n=374,001, CRC HR 0.72 [0.40-1.28] null); WCRF/AICR 2018 (“probable” for CRC, <500 g/wk)
red_meat_fatty_unprocessed	ribeye;rib eye;short rib;brisket;chuck roast;pot roast;80/20 ground beef;ground chuck;pork belly;pork shoulder;boston butt;lamb shoulder;lamb chop;leg of lamb;ground lamb;veal;oxtail;beef patty;hamburger patty;prime rib;carnitas;pulled pork;stew meat;meatball	contested	-2	moderate	Bergeron N et al., AJCN 2019 APPROACH (RCT crossover n=113; LDL-C and apoB higher with red AND white meat vs nonmeat); Li C et al., Lancet Diabetes Endocrinol 2024 (unprocessed red meat HR 1.10/100 g/d for T2D)
poultry_unbreaded	chicken breast;boneless skinless chicken;chicken thigh;chicken tenderloin;ground chicken;rotisserie chicken;roasted chicken;grilled chicken;shredded chicken;turkey breast;ground turkey;turkey cutlet;duck breast;cornish hen;poussin;chicken drumstick;chicken quarter	beneficial	+3	moderate	Zhong VW et al., JAMA Intern Med 2020 (n=29,682; poultry all-cause mortality HR 0.99 [0.97-1.02]); Li C et al., Lancet Diabetes Endocrinol 2024 (poultry HR 1.08/100 g/d, weakest of the meats)
poultry_breaded_fried	breaded chicken;fried chicken;chicken nugget;chicken tender;popcorn chicken;chicken katsu;karaage;chicken schnitzel;buttermilk fried chicken;crispy chicken;chicken cutlet breaded;nashville hot chicken;chicken patty	harmful	-3	moderate	Zhong VW et al., JAMA Intern Med 2020 (poultry-CVD signal attributed substantially to fried chicken; ARD 1.03% per 2 servings/wk)
eggs_whole	egg;eggs;whole egg;scrambled egg;hard-boiled egg;soft-boiled egg;poached egg;fried egg;omelet;omelette;frittata;quiche;shakshuka;egg patty;deviled egg;egg bite;jammy egg	contested	+2	moderate	Drouin-Chartier JP et al., BMJ 2020;368:m513 (n=215,618 + meta 1,720,108; >=1 egg/d HR 0.93 [0.82-1.05]); Zhong VW et al., JAMA 2019;321(11):1081-1095 (n=29,615; egg assoc. null after adjusting for dietary cholesterol, HR 0.99 [0.93-1.05]); Li MY et al., Nutrients 2020 (17 RCTs; LDL-C +8.14 mg/dL [4.46-11.82])
egg_whites	egg white;egg whites;liquid egg white;egg white omelet;all-white omelet	beneficial	+2	moderate	van Vliet S et al., AJCN 2017 (crossover RCT n=10; whole egg > egg white for postexercise myofibrillar protein synthesis, P=0.04) — whites are lipid-neutral but micronutrient-poor
fish_oily	salmon;atlantic salmon;sockeye;coho;chinook;king salmon;wild salmon;farmed salmon;sardine;sardines;mackerel;saba;spanish mackerel;anchovy;anchovies;herring;kipper;trout;rainbow trout;steelhead;arctic char;sablefish;black cod;bluefish	beneficial	+7	high	Jensen IJ et al., Foods 2020 (farmed vs wild Atlantic salmon: EPA+DHA 1.4 vs 1.2 g/100 g; dioxins+dl-PCB 0.51 vs 1.48 ng TEQ/kg vs EU limit 6.5; Hg 0.018 vs 0.056 mg/kg); Mohan D et al., JAMA Intern Med 2021 (n=191,558; benefit concentrated in secondary prevention); Abdelhamid AS, Cochrane 2020 (86 RCTs, n=162,796; CV mortality RR 0.92 [0.86-0.99])
fish_white	cod;atlantic cod;pacific cod;haddock;pollock;alaska pollock;tilapia;halibut;flounder;sole;dover sole;barramundi;branzino;snapper;red snapper;grouper;mahi;mahi-mahi;swai;basa;catfish;hake;monkfish;rockfish;sea bass;whiting;turbot;plaice;fish fillet;white fish	beneficial	+5	moderate	FDA Mercury Levels in Commercial Fish 1990-2012 (cod 0.111 ppm, tilapia 0.013, pollock 0.031, halibut 0.241); Zhong VW et al., JAMA Intern Med 2020 (fish CVD HR 1.00 [0.98-1.02], mortality 0.99 [0.97-1.01])
shellfish	shrimp;prawn;scallop;bay scallop;sea scallop;mussel;clam;littleneck;oyster;crab;lump crab;crab meat;lobster;langoustine;calamari;squid;octopus;crawfish;crayfish;cockle	beneficial	+4	moderate	FDA Mercury Levels 1990-2012 (shrimp 0.009 ppm, scallop 0.003 ppm — lowest measured); docked from +5 for STPP-brine sodium in commercial supply
tuna_light_canned	chunk light tuna;light tuna;skipjack;canned tuna;tuna in water;tuna pouch;tuna salad	beneficial	+4	moderate	FDA Mercury Levels 1990-2012 (canned light/skipjack 0.126 ppm mean; ~444 g/wk to reach EPA RfD for an 80 kg adult)
tuna_albacore_or_steak	albacore;solid white tuna;white tuna;tuna steak;ahi;yellowfin;bigeye;poke;seared tuna;bonito	beneficial	+2	moderate	FDA Mercury Levels 1990-2012 (canned albacore 0.350 ppm, yellowfin 0.354, bigeye 0.689; albacore ~160 g/wk = ~1.9 servings reaches the 80 kg EPA RfD of 56 ug/wk)
fish_smoked_cured	smoked salmon;lox;gravlax;nova;smoked trout;smoked mackerel;kippers;salt cod;bacalao;cured fish;anchovy paste	contested	+1	moderate	Omega-3 retained but cured/high-sodium; USDA (canned white tuna in oil 3 oz = 337 mg Na) as sodium reference; no IARC Group 1 evaluation exists for processed fish
imitation_seafood	imitation crab;krab;surimi;seafood sticks;crab stick;kanikama	harmful	-3	moderate	Starch-bound, high-sodium, low-protein processed product; no primary nutrition citation retrieved this session — scored on composition, not epidemiology
soy_whole_foods	tofu;firm tofu;extra firm tofu;silken tofu;tofu puffs;tempeh;edamame;soybean;soy beans;natto;bean curd;yuba;tofu skin;soy curls;miso	beneficial	+5	high	Reed KE et al., Reprod Toxicol 2021;100:60-67 (41 clinical studies; no effect on TT/FT/E2/E1/SHBG in men); Otun J et al., Sci Rep 2019 (18 RCTs; TSH +0.248 mIU/L, fT3/fT4 unchanged); Blanco Mejia S et al., J Nutr 2019 (46 controlled trials; LDL-C -4.76 mg/dL); Shu XO et al., JAMA 2009 (n=5,042 survivors; recurrence HR 0.68 [0.54-0.87])
soy_protein_isolate_tvp	soy protein;soy protein isolate;soy protein concentrate;textured vegetable protein;textured soy protein;tvp;isolated soy protein	beneficial	+3	moderate	Blanco Mejia S et al., J Nutr 2019 (LDL-C -4.76 mg/dL); Davis BE et al., J Diet Suppl 2026 (12 RCTs, n=261; whey vs soy no LBM difference)
pea_protein	pea protein;pea protein isolate;yellow pea protein;plant protein blend	beneficial	+3	moderate	Babault N et al., JISSN 2015;12(1):3 (RCT n=161, 12 wk, pea vs whey vs placebo; between-group P=0.09 on primary outcome, positive only in post-hoc weakest-participant subgroup; Roquette-funded); Santini MH et al., JISSN 2025 (RCT n=44; plant vs animal protein blends produced similar ~2.4-2.5 kg lean mass gains); Korzepa et al., Eur J Nutr 2025 (lower leucine iAUC vs whey, clinical relevance undetermined)
seitan_wheat_gluten	seitan;vital wheat gluten;wheat gluten;wheat protein;mock duck;wheat meat	neutral	+1	moderate	Lebwohl B et al., BMJ 2017 (n=110,017, 26 y, 2,273,931 person-years; highest vs lowest gluten CHD HR 0.95 [0.88-1.02]); lysine-limiting and usually braised in soy sauce (sodium vector)
plant_meat_analog	beyond meat;impossible;plant-based burger;plant-based sausage;meatless crumbles;veggie burger;vegan chicken;vegan sausage;plant-based ground;meatless meatball	contested	0	low	No verified primary citation retrieved this session; scored on composition (typically 350-450 mg sodium/serving, refined/coconut oils, ultra-processed) — better than processed meat, worse than whole soy
yogurt_kefir	yogurt;yoghurt;greek yogurt;plain yogurt;skyr;kefir;labneh;icelandic yogurt;tzatziki;raita;dahi;curd (indian)	beneficial	+6	moderate	Companys J et al., Adv Nutr 2020 (20 cohorts + 52 RCTs; yogurt-T2D RR 0.73 [0.70-0.76]); Zhang K et al., Crit Rev Food Sci Nutr 2019 (10 cohorts, n=385,122; yogurt-CVD OR 0.78 [0.67-0.89]); Dehghan M et al., Lancet 2018 PURE (n=136,384; yogurt HR 0.86 [0.75-0.99]); USDA (plain low-fat yogurt 6 oz = 119 mg Na)
cottage_cheese_ricotta	cottage cheese;ricotta;quark;farmer cheese;pot cheese;paneer	beneficial	+4	moderate	USDA SR Legacy (cottage cheese lowfat 2%, 4 oz = 348 mg Na; 1% = 459 mg); high casein protein-per-calorie for energy-restricted lean-mass preservation
cheese_aged_salty	cheese;cheddar;mozzarella;parmesan;parmigiano;pecorino;romano;asiago;feta;goat cheese;chevre;blue cheese;gorgonzola;halloumi;provolone;swiss cheese;gruyere;manchego;monterey jack;pepper jack;brie;burrata;fresh mozzarella;cotija;queso fresco;gouda;havarti;fontina	contested	0	moderate	Dehghan M et al., Lancet 2018 PURE (cheese HR 0.88 [0.76-1.02], null); Rooney et al., Atherosclerosis 2025 (RCT n=197; cheese lowered LDL vs deconstructed cheese) vs O’Connor et al., Food Funct 2024 (RCT n=162; no difference) — matrix hypothesis unsettled; USDA (parmesan 1,804 mg Na/cup; feta 1,708 mg Na/cup)
cheese_processed	american cheese;cheese sauce;nacho cheese;queso;cheese spread;cheese product;cheese food;velveeta;cheese whiz;cheese powder	harmful	-3	moderate	USDA SR Legacy (pasteurized process American cheese spread, 1 cup diced = 2,275 mg Na)
milk	milk;whole milk;2% milk;1% milk;skim milk;nonfat milk;buttermilk;evaporated milk;milk powder (dairy)	neutral	+2	moderate	Dehghan M et al., Lancet 2018 PURE (milk >1 serving/d HR 0.90 [0.82-0.99]); USDA (fluid milk ~120 mg Na/cup)
whey_casein_protein	whey protein;whey isolate;whey concentrate;casein;micellar casein;milk protein isolate;protein powder;protein shake;protein blend	beneficial	+4	moderate	Morton RW et al., Br J Sports Med 2018 (49 RCTs, n=1,863; FFM +0.30 kg [0.09-0.52], 1RM +2.49 kg [0.64-4.33], plateau ~1.6 g/kg/d); Khalafi et al., Healthcare 2025 (25 studies, n=1,454)
butter_cream_dairy_fat	butter;heavy cream;heavy whipping cream;half and half;ghee;clarified butter;creme fraiche;sour cream;mascarpone	contested	-2	low	Dehghan M et al., Lancet 2018 PURE (butter HR 1.09 [0.90-1.33], null); Imamura F et al., PLoS Med 2018 (dairy-fat biomarker 15:0/17:0 sum HR 0.71 [0.63-0.79] for T2D) vs Steffen et al., Front Nutr 2026 (two-sample MR: no causal effect of C15:0) — biomarker not established as causal; scored down on energy density for this subject’s deficit
legumes_pulses	black beans;black bean;pinto beans;kidney beans;red beans;cannellini;white beans;navy beans;great northern;butter beans;lima beans;fava beans;broad beans;chickpea;chickpeas;garbanzo;lentil;lentils;red lentil;green lentil;du puy;beluga lentil;split pea;split peas;black eyed pea;black-eyed peas;adzuki;mung bean;dal;daal;dahl;chana	beneficial	8	high	Back S et al., J Am Heart Assoc 2026;15(10):e046659 (MA-RCT, 38 trials, n=2095, LDL -0.14 mmol/L)
edamame_soy	edamame;soybeans;soy beans;green soybeans	beneficial	7	moderate	Viguiliouk E et al., Adv Nutr 2019;10(Suppl_4):S308-S319
barley	barley;pearl barley;pearled barley;hulled barley;barley groats	beneficial	6	high	Ho HV et al., Eur J Clin Nutr 2016;70(11):1239-1245 (MA-RCT, 14 trials, LDL -0.25 mmol/L)
oats_intact	oats;rolled oats;steel cut oats;steel-cut oats;oatmeal;oat groats;porridge	beneficial	6	high	Whitehead A et al., Am J Clin Nutr 2014;100(6):1413-21 (MA-RCT, 28 trials, LDL -0.25 mmol/L)
bulgur	bulgur;bulghur;cracked wheat	beneficial	5	moderate	USDA SR Legacy 170287 (4.5 g fibre/100 g cooked) + Reynolds A et al., Lancet 2019;393:434-445
quinoa	quinoa;red quinoa;tricolor quinoa	beneficial	5	moderate	USDA SR Legacy 168917 (lysine 54 mg/g protein, Mg 64 mg/100 g)
legume_pasta	chickpea pasta;lentil pasta;red lentil pasta;edamame pasta;banza;bean pasta	beneficial	5	moderate	Kanata MC et al., Food Funct 2025;16(11):4548-4561 (XO RCT n=15)
farro_spelt	farro;spelt;emmer;einkorn;kamut;khorasan;wheat berries;wheat berry	beneficial	4	low	USDA SR Legacy 169746 (5.5 g protein, 3.9 g fibre/100 g cooked); no RCT evidence specific to farro
freekeh	freekeh;frikeh;farik	beneficial	4	low	compositional only; no human trials located
buckwheat	buckwheat;kasha;buckwheat groats	beneficial	4	low	USDA SR Legacy 170686 (lysine 51 mg/g protein, Mg 51 mg/100 g)
wild_rice	wild rice;wild rice blend	beneficial	4	low	USDA SR Legacy 168897 (3.99 g protein, 1.8 g fibre/100 g); NOT Oryza, no rice arsenic burden
hummus	hummus;houmous;hommus	beneficial	4	moderate	USDA SR Legacy 174289 (426 mg sodium/100 g offsets pulse benefit for a sodium-reducing subject)
brown_rice	brown rice;brown basmati;brown jasmine;germinated brown rice	beneficial	3	moderate	Sun Q et al., Arch Intern Med 2010;170(11):961-9 (RR 0.89) vs FDA 2016 RA (160.5 ppb iAs vs 103.3 white)
sweet_potato	sweet potato;sweetpotato;sweet potatoes;kumara;garnet yam	beneficial	3	moderate	USDA SR Legacy 168483 (961 ug RAE vit A, 475 mg K, 3.3 g fibre/100 g)
whole_grain_bread_pasta	whole wheat bread;whole grain bread;whole wheat pasta;whole wheat tortilla;whole wheat pita;whole grain wheat flour;whole wheat flour	beneficial	2	moderate	Zafar TA et al., Int J Food Sci 2020;2020:8834960 (whole wheat bread glycemic response similar to white)
potato_plain	potato;potatoes;russet;yukon gold;red potato;baby potato;fingerling;new potatoes;mashed potato;roasted potato;boiled potato;potato wedges	beneficial	2	moderate	Muraki I et al., Diabetes Care 2016;39(3):376-384 (baked/boiled/mashed HR 1.04 per 3 svg/wk) + Holt SHA et al., Eur J Clin Nutr 1995;49:675-690 (satiety index 323)
corn_masa	corn tortilla;masa;masa harina;nixtamal;hominy;arepa	beneficial	2	moderate	USDA SR Legacy 173241 (5.2 g fibre, 11 mg sodium/100 g)
sourdough_whole	sourdough;whole grain sourdough;levain	beneficial	2	low	Ozer YE et al., Wien Klin Wochenschr 2023;135(13-14):349-357 (XO, 45.5% less insulin, 9.6% lower 1h glucose)
sweet_corn	sweet corn;corn kernels;corn on the cob;elote	beneficial	2	moderate	USDA SR Legacy 169999 (218 mg K, 2.4 g fibre/100 g)
millet	millet;pearl millet;proso millet	beneficial	2	low	USDA SR Legacy 168871 (only 1.3 g fibre/100 g cooked; lysine 19 mg/g protein)
semolina_pasta	pasta;spaghetti;penne;rigatoni;fusilli;linguine;fettuccine;macaroni;farfalle;rotini;bucatini;cavatappi;lasagna;orzo;couscous	neutral	1	moderate	Chiavaroli L et al., BMJ Open 2018;8(3):e019438 (MA-RCT, 32 comparisons, weight -0.63 kg); Atkinson FS et al., Am J Clin Nutr 2021;114:1625-1632 (pasta consistently low-GI)
dried_fruit_unsweetened	raisins;golden raisins;dates;medjool;dried apricots;dried figs;prunes;currants;dried cherries	neutral	1	moderate	Muraki I et al., BMJ 2013;347:f5001 (grapes and raisins HR 0.88 per 3 svg/wk)
white_rice	white rice;jasmine rice;basmati;basmati rice;parboiled rice;converted rice;long grain rice;long-grain rice;sushi rice;short grain rice;arborio;calrose;sticky rice;glutinous rice;coconut rice;rice pilaf	contested	-1	moderate	Hu EA et al., BMJ 2012;344:e1454 (Western RR 1.12, 0.94-1.33, ns; Asian RR 1.55); PURE China HR 1.04 ns
flour_tortilla	flour tortilla;wheat tortilla;wrap;burrito tortilla	contested	-1	low	USDA SR Legacy 167535/175037 (refined flour + added fat + sodium)
refined_grains	enriched flour;enriched wheat flour;all-purpose flour;white flour;bleached flour;unbleached flour;white bread;white bun;brioche;panko;breadcrumbs;croutons;white rice flour;rice flour;polenta;grits;instant mashed potato	harmful	-2	moderate	Mozaffarian D et al., NEJM 2011;364(25):2392-2404 (refined grains +0.39 lb/4 y per daily serving)
rice_cakes_puffed	rice cake;rice cakes;puffed rice;rice crisps	harmful	-2	moderate	Rondanelli M et al., J Med Food 2023;26(6):422-427 (GI 83.3 classic, 102.2 brown)
sweetened_dried_fruit	sweetened dried cranberries;craisins;candied;glace;sweetened dried mango;fruit-juice infused	harmful	-2	moderate	added sugar counts against 36 g/day target; USDA labelling rule
fruit_juice	orange juice;apple juice;fruit juice;100% juice;juice concentrate;apple juice concentrate;fruit juice concentrate	harmful	-3	moderate	Muraki I et al., BMJ 2013;347:f5001 (juice HR 1.08 per 3 svg/wk vs whole fruit 0.98)
added_sugar_sauces	honey;maple syrup;agave;brown rice syrup;cane sugar;brown sugar;molasses;corn syrup;high fructose corn syrup;date syrup;hoisin;teriyaki;bbq sauce;barbecue sauce;sweet chili;sweet and sour;honey mustard;glaze;orange sauce	harmful	-4	moderate	8-18 g added sugar per serving = 22-50% of the 36 g/day target; also 400-900 mg sodium
french_fries	french fries;fries;steak fries;curly fries;tater tots;hash browns;home fries;potato skins	harmful	-5	moderate	Mozaffarian D et al., NEJM 2011 (+3.35 lb/4 y); Muraki I et al., Diabetes Care 2016 (HR 1.19 per 3 svg/wk)
potato_chips	potato chips;crisps;kettle chips;tortilla chips;corn chips	harmful	-5	moderate	Mozaffarian D et al., NEJM 2011;364(25):2392-2404 (+1.69 lb/4 y per daily serving)
extra_virgin_olive_oil	extra virgin olive oil; extra-virgin olive oil; evoo; cold pressed olive oil	beneficial	+6	moderate	Guasch-Ferré, JACC 2022;79:101-112 (n=92,383, HR 0.81 total mortality); Estruch, NEJM 2018;378:e34 (PREDIMED, republished)
olive_oil_refined	olive oil; pure olive oil; light olive oil; classic olive oil	beneficial	+4	moderate	Guasch-Ferré, JACC 2022;79:101-112
olive_pomace_oil	olive pomace oil; pomace oil; olive-pomace oil	neutral	+2	moderate	MOH speciation study: pomace 33-205 mg/kg MOSH, 2-55 mg/kg MOAH, but 3+-ring genotoxic species ABSENT in all samples; downgrade vs EVOO is polyphenol loss, not contaminants
whole_nuts	almond; walnut; pecan; pistachio; cashew; hazelnut; macadamia; peanut; pine nut; mixed nuts; slivered almonds; toasted almonds	beneficial	+6	high	Novotny, Am J Clin Nutr 2012 (almond ME 32% below Atwater); Bao, NEJM 2013;369:2001
nut_butter	peanut butter; almond butter; cashew butter; nut butter; almond flour; almond meal; peanut sauce	beneficial	+3	high	Gebauer/Novotny USDA: almond butter ME 6.53 kcal/g = Atwater prediction (P=0.08)
candied_or_salted_nuts	honey roasted; candied pecan; candied walnut; salted peanuts; praline	neutral	+1	moderate	added sugar + sodium offset
seeds	chia; flaxseed; flax seed; ground flax; pumpkin seed; pepita; sunflower seed; hemp heart; hemp seed; sesame seed; poppy seed	beneficial	+4	moderate	flax RCT BP meta-analyses (effect size unverified); high ALA/fibre/mineral density
tahini_seed_butter	tahini; sesame paste; sunflower butter; sunbutter	beneficial	+3	moderate	as seeds, but full calorie absorption (no matrix effect)
avocado	avocado; guacamole; avocado slices; hass avocado	beneficial	+4	high	Lichtenstein, J Am Heart Assoc 2022;11:e025657 (n=1,008, VAT null P=0.405, LDL -2.47 mg/dL)
avocado_oil	avocado oil; high oleic avocado oil	beneficial	+3	low	high MUFA, high smoke point; little outcome data; frequent adulteration reports
coconut_oil	coconut oil; refined coconut oil; virgin coconut oil; copha	harmful	-4	high	Neelakantan, Circulation 2020;141:803-814 (16 RCTs, LDL +10.47 mg/dL, 95% CI 3.01-17.94)
coconut_milk_cream	coconut milk; coconut cream; creamed coconut; coconut water is NOT this	harmful	-2	moderate	as coconut oil, diluted; fewer g SFA/serving
butter	butter; unsalted butter; salted butter; cultured butter; browned butter; beurre	contested	-2	moderate	Hooper, Cochrane 2020 CD011737.pub3 (RR 0.83, 12 trials, n=53,758, CV events; no mortality effect)
ghee	ghee; clarified butter	contested	-2	moderate	as butter, marginally more concentrated SFA
animal_fat	lard; tallow; beef tallow; duck fat; bacon fat; schmaltz; suet	contested	-2	low	as butter; SFA-raising, weak event evidence
seed_oils_neutral	canola oil; rapeseed oil; sunflower oil; safflower oil; soybean oil; grapeseed oil; corn oil; rice bran oil; vegetable oil; cottonseed oil	contested	0	moderate	Johnson & Fritsche, J Acad Nutr Diet 2012 (15 RCTs, no inflammatory marker effect); Marklund, Circulation 2019;139:2422 (30 cohorts, ~69,000, LA inversely associated with CVD)
high_oleic_oils	high oleic sunflower; high-oleic safflower; high oleic canola	neutral	+1	moderate	as above, better oxidative stability
sesame_oil	sesame oil; toasted sesame oil; sesame seed oil	beneficial	+1	low	lignan content; used in flavour quantities
peanut_oil	peanut oil; groundnut oil; arachis oil	neutral	0	moderate	high MUFA, neutral profile
palm_oil_refined	palm oil; palm olein; palm kernel oil; refined palm	harmful	-3	moderate	~50% palmitic acid; highest 3-MCPD/GE among refined oils (EU limits 2.5 mg/kg 3-MCPD, 1.0 mg/kg GE)
margarine_modern	margarine; plant butter; vegetable oil spread; buttery spread	neutral	0	moderate	post-PHO-ban trans-free; sterol-fortified variants lower LDL ~8-10% at 2 g/d
trans_fat	partially hydrogenated; hydrogenated oil; shortening; vanaspati	harmful	-8	high	the one fat with unambiguous large harm evidence; should be near-absent in current products
mct_oil	mct oil; medium chain triglyceride; c8 oil; caprylic acid oil	neutral	0	moderate	distinct from coconut oil (C8/C10 vs C12 lauric); no visceral-fat benefit
cruciferous_vegetables	broccoli; broccolini; broccoli rabe; cauliflower; brussels sprout; cabbage; napa cabbage; red cabbage; savoy; bok choy; kale; collard; mustard green; arugula; rocket; watercress; kohlrabi; turnip; daikon; radish; horseradish; wasabi	beneficial	+6	moderate	Pollock, JRSM Cardiovasc Dis 2016;5:2048004016661435 (8 cohorts, RR 0.842, 0.753-0.941); Qidong broccoli-sprout trials are biomarker-only
leafy_greens_dark	spinach; baby spinach; kale; swiss chard; chard; collard greens; mixed greens; spring mix; mesclun; romaine; escarole; mustard greens; beet greens; watercress	beneficial	+7	moderate	nitrate/NO pathway; nutrient density per kcal; Pollock 2016 as above
lettuce_light	iceberg; butter lettuce; bibb; green leaf lettuce; shredded lettuce	beneficial	+2	moderate	volume/satiety only; low nutrient density
beets	beet; beets; roasted beet; golden beet; beetroot	beneficial	+4	moderate	Eroglu, Nutrients 2026;18:2513 (23 RCTs, n=401 — largely NULL for sprint/repeated-sprint performance, low certainty)
alliums_fresh	garlic; roasted garlic; minced garlic; onion; red onion; yellow onion; shallot; leek; scallion; green onion; chive; ramp	beneficial	+2	moderate	Ma, Asian Biomed 2025;19:131-140 (12 reports, 405 hypertensive pts, DBP -4.26 mmHg) — but SUPPLEMENT doses in HYPERTENSIVE patients
alliums_dried	garlic powder; onion powder; granulated garlic; dehydrated onion	beneficial	+1	low	flavour-per-sodium value; allicin chemistry largely absent
tomatoes	tomato; cherry tomato; grape tomato; roma; diced tomatoes; crushed tomatoes; tomato paste; tomato puree; sun-dried tomato; passata	beneficial	+4	moderate	lycopene bioavailability higher when cooked; FDA rejected the lycopene-prostate health claim
mushrooms	mushroom; cremini; crimini; button mushroom; portobello; shiitake; oyster mushroom; maitake; enoki; porcini; king trumpet	beneficial	+5	moderate	umami/sodium-substitution literature (mushroom blend); ergothioneine hypothesis unproven; agaritine not a real concern
peppers_capsicum	bell pepper; red pepper; poblano; jalapeno; serrano; chili pepper; chipotle; habanero; padron; shishito; banana pepper	beneficial	+4	moderate	high vitamin C; capsaicin thermogenesis real but trivial (tens of kcal/day)
note_nightshade_no_penalty	(do not encode - no evidence)	neutral	0	high	Childers hypothesis rests on an uncontrolled self-selected questionnaire; no RCT support
root_and_squash_veg	carrot; butternut squash; acorn squash; kabocha; parsnip; rutabaga; celery root; celeriac	beneficial	+4	moderate	carotenoids + fibre; CARET/ATBC: isolated beta-carotene supplements HARMED smokers
other_vegetables	zucchini; summer squash; asparagus; green bean; haricot vert; snap pea; snow pea; pea; artichoke; eggplant; aubergine; cucumber; celery; fennel; okra; hearts of palm; jicama	beneficial	+3	moderate	energy density and fibre; standard cohort caveats
berries	blueberry; strawberry; raspberry; blackberry; cranberry; acai	beneficial	+5	moderate	Cassidy/Rimm NHS II anthocyanin cohort work (effect estimate unverified)
whole_fruit	apple; pear; orange; peach; mango; pineapple; banana; grape; melon; cherry; plum; kiwi; pomegranate	beneficial	+4	moderate	whole fruit inversely associated with T2D while juice is positively associated - matrix effect
fermented_live_possible	kimchi; raw sauerkraut; unpasteurized sauerkraut; kefir; live cultures	contested	0	moderate	Wastyk, Cell 2021;184:4137 (n=18/arm, 17 wk; PRIMARY OUTCOME NULL; fermented arm raised diversity, lowered inflammatory markers) - offset by 400-700 mg sodium/100 g
fermented_cooked	tempeh; natto; cooked kimchi; sauerkraut (cooked); sourdough	beneficial	+1	moderate	cultures dead; fermentation benefits (phytate reduction, peptides, K2) survive
miso	miso; miso paste; white miso; red miso; shiro miso	neutral	-1	moderate	~600-900 mg sodium/tbsp dominates any fermentation benefit for a sodium-reducing subject
soy_sauce_high_sodium	soy sauce; shoyu; tamari; fish sauce; oyster sauce; hoisin; teriyaki	harmful	-2	high	~900-1,000 mg sodium/tbsp; ~2 tbsp = 80% of the AHA 1,500 mg ideal daily intake
vinegar	vinegar; rice vinegar; balsamic; apple cider vinegar; red wine vinegar; sherry vinegar	beneficial	+2	moderate	sodium-free acidity for flavour; modest RCT evidence for blunted postprandial glycaemia
kombucha	kombucha	neutral	0	low	usually pasteurized or low-viability at retail; often sweetened
turmeric	turmeric; ground turmeric; curcumin; haldi	neutral	0	high	Nelson, J Med Chem 2017;60:1620-1637: curcumin is a PAINS/IMPS compound; “>120 clinical trials”; “No double-blinded, placebo controlled clinical trial of curcumin has been successful.” Meal dose is 2-3 orders of magnitude below trial dose
ginger	ginger; fresh ginger; ground ginger; galangal	beneficial	+1	moderate	best-evidenced culinary spice, but specifically for nausea at ~1 g/day
cinnamon	cinnamon; cassia; ceylon cinnamon	neutral	0	moderate	glycaemic effects small, mixed, mostly in T2D; cassia coumarin TDI 0.1 mg/kg bw/day only relevant at supplement doses
culinary_herbs	rosemary; thyme; oregano; basil; parsley; cilantro; dill; sage; mint; tarragon; marjoram; bay leaf; herbes de provence; italian seasoning	beneficial	+1	moderate	scored for SODIUM SUBSTITUTION, not pharmacology; rosemary extract does retard lipid oxidation in cooked meat
culinary_spices	cumin; coriander; paprika; smoked paprika; chili powder; cayenne; black pepper; peppercorn; cardamom; clove; nutmeg; allspice; fenugreek; sumac; za’atar; garam masala; curry powder; harissa	beneficial	+1	moderate	same rationale: flavour intensity without sodium
seaweed_nori	nori; wakame; seaweed; dulse	beneficial	+2	low	iodine; nori modest
seaweed_kelp	kelp; kombu; kelp powder; bladderwrack	harmful	-1	low	iodine can far exceed the 1,100 ug/day UL; chronic excess causes thyroid dysfunction
olives	olive; kalamata; castelvetrano; green olives; black olives; olive tapenade	beneficial	+1	moderate	good fat + polyphenols, offset by cure sodium (~250-450 mg per 10)
pickles_brined	pickle; dill pickle; cornichon; pickled onion; giardiniera; capers; pepperoncini	harmful	-1	moderate	sodium load with little nutritional return; capers especially concentrated
prep_deep_fried	deep fried; deep-fried; battered; tempura; breaded and fried; crispy fried; french fries; fried chicken	harmful	-5	moderate	penalty is for energy density, added fat, and oil oxidation products — NOT acrylamide
prep_charred_blackened	charred; blackened; burnt ends; heavily grilled; caramelized crust on meat	harmful	-1	moderate	NCI: population studies have not established HCA/PAH-cancer link; rodent doses 1000x dietary
prep_grilled_roasted	grilled; roasted; broiled; seared; pan-seared; baked	neutral	0	high	Filippini, Front Nutr 2022;9:875607 (16 studies, n=1,151,189, no acrylamide-cancer association)
prep_steamed_braised_poached	steamed; braised; poached; simmered; slow-cooked; stewed	beneficial	+1	moderate	Sohouli, Adv Nutr 2021;12:766-776 (13 RCTs, low-AGE diet: HOMA-IR -1.20, LDL -6.26 mg/dL) — but effect likely confounded with overall diet quality
note_acrylamide_no_penalty	potato chips; crisps; hash brown; roasted potato; toast; crispbread	neutral	0	high	Pelucchi, Int J Cancer 2015;136:2912 (continuous estimates 0.95-1.03, all non-significant)
note_rice_arsenic_no_penalty	brown rice; rice cake; brown rice syrup; rice cereal	neutral	0	high	Food Chem Toxicol 2020;141:111420 (brown 189 vs white 132 ug/kg total As); IJERPH 2022;19:16460 (ILCR 6.0e-6)
dark_chocolate_cacao_heavy	dark chocolate; cacao nibs; cocoa powder (as primary component)	neutral	-1	moderate	Consumer Reports 2022 cadmium/lead testing vs Prop 65 MADL (conservative benchmark, not a health threshold)
protein_powder_unspecified	protein powder; protein blend; whey isolate powder; plant protein blend	neutral	-1	low	Clean Label Project 2018/2024 — advocacy-grade, not peer-reviewed; prefer NSF Certified for Sport
note_canned_bpa_no_penalty	canned; can of; tinned; canned tomatoes; canned beans; canned tuna	neutral	0	low	FDA: “BPA is safe at current levels”; EFSA 2023 lowered TDI from the 2015 temporary 4 ug/kg bw/day; BfR and EMA objected. Do NOT penalise — the foods involved are high-value
note_plastic_tray_is_a_constant	(applies to all meals — do not encode per-meal)	neutral	0	moderate	Hussain, Environ Sci Technol 2023;57:9782 (microwave = highest release; modelled EDI 20.3-22.1 ng/kg/day); Zota, EHP 2016;124:1521 (n=8,877, DEHP +23.8%, DiNP +39.0% in high fast-food consumers)

Applying This Table to 800 Meals — Integration Notes

1. Rows that are NOT foods

Five rows are prefixed note_. They carry score_adjustment = 0 and exist to record an explicit decision not to penalise something that a naive system would penalise. Do not delete them, and do not let a later contributor “fix” them by adding a penalty:

Row	Decision recorded
`note_acrylamide_no_penalty`	1,151,189-participant meta-analysis found no acrylamide-cancer association. Penalise fries for energy density and fat, never for acrylamide.
`note_rice_arsenic_no_penalty`	Incremental lifetime cancer risk ~6×10⁻⁶ against a ~40% baseline. Also brown rice has more arsenic than white, so an arsenic penalty would invert the fibre ranking.
`note_plastic_tray_is_a_constant`	Every meal ships in a plastic tray, so it cancels across all 800 and adds only noise. Handle as a standing recommendation, not a score.
`note_nightshade_no_penalty`	The nightshade-inflammation claim traces to an uncontrolled self-selected questionnaire. No RCT support.
`note_canned_bpa_no_penalty`	FDA and EFSA actively disagree; meanwhile canned beans and tomatoes are among the highest-value foods in the dataset. Penalising them would be a large net error.

2. Known overlapping rows — do not sum them

Several concepts are legitimately represented at more than one granularity. Match the most specific row and stop; do not add both.

extra_virgin_olive_oil → olive_oil_refined → olive_pomace_oil. All three contain the substring olive oil. Match longest-first.
legumes_pulses / edamame_soy / soy_whole_foods / hummus / legume_pasta — a chickpea can match several. Take the most specific.
prep_deep_fried vs french_fries / potato_chips / poultry_breaded_fried — the specific food rows already price in the frying. Applying prep_deep_fried on top double-counts. Use prep_deep_fried only when a deep-fried item has no dedicated row.
processed_red_meat vs uncured_celery_cured_processed_meat — the same score, deliberately. Matching either is sufficient; matching both must not double the penalty.
whole_nuts vs nut_butter vs tahini_seed_butter — mutually exclusive by preparation, but almond substring-matches almond butter. Match the butter forms first.

3. The three things that should outweigh this entire table

For this subject, ranked by expected effect on the adult’s stated goals:

Energy density and total calories. The adult is in a deficit at [calorie target removed]. A 900 kcal tray and a 550 kcal tray with identical ingredient scores are not equivalent meals.
Sodium. The adult is actively reducing it. Sodium is embedded in cured meat, cheese, soy sauce, miso, olives, pickles, canned beans, bread, and nearly every sauce. Count it once, as milligrams, at the meal level — do not try to reconstruct it from ingredient scores, which will systematically undercount.
Protein grams and quality. The adult is preserving lean mass while cutting. A meal below ~30 g of protein is a problem regardless of how well its vegetables score.

A defensible weighting is roughly: 40% macro/energy fit, 25% sodium, 20% ingredient-quality score (this table), 15% fibre and micronutrient density. This table should not be the whole model.

4. Where the scores in this table are least trustworthy

Stated plainly so downstream users can discount appropriately:

Everything scored confidence: low — avocado oil, ergothioneine-driven mushroom scoring, kombucha, kelp, protein powder contaminants, animal fats.
All between-vegetable differentials. The evidence that vegetables beat no vegetables is far stronger than the evidence that kale beats zucchini. If you compressed every vegetable to a single +4, you would lose very little real information.
Anything whose score rests on a single cohort’s hazard ratio in the 0.8–1.25 band. That is the noise floor of observational nutrition (framework §C).
Quantity blindness. The table scores presence, not dose. It cannot tell a drizzle of olive oil from a quarter-cup, and it will therefore systematically misprice calorie-dense ingredients. Use ingredient-list position as a partial correction.

5. Unverified items carried in this document

Each section contains its own explicit ledger of claims that could not be verified in this session. They are marked inline with “unverified recall” or “unverified.” The largest concentrations are: EFSA’s acrylamide BMDL10 value, EFSA’s 2023 BPA TDI figure, per-variety potato and rice glycaemic index numbers, several composition-table sodium and iodine values, and a handful of sample sizes for well-established papers whose existence and direction are not in doubt. No citation, sample size, effect estimate, or DOI in this document was invented. Where a number could not be confirmed, the number is flagged separately from the qualitative claim so the two can be discounted independently.