Frontier evidence 2024 to 2026

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.

The 2024-2026 frontier: processed food, additives and diet quality

What changed in the last two years that most summaries have not caught up with.

Scope. Written 2026-09-06 as a curiosity-driven companion to /research/health/rubric/. The rubric is the baseline; this document covers only what is new or newly contested relative to it. Section 2 states each conflict explicitly as “rubric says X, new evidence says Y”.

Verification marks. [V] = citation fetched and read during the 2026-09-06 research session (journal page, Europe PMC record, PMC full text, ClinicalTrials.gov API v2, Crossref, or a primary agency document). [R] = reported second hand or recalled and not independently fetched. Treat [R] as a lead, not proof.

Evidence tiers used throughout.

Tier	Meaning
A	RCT with a clinical or hard intermediate outcome, adequately powered
B	RCT with a surrogate or mechanistic outcome, or a small/underpowered RCT
C	Prospective cohort, or biomarker-based human observational study
D	Cross-sectional human, post hoc, modelled, or ecological
E	Animal, organoid, or cell only
F	Detection-only: a substance was measured in human tissue, no outcome attached
G	Regulatory or policy action; not itself evidence

---

1. Topic by topic

1.1 The Lancet 2025 UPF Series and its critics

What changed. The field acquired an agreed-upon canonical statement of the pro-UPF-policy case, and, three months later, an agreed-upon list of what its authors will and will not concede. Before November 2025 the debate was diffuse. It is now pinned to specific text.

The Series. Three peer-reviewed Reviews, The Lancet 406(10520), published 18 Nov 2025 (issue dated 6 Dec 2025) [V].

Paper	Citation	Core claim
1	Monteiro CA, Louzada ML, Steele-Martinez E, Cannon G, … Srour B, Swinburn B, Touvier M. “Ultra-processed foods and human health: the main thesis and the evidence.” pp 2667-2684. 10.1016/S0140-6736(25)01565-X, PMID 41270766	Three hypotheses: UPF displaces whole-food diets; degrades diet quality; raises chronic disease risk. Claims “more than 100 prospective studies”
2	Scrinis G, Popkin BM, Corvalan C, Duran AC, Nestle M, Lawrence M, Baker P, Monteiro CA, Millett C, Moubarac JC, Jaime P, Khandpur N. “Policies to halt and reverse the rise in UPF production, marketing, and consumption.” pp 2685-2702. 10.1016/S0140-6736(25)01566-1, PMID 41270767	Four policy domains: products, food environments, corporations, supply chains
3	Baker P, Slater S, White M, Wood B, … Van Tulleken C, Nestle M, Barquera S. “Towards unified global action on ultra-processed foods.” pp 2703-2726. 10.1016/S0140-6736(25)01567-3, PMID 41270764	Commercial determinants; corporate political activity as the main barrier

Accompanied by a Lancet editorial “Ultra-processed foods: time to put health before profit” (10.1016/S0140-6736(25)02322-0), a UNICEF Comment, and a WHO/PAHO Comment [V].

Funding and conflicts [V, from the PubMed COI fields]. The Series was funded by Bloomberg Philanthropies through Deakin University, subcontracted to Melbourne, Sydney and Sao Paulo. Monteiro, Louzada, Cannon, Moubarac and Levy are the authors of the NOVA classification being evaluated. Van Tulleken declares book royalties on a UPF book; Khandpur is a paid UNICEF and PAHO consultant. This is a self-assessment of a framework by the people who built it, funded by an advocacy foundation, and should be read as such.

Correspondence, The Lancet 407, March 2026 [V]. Five letters plus an authors’ reply, all titled “Ultra-processed foods in research and policy”: Ludwig DS (10.1016/S0140-6736(26)00132-7), Kuhnle GGC (10.1016/S0140-6736(26)00106-6), Jacobs D & Sampson R for FoodDrinkEurope, a trade association (10.1016/S0140-6736(26)00107-8), Campos T & Esturo A for the International Fruit and Vegetable Juice Association (10.1016/S0140-6736(26)00108-X), and a supportive letter from Brazil’s Ministry of Social Development (10.1016/S0140-6736(26)00109-1). Authors’ reply: Monteiro, Rezende, Baker, Nestle, Corvalan, Popkin, 10.1016/S0140-6736(26)00292-8, PMID 41794433.

What each side concedes

The Series authors concede, in writing [V, verbatim from the reply]:

• NOVA “does not replace nutrient science, but adds a complementary layer,” and paper 2 “explicitly proposed that all regulations should combine criteria on crucial nutrients with markers of food ultra-processing, rather than treating processing as a stand-alone metric.” This is a significant retreat from the strong form of the NOVA thesis and is almost never reported.
• The measurement objection is granted outright: “We agree with Gunter G C Kuhnle that evidence on the harms of UPFs was generated using dietary instruments not designed to capture Nova groups.” Their defence is that the misclassification is non-differential and therefore biases toward the null.
• Heterogeneity within NOVA group 4 is real: “Some UPFs might perform better than others in specific comparisons, and relative harms might be modest in narrow contrasts.” They then add, “policy, however, cannot be built on marginal cases.”

What they refuse. They reject UPF subgroup analyses as “conceptually and methodologically flawed,” citing their own BMJ Analysis: Rezende LFM et al., “Misleading narrative of ‘healthy’ ultraprocessed foods,” BMJ 2026;392:e087538, 27 Jan 2026, 10.1136/bmj-2025-087538 [V article page; body paywalled]. They assert that 92.4% of US diets excessive in added sugar, saturated fat and energy density and insufficient in fibre are attributable to UPF consumption (citing Martinez Steele et al., Nutr Metab Cardiovasc Dis 2022;32:2739-50) [V that the claim is made; underlying paper not checked].

What the critics concede. In the parallel Nature Medicine exchange over the Dicken trial, Ludwig, Willett and Putt open by accepting that “extensive observational data link consumption of ultra-processed food … to obesity and other chronic health conditions” [V]. Their objection is to effect size and to the leap to guidelines, not to the existence of the association. In the Science Media Centre roundup on the Series [V], Kuhnle-adjacent skeptics concede the same: Beaumont (Sheffield Hallam, no declared COI) says there is “little convincing, high-quality evidence that ultra-processed foods are inherently unhealthy,” while Warren (Quadram) frames it as “strength of causal claims exceeds perhaps what the underlying data can fully support” and Griffin (Rowett) asks only for more RCTs. Industry bodies (FDF, FoodDrinkEurope, IFBA) concede the shared diet-quality goal and, in FDF’s case, that reformulation since 2015 was necessary [V via FoodNavigator, 20 Nov 2025].

Note on attribution. No Series-specific published response from Katz, Mozaffarian, Gibney or Astrup was located in Crossref, PubMed or the SMC roundup. Astrup’s known position is the 2022 AJCN head-to-head with Monteiro (116:1482-1488), which predates the Series. Do not attribute a Series critique to any of them without checking.

Tier: the Series is a narrative and systematic review of tier C evidence with tier A/B feeding trials layered in; the correspondence is tier G discourse.

---

1.2 The 2025-2026 whole-diet processing RCT wave

This is the single largest gap between the rubric and the current literature. The rubric knows Hall 2019, Dicken 2025 and the NIH factorial. At least ten further processing-level randomised trials published in 2025 and 2026, and they do not all point the same way.

1.2.1 The NIH four-arm factorial is complete, and processing per se was null

NCT05290064, “Effect of Ultra-processed Versus Unprocessed Diets on Energy Metabolism,” NIH Clinical Center / NIDDK. Status COMPLETED, actual enrolment 38, n=36 analysed, primary completion 2025-08-15, results posted to the registry. Not yet peer-reviewed; no linked PubMed record. [V, read directly from the ClinicalTrials.gov API v2 results section.]

Randomised crossover, four one-week inpatient ad libitum diets. LS-mean energy intake, kcal/day:

Arm	Energy intake
UPF, high energy density + high hyperpalatable (hh)	3424.3
UPF, high ED + low HP (hl)	3265.9
UPF, low ED + low HP (ll)	2604.4
Unprocessed, low ED + low HP (ll)	2476.0

Prespecified contrasts:

Contrast	Isolates	Difference (95% CI)	p
UPF-hh vs UPF-hl	hyperpalatability	+158.4 (21.9 to 294.9)	0.023
UPF-hl vs UPF-ll	energy density	+661.6 (524.9 to 798.3)	<0.0001
UPF-ll vs UNF-ll	processing alone	+128.4 (-8.2 to 265.0)	0.065
UPF-hh vs UNF-ll	everything	+948.4 (811.9 to 1084.8)	<0.0001

Two secondary results that nobody quotes. Measured eating rate was similar across all four arms: 39.21, 39.02, 39.04 and 38.43 g/min, every pairwise p = 0.99. Palatability VAS was likewise similar: 65.24, 62.41, 66.67, 64.18, every p >= 0.6.

[Corrected on review: an earlier version of this section read those numbers as “flat”, said the mediators “did not move at all”, and concluded that “neither can be carrying the 948 kcal”. That treated a limitation as a null.] Two reasons the reading fails:

1. The contrasts are imprecise, not zero. The eating-rate contrasts are 0.19, -0.02, 0.61 and 0.78 g/min with confidence intervals of roughly ±2.8 g/min, about ±7% of the mean. The palatability contrasts are +2.8 (-3.8, 9.5), -4.25 (-10.8, 2.3), +2.49 (-4.0, 9.0) and +1.06 (-5.5, 7.6). Intervals that wide exclude nothing of mechanistic interest.
2. g/min was measured while energy density roughly doubled. At equal g/min, kcal/min roughly doubles. A flat mass eating rate across an energy-density manipulation is what the energy-density account predicts, not evidence against an eating-rate mediator.

1.2.2 A 369 kcal/day contrast within ultra-processed food, attributed to eating rate

Forde CG, Heuven LA, van Bruinessen M, Liu Z, Stieger M, de Graaf K, Lasschuijt MP. “Eating rate has sustained effects on energy intake from ultraprocessed diets: a 2-week ad libitum dietary randomized controlled crossover trial.” Am J Clin Nutr 2026; 10.1016/j.ajcnut.2025.11.012, PMID 41314613, NCT06113146. Tier A/B. [V]

n=41 completers, block-randomised crossover, two 14-day ad libitum diets, 2-week washout. Both arms were ultra-processed. The intended contrast was food texture, and the diets were matched for palatability, portion served, total energy served, non-beverage energy density and meal variety. Registry masking is double, not single-blind.

[Corrected on review: the two arms did not differ only in texture.] Table 1, as served, gives fat 22 vs 33 EN%, carbohydrate 53 vs 47, mono- and disaccharides 15 vs 19, protein 21 vs 16, fibre 1.54 vs 1.48 g/100 kcal and UPF share 97 vs 94 EN%. As consumed, fat differed by 9 percentage points (F=1288, P<0.0001) and water intake differed by 132 g/day (P=0.019). The authors write: “This resulted in some differences in the macronutrient matching.” Source: https://europepmc.org/article/PMC/PMC13084570

• Daily energy intake 369 kcal/d lower on the slow-eating-rate UPF diet (95% CI 221 to 517), F(1,1051)=23.98, p<0.001.
• Effect constant across the 14 days (diet x time p=0.486).
• The claim that the effect was “not explained by macronutrients (all p>0.05)” is a post hoc covariate model, not a design property. Fat EN% in that model has P=0.082 and is collinear with arm, so the model cannot separate texture from fat.
• No body-weight difference; fat mass fell 0.43 kg on the slow arm (p=0.0002).

This is still one of the most informative trials in the document, but the defensible reading is narrower: texture-driven eating rate is a candidate lever within ultra-processed food, not a demonstrated sole cause. The clean texture manipulation is Lasschuijt 2023 (n=18, texture varied with composition held); Forde 2026 is larger and longer but confounded with fat, sugar and protein.

1.2.3 Match energy density and the effect disappears in young adults

Rego MLM, Leslie E, … DiFeliceantonio AG, Davy BM. “The Influence of Ultraprocessed Food Consumption on Energy Intake in Emerging Adulthood: A Controlled Feeding Trial.” Obesity 2026; 10.1002/oby.70086, PMID 41255123, NCT05550818. Tier B. [V]

n=27, aged 18-25, randomised crossover, two 14-day eucaloric controlled feeding periods (81% vs 0% energy from UPF), 4-week washout, matched for macronutrients, fibre, added sugar, diet quality and energy density. Diet compliance ~99%. Outcome was an ad libitum buffet after each period.

No effect of diet condition on total kcal or grams consumed in the full sample (all p>0.05). An exploratory diet-by-age interaction (p<0.001) showed increased intake after the UPF diet in late adolescents aged 18-21 (p=0.03, d=0.79) but not in 22-25 year olds.

1.2.4 A 12-month free-living UPF-restriction RCT that came out nearly null

Macena ML, Pereira MR, … Bueno NB. “Effectiveness and metabolic impacts of restricting the consumption of ultra-processed foods in individuals with obesity submitted to energy restriction: a randomized clinical trial.” Nutr Metab Cardiovasc Dis 2026; 10.1016/j.numecd.2025.104426, PMID 41381310. Tier A. [V]

n=148, randomised parallel, 12 months. Energy restriction alone (ER-G) vs energy restriction plus UPF restriction (ER-UPF). The ER-UPF arm did reduce its NOVA-UPF Score (2.74 to 1.86 vs 2.62 to 2.47, p=0.03) and lost more weight (82.9 vs 86.3 kg final, p=0.01), but the authors’ own verdict is that this was “only a statistically, but non-clinically significant, greater weight loss.” No other outcome moved. Baseline UPF intake was only ~21-24% of energy, which the authors flag as the likely reason.

This is the longest free-living processing RCT in existence and it is the strongest counterweight to Dicken 2025.

1.2.5 A eucaloric trial where glycaemia moved without weight moving

Capra BT, Hudson S, … Davy BM. “Chemical Analysis of Controlled Diets High in and Free of Ultraprocessed Foods and Proof-of-Concept Findings: Reducing Ultraprocessed Food Consumption May Lower Diabetes Risk in Midlife Adults.” J Nutr 2026; 10.1016/j.tjnut.2026.101370, PMID 41577034. Tier B (pilot). Peer reviewed in J Nutr. [V]

Adults 40-65, 2-week lead-in at 59% UPF, then randomised to eucaloric 81% UPF or 0% UPF for 6 weeks. Diets chemically assayed, not just modelled. Design correction on review: this is a parallel-group pilot, 18 participants in total, roughly 9 per arm.

• Matsuda and HOMA-IR: no change.
• Glucose AUC and MAGE both tended to worsen on high-UPF (p=0.054 and p=0.055, ES 0.52 and 0.51). Underpowered; directional only.
• [Corrected on review: the arms did not start from the same place.] Glucose AUC at baseline was 13,431 in the high-UPF arm versus 15,349 in the non-UPF arm. The non-UPF arm started worse and improved; the high-UPF arm started better and barely moved. With about 9 per arm and p = 0.054/0.055, regression to the mean is a plausible complete explanation.
• The novel bit, stated more carefully: the non-UPF arm showed reductions in the food-contact chemical 2,4-di-tert-butylphenol and in the thermal processing by-product N6-carboxymethyllysine. The abstract reports only that “reductions were observed”, with no effect size and no p value. Treat it as exposure measurement inside a feeding trial, not as a quantified result.

1.2.6 UPF moved reproductive and lipid markers, but weight moved too

Preston JM, Iversen J, … (Cell Metab consortium). “Effect of ultra-processed food consumption on male reproductive and metabolic health.” Cell Metabolism 2025; 10.1016/j.cmet.2025.08.004, PMID 40882621, NCT05368194. Tier B, contested. [V for the paper; per-arm p-values below come from secondary sources and are not verified]

Design. n=43 men aged 20-35, controlled 2 x 2 crossover crossing processing level with caloric load, 3 weeks per diet, 12-week washout, with a +500 kcal excess arm alongside the “adequate” arm. The registered primary outcome is sperm DNA methylation (NCT05368194), not body weight and not lipids. Everything quoted below is secondary.

What was reported: increased body weight and increased LDL:HDL ratio on the UPF diet; decreased GDF15 and FSH; sperm total motility trended down. Differential pollutant accumulation was detected between the diets, including decreased plasma lithium and a trend toward increased serum mono(4-methyl-7-carboxyheptyl)phthalate (cxMINP) on the UPF diet.

[Corrected on review: this section previously called the trial “the strongest published counter-example to a pure energy-density account”. Withdrawn.] Body weight rose 1.3 to 1.4 kg on UPF in both arms, so the “adequate” arm was not eucaloric in outcome whatever it was in design: calories, or absorbed energy, moved. The findings are also arm-split rather than uniform. The LDL:HDL rise is reported in the adequate arm only, the FSH fall in the excess arm only, and the motility change was not significant after correction.

The published correspondence, omitted from the earlier version of this document (Cell Metabolism 38(1), January 2026):

Item	Authors	PMID
Letter	Ludwig DS	41500195
Letter	Matthiessen, Rosane, Mosig, Ahrné, Magkos, Bügel	41500197
Authors’ reply	Barrès, Simpson, Nóbrega, Preston	41500199

1.2.7 Protein-fortifying ultra-processed food does not fix it

Hagele FA, Herpich C, … Bosy-Westphal A. “Short-term effects of high-protein, lower-carbohydrate ultra-processed foods on human energy balance.” Nature Metabolism 2025; 10.1038/s42255-025-01247-4, PMID 40082711. Tier B. [V, full text via PMC12021659]

n=21, single-blind randomised crossover, 54 h in a whole-room calorimeter, both arms 84% UPF, ad libitum, matched palatability.

	HPLC-UPF (30% protein)	NPNC-UPF (13% protein)
Energy intake	2,465 ± 672 kcal/d	2,832 ± 688 kcal/d
Energy expenditure	+128 ± 98 kcal/d vs NPNC	reference
Energy balance	+18%	+32%

Protein enrichment reduced intake by 196 ± 396 kcal/d and raised expenditure, but both arms were in positive energy balance. Even at >3 g/kg body weight protein the participants overate. The accompanying commentary is titled “More protein in ultra-processed foods: no shortcut to eating less” (10.1038/s42255-025-01258-1) [V].

1.2.8 Acute Brazilian factorial trials: null on intake, positive on insulin

Two reports from one 2 x 2 crossover (ReBEC RBR-10gfr3fb, n=19 young lean adults, meals matched for energy, macronutrients, fibre, saturated fat, free sugar and sodium) [V]:

• Barros LM et al., Appetite 2026; 10.1016/j.appet.2026.108761, PMID 42633852. Preload meals at 0% vs >=80% UPF crossed with energy density <=1.2 vs >=2.0 kcal/g. No main effect and no interaction on subsequent ad libitum intake (UPF -24.5 kcal, 95% CI -87 to 38; ED -41.6, -104 to 20.9). High-ED preloads moved appetite VAS by 3-5 mm without moving intake.
• Gama IRS et al., Nutrients 2026; 10.3390/nu18152534, PMID 42588167. Same design, hormonal endpoints. UPF raised postprandial insulin across the day (p=0.016) with no ED main effect and no interaction. Exploratory secondary analysis; authors call for confirmation.

1.2.9 Mechanism: micronutrient deleveraging

Brunstrom JM, Schatzker M, Rogers PJ, Courville AB, Hall KD, Flynn AN. “Consuming an unprocessed diet reduces energy intake: a post-hoc analysis of a randomized controlled trial reveals a role for human nutritional intelligence.” Am J Clin Nutr 2026; 10.1016/j.ajcnut.2025.101183. Tier D (post hoc). [V]

Re-analysis of Hall 2019 (n=20). Unprocessed meals were 665.5 g vs 423.5 g and 1.08 vs 1.96 kcal/g. Two predictors, a carbohydrate-fat “blend index” and a fruit/vegetable score, accounted for 87.6% of the 330 kcal per-meal difference (r=0.78). The proposed mechanism, “micronutrient deleveraging,” is the mirror of protein leverage: on an unprocessed diet you must eat low-energy-density, micronutrient-dense items to meet micronutrient needs, and that constraint suppresses energy intake. On a UPF diet, fortification co-locates micronutrients with calories and the constraint vanishes.

1.2.10 An outcome trial in liver disease

Bo S, Armandi A, … Bugianesi E. “Impact of weight loss and reduction of ultra-processed foods on liver fat content in MASLD: A randomized controlled trial.” JHEP Reports 2026; 10.1016/j.jhepr.2026.101929, PMID 42331287. Tier A for the randomised comparison, tier D for the UPF finding. [V]

n=173 enrolled, 148 completed, 6 months, three arms (Mediterranean, low-carb/high-protein, standard advice). No between-group difference in liver fat (CAP). Structural equation modelling showed no direct or total effect of diet pattern, but reductions in BMI (beta 14.34, 95% CI 10.55-18.13) and in UPF intake (beta 0.64, 95% CI 0.25-1.03) independently predicted CAP improvement. UPF reduction was not randomised, so this is an observational finding inside a trial.

1.2.11 Cohort side: an updated dose-response picture

• Umbrella review through Dec 2025: Nutrition Reviews 2026, 10.1093/nutrit/nuag099, PMID 42341183. Twelve lead meta-analyses, 13 outcomes. GRADE certainty high for 1 outcome, moderate for 6, low for 6. T2D shows ~10% higher risk per 10% increase in UPF; CCVD curvilinear; NAFLD not significant. [V]
• All-cause mortality dose-response: Syst Rev 2025, 10.1186/s13643-025-02800-8, PMID 40033461. 18 cohorts, n=1,148,387, 173,107 deaths. Highest vs lowest HR 1.15 (1.09-1.22), 10% per 10% increment, linear. I² = 83-91%. [V]
• Digestive cancers: Front Nutr 2026, 10.3389/fnut.2026.1901660, PMID 42528625. Nine cohorts. Overall HR 1.12 (1.05-1.20), but ultra-processed meat/protein products HR 1.33 (1.15-1.53). [V] The subgroup again carries the aggregate.

1.2.12 Objective biomarkers of UPF intake now exist

Abar L, Martinez Steele E, Lee SK, Kahle L, Moore SC, Watts E, et al. “Identification and validation of poly-metabolite scores for diets high in ultra-processed food: An observational study and post-hoc randomized controlled crossover-feeding trial.” PLoS Medicine 2025; 10.1371/journal.pmed.1004560. Tier C plus tier B validation. [V]

IDATA, n=718, serum and urine metabolomics. Continuous UPF prediction r >= 0.47 in all three matrices. Binary high-vs-low classification AUC 0.75 train / 0.66 test (serum), 0.78 / 0.72 (24-h urine), 0.77 / 0.68 (first morning void). Validated in the n=20 Hall crossover: scores differed within-individual between the 80% and 0% UPF phases (all paired-t p<0.001).

Modest discrimination, but it is the first objective UPF exposure measure, and it is precisely the instrument that answers Kuhnle’s objection, which the Series authors conceded.

1.2.13 A continuous processing metric that dissolves the categories

Roos YH, Forde CG. “Exposure to processing from ultraprocessed diets and in feeding studies.” Current Research in Food Science 2026; 10.1016/j.crfs.2026.101504, PMID 42502741. Tier D (methodological). [V]

Proposes Food Processing Levels (FPL), a transformation-based scale separating physical from chemical modification, plus Processed Food Intake (PFI) descriptors that are item-, energy- or nutrient-weighted. Applied to the feeding-study corpus, the finding is that “diets classified as UPF result in a wide variation in cumulative processing exposure, while substantial overlap is observed between UPF and non-UPF diets when processing is expressed quantitatively.”

If that holds, several of the trials in section 1.2 are comparing overlapping distributions and calling them categories.

---

1.3 Sweeteners after aspartame 2B and the erythritol/xylitol signals

What changed. The erythritol story moved decisively toward reverse-causation, and the definitive RCT is still not published. Meanwhile EFSA quietly raised two sweetener ADIs.

1.3.1 The erythritol confound is now well documented

The rubric’s Tier 2 entry weights erythritol at 0.5 on the ground that “the observational signal is likely reverse causation; the pharmacodynamics survive that critique.” The 2024-2026 work supports the first clause more strongly than the rubric does, and puts real pressure on the second.

The only human platelet data remain Hazen’s own uncontrolled arms. Witkowski et al., ATVB 2024;44(9), 10.1161/ATVBAHA.124.321019, PMID 39114916: n=10 erythritol vs n=10 glucose, single 30 g bolus, prospective interventional, within-subject, not randomised, not placebo-controlled. Plasma erythritol rose >1000-fold (6,480 vs 3.75 umol/L). Tier B at best. [V] Xylitol: Eur Heart J 2024;45(27):2439, PMID 38842092, discovery n=1,157 / validation n=2,149, MACE HR 1.57 (1.12-2.21) tertile 3 vs 1, with an interventional arm of n=10, again unrandomised and uncontrolled. [V]

Evidence that plasma erythritol is a marker of endogenous pentose-phosphate flux, not of intake:

Study	Design / n	Finding
ATBC, Nutrients 2024;16:3099, PMID 39339699 [V, https://pubmed.ncbi.nlm.nih.gov/39339699/]	n=4,468 Finnish male smokers, serum banked 1985-1993, 19.1 y follow-up	Total mortality HR 1.50 (1.17-1.92); CVD mortality 1.86; cancer mortality 1.54
ARIC, JACC Advances 2025, 10.1016/j.jacadv.2025.101605 [V, https://pmc.ncbi.nlm.nih.gov/articles/PMC11889355/]	n=4,006; “median 8.4 y” is not verified	Total mortality: erythritol 1.18 (1.10-1.26), erythronate 1.61. CVD death: erythritol 1.19, erythronate 1.72 (1.43-2.06). Intake never measured
IJMS 2025;26:9763, PMID 41097029 [V]	n=218 across BMI strata	Age, not BMI or glycaemia, predicted fasting erythritol; post-bariatric, BMI change predicted erythritol change

Serum drawn largely before the sweetener was in wide use still predicts mortality, and in ARIC the pure pentose-phosphate metabolite outperforms erythritol itself. That is strong support for the marker interpretation of the cohort signal.

One correction on review. The “before commercial erythritol existed” framing is tighter than the facts. Commercial erythritol sweetener use began in the 1990s, and erythritol also occurs naturally in fruit, mushrooms and fermented foods. So “these serum levels are endogenous only” is an inference from timing and fasting status, not a measurement: no cohort in this table measured intake.

FDA’s own 15 June 2023 Office of Food Additive Safety memo [V, https://www.fda.gov/media/182122/download] says so explicitly: ~80% of the Witkowski 2023 samples were collected 2001-2007 before widespread US erythritol use, all subjects were fasting, so “detected erythritol levels are likely a reflection of endogenous production”; no dietary exposure was measured in any cohort; the mouse FeCl3 model reached ~290 uM, far above human observational levels; and at the 6 mM actually reached after a 30 g drink, aggregation was “only slightly higher” than at micromolar doses. FDA did not change erythritol’s GRAS status.

The sentence the earlier version of this document omitted, from the same memo [V, https://www.fda.gov/media/182122/download]: the mechanistic studies “suggest that erythritol can augment platelet activation and enhance thrombosis potential”, and the memo recommends coagulation testing after ingestion. The memo is therefore not a clean exculpation: it rejects the cohort exposure interpretation while taking the pharmacodynamic question seriously enough to ask for more testing.

Mendelian randomisation is split, and every arm of it is weak. The null study is Khafagy, Paterson, Dash, Diabetes 2024;73(2):325, PMID 37939167, 10.2337/db23-0330, bidirectional MR, no support for erythritol raising CAD (b = -0.033 ± 0.02, P = 0.14) [V]. [Corrected on review: this was previously called “the best-designed” MR. Withdrawn.] It draws its instruments from the same small erythritol GWAS that the positive MRs use, so it inherits the identical weak-instrument problem: a null there is low-powered, not clean. Two positive MRs exist (Am J Prev Cardiol 2025, 10.1016/j.ajpc.2025.101325, CHD OR 1.0020; Medicine 2025, PMID 41137360, CHD OR 1.077) but rest on erythritol GWAS of n=291 [not verified] and n=8,167 respectively, produce near-unity odds ratios, and show detected horizontal pleiotropy for the venous outcomes [V]. The honest summary is that MR is uninformative here in both directions.

Published rebuttals to the xylitol paper [V]: Wolnerhanssen, Meyer-Gerspach, Arduini, Eur Heart J 2025;46(3):328, 10.1093/eurheartj/ehae730, noting that intravenous xylitol is licensed at 3 g/kg/day in Germany and Japan with no thrombotic signal. COI: Arduini is founder and major shareholder of CoreQuest and Iperboreal Pharma. Companion editorial: Bonomini, Masola, Gronda, Eur Heart J 2025;46(3):326, “an active component or an innocent bystander?”

The one new mechanistic study that cuts the other way. Berry et al., J Appl Physiol 2025;138(6):1571, 10.1152/japplphysiol.00276.2025, PMID 40459966 (DeSouza lab). Human cerebral microvascular endothelial cells at 6 mM erythritol, the exact plasma level after 30 g, for 3 h: ROS 204% vs 105%, p-eNOS Ser1177 down and Thr495 up, NO down (5.8 vs 7.3 umol/L), ET-1 up, thrombin-stimulated t-PA release abolished. Tier E. A caution letter followed (J Appl Physiol 2025;139, PMID 41379634) [V].

The trial that settles it is still pending. NCT05967741, UC Davis, randomised triple-blind crossover, n=24, erythritol- vs aspartame-sweetened beverages 2 weeks each, primary outcomes P-selectin, PAC-1, annexin V, agonist-stimulated aggregation and platelet-leukocyte aggregates. Dose 1 g/kg/day, roughly 70 g/day, which is well above ordinary dietary exposure and above the 30 g bolus used by Witkowski. Status RECRUITING, estimated completion 2026-05-31, registry last updated 2025-12-05, so a 2026 readout should not be assumed. [V, CT.gov API]

Note on a recurring mix-up: Wagner 2024 (BMC Medicine) is a carrageenan trial (section 1.6.5). It has no bearing on erythritol.

1.3.2 Non-nutritive sweeteners generally

• NutriNet-Sante, Diabetes Care 2023;46:1681, PMID 37490630 [V]: n=105,588, 9.1 y, 972 incident T2D. Total sweeteners HR 1.69 (1.45-1.97); aspartame 1.63; acesulfame-K 1.70; sucralose 1.34. One of very few cohorts that separates by compound.
• UK Biobank, Cardiovasc Diabetol 2024;23, PMID 38965574 [V]: n=133,285. Per-teaspoon CVD HR 1.012 (1.008-1.017). 70% of the CVD association was mediated by prior T2D, a strong reverse-causation flag.
• The best substitution RCT is industry-funded. Harrold et al., Int J Obes 2023;47, 10.1038/s41366-023-01393-3, NCT02591134 [V]: n=493, 52 weeks, open-label. NNS beverages -7.5 kg vs water -6.1 kg, between-group 1.4 kg favouring NNS; the groups were formally non-equivalent. 53% completion. Funded by the American Beverage Association.
• The UK government pushed back on WHO. SACN position statement on non-sugar sweeteners, April 2025 [V, https://assets.publishing.service.gov.uk/media/67ea97b3ea9f8afd8105627d/sacn-position-statement-on-non-sugar-sweeteners.pdf]. Core objection: “WHO gave greater weight to PCS evidence over RCT evidence.” SACN concludes RCTs “consistently suggest NSS, compared with free sugars, reduce energy intake and therefore body weight,” and that NSS “may be some value … in the short to medium term, but is not essential.”
• Sucralose brain imaging. Chakravartti et al., Nature Metabolism 2025;7, 10.1038/s42255-025-01227-8, PMID 40140714, NCT02945475 [V]. Randomised crossover, n=75. Sucralose vs sucrose raised hypothalamic blood flow (p<0.018) and hunger (p<0.001), and increased hypothalamic connectivity to motivation and somatosensory regions; effects largest in obesity. Tier B. This is the first well-powered human imaging evidence for the long-hypothesised uncoupling of sweetness from calories.
• EFSA raised two sweetener ADIs on re-evaluation (see 1.6).

---

1.4 Nitrite and nitrate: the source split survives, the mechanism does not

What changed. The additive-vs-vegetable split held up in new cohort work, but the two studies designed to isolate nitrite as the causal agent both came out negative. The rubric’s position (“uncured is marketing, the same 156 ppm ceiling applies”) is unaffected and arguably strengthened, but the reason is different from the one usually given.

1.4.1 The cohort split is reproducible

Study	n / design	Result
Chazelas, Int J Epidemiol 2022;51, PMID 35303088 [V]	NutriNet-Sante n=101,056, 6.7 y, 3,311 cancers	Additive nitrates and breast cancer HR 1.24 (1.03-1.48); additive nitrite (E250) and prostate cancer HR 1.58 (1.14-2.18). “No association was observed for natural sources”
Srour, PLOS Med 2023;20, 10.1371/journal.pmed.1004149 [V]	n=104,168, 7.3 y, 969 T2D	Additive nitrites HR 1.53 (1.24-1.88); natural-source nitrites 1.26 (1.03-1.54); no association for nitrates from any source
Nature Communications, 7 Jan 2026, 10.1038/s41467-025-67360-w [V]	NutriNet-Sante n=108,723, 8.05 y, 1,131 T2D	Sodium nitrite E250 HR 1.50 (1.26-1.78); potassium nitrate E252 null
PLOS Med 2025;22, PMID 40198579 [V]	n=108,643, additive mixtures	2 of 5 NMF-derived mixtures raised T2D risk; one sweetener-containing mixture HR 1.13 per SD

The buried problem in the 2026 Nature Communications paper: potassium sorbate scored HR 2.15, sodium ascorbate 1.41 and sodium erythorbate 1.43, equal to or higher than nitrite. Sodium ascorbate and erythorbate are the cure accelerators added specifically to block nitrosamine formation. When the nitrosation inhibitor scores as badly as the nitrosating agent, the model is reading “cured meat,” not chemistry.

1.4.2 The two studies that tried to isolate nitrite both failed to incriminate it

EPIC, nitrosyl-heme, and colorectal cancer. Rizzolo-Brime et al., “Dietary nitrosyl-heme from processed meats and its association with colorectal cancer risk,” Nutrition Journal 2025, PMID 41422247, 10.1186/s12937-025-01266-7. Tier C, and the best available direct test so far. [V] n=367,463 across seven countries, 15-year median, 5,115 incident CRC. HR T3 vs T1 = 1.01 (0.93-1.09) for incident colorectal cancer, null overall, at every subsite and in every lifestyle subgroup. WHO/ISCIII funded. https://pubmed.ncbi.nlm.nih.gov/41422247/

Two limits added on review. First, the exposure was assayed in 52 Spanish products and then extrapolated onto the FFQ items of six other countries, so the exposure contrast in most of the cohort is modelled, not measured. Second, nitrosyl-heme is one route of the nitrite hypothesis, not “the exact molecular species” it names [phrase withdrawn on review]: endogenous nitrosation in the colon, volatile nitrosamines formed on cooking and nitrosothiols are separate routes that this exposure does not capture. The HR is for one product of one route against one outcome.

Removing nitrite from cured meat did not help, in a controlled animal experiment. Gueraud et al., npj Science of Food 2023;7, 10.1038/s41538-023-00228-9. Tier E. [V] Azoxymethane-treated F344 rats, six diets, n=11-12/group, 98 days. Cutting nitrite from 120 to 90 mg/kg reduced mucin-depleted foci; complete removal produced no further benefit because it “induced a strong increase of lipid peroxidation” (TBARS, urinary DHN-MA). Vegetable-stock replacement behaved like full nitrite for N-nitroso compound formation, so “clean label” celery-powder cures are not mechanistically different. 90 mg/kg controlled Listeria as well as 120. COI: 38% co-financed by IFIP and an industry consortium, with IFIP employees among the authors, so treat directionally.

1.4.3 Why the source matters, and the vegetable side is weaker than advertised

The exposure arithmetic reframes the debate. ANSES revised opinion, 12 July 2022 [V]: roughly two-thirds of nitrate exposure comes from plant foods and about a quarter from water, with under 4% from charcuterie additives; for nitrite, over half of exposure is charcuterie additives. About 99% of the French population sits below the ADI.

The “nitrate from vegetables is good for you” claim is real but smaller than usually quoted, and it is matrix-dependent:

• Beetroot juice in hypertensives: Gronroos et al., Nutr Metab Cardiovasc Dis 2024;34, PMID 39069465 [V]. 11 RCTs, n=349, 200-800 mg nitrate/day. Clinic SBP -5.31 mmHg (-7.46 to -3.16), I²=64%, but no effect on DBP and no effect on any 24-hour ambulatory outcome. GRADE certainty low.
• Isolated sodium nitrate salt: Forster et al., Nutrition Research 2026;146, PMID 41619653 [V]. 6 RCTs, n=181, >=1 week, beetroot excluded. SBP -3.81 (-10.05 to 2.43) and DBP -2.00 (-4.37 to 0.38), both null.

Beetroot works and the isolated ion does not. That is the same matrix argument (ascorbate and polyphenols) running in the opposite direction from the nitrosation-blocking argument on the cancer side, and it is the closest thing to a mechanistic resolution the field has.

---

1.5 Microplastics, PFAS and packaging migration

What changed. The headline human outcome study is now under a specific, technical, and largely unanswered analytical challenge, and the challenge lands on exactly the two polymers it reported.

1.5.1 The tier structure, stated bluntly

• Detected in human tissue (tier F): abundant and growing, analytically contested.
• Associated with an outcome in humans (tier C/D): roughly four studies, all cross-sectional or single-cohort, all sharing the same exposure-measurement problem.
• Causal in humans (tier A/B): zero. No RCT, no Mendelian randomisation, no dose-response with a validated exposure measure.

1.5.2 Marfella 2024, and the method critique that undermines it

Marfella R et al., NEJM 2024;390(10):900-910, 10.1056/NEJMoa2309822, PMID 38446676, NCT05900947 [V]. Prospective multicentre observational; 304 enrolled, 257 followed a mean 33.7 ± 6.9 months after carotid endarterectomy. Polyethylene detected in 150 (58.4%), PVC in 31 (12.1%). Composite MI/stroke/death HR 4.53 (2.00-10.27), p<0.001. Method: pyrolysis GC-MS plus stable isotope plus electron microscopy.

The decisive technical hit [V]: Rauert C, Charlton N, Bagley A, Dunlop SA, Symeonides C, Thomas KV. Environ Sci Technol 2025;59(4):1984-1994, 10.1021/acs.est.4c12599, PMID 39851066. With validated extraction, realistic limits of detection were up to 20x higher than nominal and recoveries ranged 7-109%. Verbatim conclusion: “Py-GC-MS is currently not a suitable analysis method for PE and PVC in biological matrices due to the presence of interferences and nonspecific pyrolysis products.” Those are precisely and only the two polymers Marfella reported. Corroborated by Almeida et al., J Hazard Mater 2026;501:140658, PMID 41352016: tissue lipids generate PE false positives, and eliminating them by using >C22 markers raises the PE limit of quantification to 50 ug/g versus 1 ug/g for PS, PP and PMMA.

No independent replication exists. The nearest is the same group: Paolisso, Scisciola, … Marfella, Barbato, Eur Heart J 2026, ehag447, 10.1093/eurheartj/ehag447, PMID 42447841 [V]. Cross-sectional, n=61. Detection 84.2% (STEMI) vs 40% (CCS) vs 31.8% (controls), p=0.002. But in multivariable analysis, smoking history was the only independent predictor of microplastic presence (OR 5.69, 1.33-26.63, p=0.023), with PM2.5 co-associated. That points at confounding by combustion exposure rather than an independent plastics effect.

Other human outcome work, all tier C/D [V]: PIAMA n=100 (immune-cell surface markers, no association with lung function), Microplast Nanoplast 2026;6(1):39; cord blood n=151 with birth weight beta -29.4 g, Environ Pollut 2026;405:128550; systematic review of 25 studies, Environ Health 2026;25(1):26, ROBINS-E moderate-to-high risk of confounding and exposure-measurement bias, no meta-analysis attempted.

Brain detection. Nihart, Garcia, El Hayek … Campen, Nature Medicine 2025;31(4):1114-1119, 10.1038/s41591-024-03453-1, PMID 39901044 (erratum 31(4):1367) [V]. Decedent liver, kidney and brain, using Py-GC/MS plus ATR-FTIR plus EM-EDS, which is better orthogonality than Marfella. Death year 2016 vs 2024 differed (p=0.01); dementia brains higher, with the authors explicitly disclaiming causality. Reverse causation (impaired clearance, blood-brain-barrier breakdown) is at least as plausible. The viral “a spoonful of plastic in your brain” mass figures rest on the PE quantification Rauert says is not defensible. Tier F.

1.5.3 Food contact chemicals: a good exposure map, no outcomes

Geueke B, Parkinson LV, Groh KJ, Kassotis CD, Maffini MV, Martin OV, Zimmermann L, Scheringer M, Muncke J. J Expo Sci Environ Epidemiol 2025;35(3):330-341, 10.1038/s41370-024-00718-2, PMID 39285208 [V]. Systematic evidence map: >14,000 known food contact chemicals, evidence of presence in humans for 3,601 (25%); 194 from biomonitoring programmes, of which 80 have hazard properties of high concern; 59 prioritised FCCs have no hazard data at all. Other exposure routes exist for many. Tier F: exposure mapping with zero outcome data.

The one place this connects to diet-quality research is Capra 2026 (section 1.2.5), where a 0% UPF arm lowered plasma 2,4-di-tert-butylphenol, and Preston 2025 (section 1.2.6), where a UPF arm trended toward higher serum cxMINP phthalate. Those are the first feeding trials to measure food-contact chemical exposure alongside a processing intervention. [Corrected on review: an earlier version said “as a mediator”. Withdrawn. Capra measured 2,4-DTBP but ran no mediation analysis, and neither trial estimated an indirect effect. Measuring a candidate mediator is not mediation.]

1.5.4 PFAS

• EPA final drinking-water rule [V]: Federal Register 2024-04-26, doc 2024-07773, effective 2024-06-25. MCLs 4.0 ppt PFOA and PFOS; 10 ppt PFHxS, PFNA, HFPO-DA; hazard index for the mixture. https://www.federalregister.gov/documents/2024/04/26/2024-07773/
• The 2026 reversal [V]: two EPA proposed rules published 2026-05-20: rescission of the regulatory determinations for PFHxS, PFNA, GenX and the mixture (2026-10085), and extension of the PFOA/PFOS compliance deadlines (2026-10086). PFOA and PFOS MCLs retained. Tier G, deregulatory, reversing a rule with one of the better epidemiological bases in this document.
• EFSA TWI 4.4 ng/kg bw per week for the sum of PFOA + PFOS + PFNA + PFHxS, critical effect explicitly reduced antibody response to vaccination in children, EFSA J 2020;18(9):6223 [V].
• 2024-2026 human evidence is mixed on the vaccine endpoint and inconsistent on breast cancer (Ronneby, ACS CPS-II, Danish Diet Cancer and Health) [V titles, effect sizes not extracted]. Lipid associations are consistent but cross-sectional, and reverse causality via albumin binding and enterohepatic recirculation is unresolved. The C8 Science Panel remains the only large “probable link” body and is legacy high-exposure, not dietary.
• PFAS in food specifically is occurrence data only. Environ Sci Technol 2026, “PFAS in Fast Foods with Hot-Contact to Food-Contact Materials,” 10.1021/acs.est.6c03801 [V title]. No study links wrapper or popcorn-bag PFAS to any clinical outcome.

1.5.5 Bisphenols and phthalates

• EFSA’s 2023 BPA TDI of 0.2 ng/kg bw/day, a ~20,000-fold cut driven by a single mouse Th17 endpoint, was not endorsed by EMA or BfR; BfR derived 0.2 ug/kg bw/day, a thousand-fold higher [R]. Documented in Environ Health Perspect 2024;132(4), 10.1289/EHP13812; rebutted in Toxicol Sci 2024;198(2):185 [R].
• EU BPA ban: Commission Regulation (EU) 2024/3190, adopted 2024-12-19, in force 2025-01-20, prohibiting BPA in food-contact plastics, coatings, varnishes, inks, adhesives, ion-exchange resins, silicones and rubber, with 18-48 month transitions [R]. Tier G.
• The widely-circulated phthalate mortality figure is modelling, not a cohort: Hyman, Acevedo, Giannarelli, Trasande, EBioMedicine 2025;117:105730, 10.1016/j.ebiom.2025.105730, PMID 40307157 [V]. Published hazard ratios applied to IHME mortality and regional DEHP estimates, yielding “356,238 global deaths in 2018.” Funded by Bloomberg Philanthropies and Beyond Petrochemicals. Three-decimal precision on a modelled attributable fraction is a presentation red flag. Tier D.

---

1.6 Additive re-evaluations, 2024-2026

What changed. A large amount of regulatory motion, most of it not driven by new toxicology. Separating the two is the whole point of this section.

1.6.1 FDA

Action	Date / citation	Basis	Driven by
FD&C Red No. 3 revoked	FR 2025-01-16, doc 2025-00830, docket FDA-2023-N-0437; effective 2027-01-15 food, 2028-01-18 drugs [V]	Delaney Clause, triggered by male-rat thyroid tumours via a hormonal mechanism FDA states does not occur in humans	Law, not new data
Brominated vegetable oil revoked	FR 2024-07-03, doc 2024-14300, docket FDA-2023-N-0937, effective 2024-08-02 [V]	FDA/NIH studies showing bromine accumulation and thyroid effects	New data
Synthetic dye phase-out request	HHS/FDA announcement 2025-04-22 [R]	Voluntary request covering Blue 1, Blue 2, Green 3, Red 40, Yellow 5, Yellow 6 by end-2026; Orange B and Citrus Red 2 to be revoked	Precaution/politics. No authorisation revoked; all six remain legal
Post-market chemical review programme	Finalised 2026-05-12, two documents [V]	Reassessments launched for BHT (docket FDA-2026-N-2526, comments to 2026-08-31) and azodicarbonamide (FDA-2026-N-4126)	Process reform
GRAS self-affirmation reform	HHS directive 2025-03-10; proposed rule amending 21 CFR 170/570 on the Unified Agenda, slipped Oct to Dec 2025 [R]	Would end self-affirmation	Politics

State level [V for AB 418, R for the rest]: California AB 418 bans brominated vegetable oil, potassium bromate, propylparaben and Red 3 from 2027-01-01 (penalties $5,000 / $10,000). California AB 2316 removes six dyes from public-school food from 2027. West Virginia HB 2354 was signed March 2025, but in December 2025 Judge Irene Berger preliminarily enjoined the general adulteration provisions (International Association of Color Manufacturers); the school provisions stand. Texas SB 25 warning labels are preliminarily enjoined and on appeal. Roughly 118 additive bills were tracked across 2025 state sessions.

Relevant to the rubric: potassium bromate is banned in California from 2027, which the rubric already records. BHT is now under active FDA reassessment, which the rubric does not.

1.6.2 EFSA: the sweeteners were cleared, and two ADIs went up

All [V] via PubMed DOIs; four abstracts read in full.

Additive	Opinion	Outcome
Saccharin E954	EFSA J 2024;22(11):e9044	Cleared
Acesulfame K E950	2025;23(4):e9317	ADI raised 9 to 15 mg/kg bw/day. “No new studies suitable for identification of a reference point on adverse effects”
Neotame E961	2025;23(7):e9480	ADI raised 2 to 10 mg/kg bw/day
Sucralose E955	2026;24(2):e9854	ADI 15 mg/kg bw/day retained, no safety concern at current uses. Could not conclude on extension into fine bakery wares, because of possible chlorinated compound formation during baking
Silver E174	2025;23(4):e9316	Not cleared. Industry submitted inadequate physicochemical characterisation, so safety remains unassessable. This is the same data-gap route that removed titanium dioxide E171
Various gases, pullulan E1204, shellac E904, quillaia E999, guar gum E412 (infants), E472c, silicon dioxide E551	2024-2025	Cleared

Corrections to common assumptions [V]: there is no 2024-2026 EFSA opinion on E471 mono- and diglycerides (last re-evaluated 2017, infant follow-up 2021), on carrageenan E407 (2018 temporary group ADI 75 mg/kg bw/day still stands; a data call reportedly closed 2025-11-25 with no new opinion published [R]), on aspartame (no new EFSA opinion since IARC 2B and the JECFA ADI reaffirmation of July 2023), or on phosphates or sulfites (the 2025 output is an exposure update, sp.efsa.2025.EN-9754, not a re-evaluation). Titanium dioxide E171 remains banned in EU food since 2022 with no reversal, and FDA has taken no action.

New mechanistic evidence on aspartame, which no regulator has yet weighed. Wu et al., Cell Metabolism 2025;37(5), PMID 39978336: dietary aspartame at 0.15% raised insulin in mice and in monkeys and aggravated atherosclerosis in ApoE-/- mice, with the effect traced to CX3CL1 signalling. Tier E, mechanistic. Animal and non-human primate only, no human outcome data, and it does not move any weight on its own. It is listed here because it is the only substantive new aspartame biology since the 2023 IARC 2B classification and the JECFA ADI reaffirmation.

Nitrites and nitrates [V]: EFSA’s 2017 opinions set the nitrite ADI at 0.07 mg nitrite ion/kg bw/day and nitrate at 3.7 mg/kg bw/day. Commission Regulation (EU) 2023/2108 lowered maximum levels for E249-E252 by roughly 20%, applicable from 9 Oct 2025 with a second phase in Oct 2026 [R]. The recitals cite nitrosamine minimisation and the IARC 2015 processed-meat classification, not new toxicology.

1.6.3 Titanium dioxide now has human intervention data

Bischoff NS, Undas AK, van Bemmel G, van Herwijnen M, Verheijen M, Van Breda SG, Briede JJ, et al. “Dietary Titanium Dioxide (E171) Alters the Colon Transcriptome: Evidence From a Human Dietary Intervention Study.” Molecular Nutrition & Food Research 2026; 10.1002/mnfr.70583, PMID 42613907. Tier B. [V]

Randomised crossover, n=31 adults, 2 mg/kg body weight/day of E171 for 2 weeks. Faecal titanium rose from 6.3 to 377 mg/kg, confirming exposure. Median particle size 228 nm with 9% below 100 nm.

What the signal actually is (https://pmc.ncbi.nlm.nih.gov/articles/PMC13487354/): 73 upregulated KEGG pathways by gene-set enrichment analysis of colon biopsy transcriptomes (Benjamini-Hochberg FDR < 0.05). “Colorectal cancer” is one of those KEGG pathway labels, not an observed lesion. There is no comet assay, no micronucleus assay, no 8-oxo-dG and no histology in this study.

What was null or borderline: hs-CRP P=0.076; eight cytokines null; faecal calprotectin null; whole-blood superoxide geometric mean ratio 1.175 (0.995-1.389), P=0.057. Blinding is not described in the report.

This is the first human intervention study on E171, and it changes the type of evidence available (a human randomised exposure with a target-tissue molecular readout) without changing what is known about outcomes. A supporting rodent systematic review appeared the same year (Part Fibre Toxicol 2026; 10.1186/s12989-025-00651-8, 54 studies: minimal acute accumulation, but dose-dependent accumulation in liver, spleen, kidney, GI tract and brain with subacute and subchronic exposure) [V].

1.6.4 Emulsifiers: human biomarker data arrived

Three items the rubric does not have.

ENIGMA: first human quantification of polysorbate-80 metabolism. Zhang J, Hu J, Tang X, … Gut 2026; 10.1136/gutjnl-2024-333999, PMID 41951359. Tier D (cross-sectional) but biomarker-based, not FFQ-based. [V] 1,461 biosamples from 487 subjects (245 Crohn’s, 242 controls) across Australia, Hong Kong and mainland China, assayed for aspartame, sucralose, saccharin and P-80 in stool, urine and serum. Crohn’s patients had higher sweetener levels across all cohorts (all p<0.0001). Native P-80 was undetectable; it undergoes predominantly hydrolytic degradation in Crohn’s and oxidoreductive degradation in controls. CD-associated P-80 metabolites correlated with urinary sweeteners, and in vitro those metabolites increased gut permeability, enabling sweetener translocation across the epithelium. A model using sweeteners plus P-80 metabolites separated active from inactive Crohn’s with AUC 0.86 in discovery and average 0.94 in two validation cohorts.

ADDapt reached full trial size. Emulsifier restriction in active Crohn’s, n=154, multicentre, randomised, double-blind, placebo-controlled re-supplementation design, King’s College London (NCT04046913). CDAI response

=70 in 49.4% vs 30.7%. Presented at ECCO February 2025; the abstracts are published (J Crohns Colitis 2025;19(Suppl 1):i262 and i1460) but I found no full peer-reviewed primary paper as of 2026-09-06 [V for the abstracts, R for publication status]. A companion dietary analysis found emulsifier restriction did not impair nutrient intake and improved food-related quality of life. Tier A, but currently abstract-only.

Emulsifier-free diets are feasible but confounded with UPF reduction. Wellens J, Luppens M, … Vermeire S. “Feasibility assessment of an emulsifier-free diet in healthy subjects: a sub-analysis of the FOAM trial.” Clin Nutr ESPEN 2026; 10.1016/j.clnesp.2026.103321, PMID 42105859, NCT06552156 [V]. n=60 healthy volunteers, 6 weeks. Emulsifier intake fell 96.6% and adherence was 88.1%, but UPF consumption fell 37.1% as an inadvertent side effect.

That confound binds subtractive designs: a trial that simply removes emulsifiers from a free-living diet also removes ultra-processed food. It does not bind ADDapt, which is a placebo-controlled re-supplementation design: every participant eats the same fixed low-emulsifier diet, and the randomisation adds back emulsifier or placebo on top of it. Processing level is held constant by construction, so the contrast is the emulsifier.

A new mechanistic challenge to lecithin, which the rubric treats as benign. Yazici D, Hayashi Y, Pat Y, … Allergy 2026; 10.1111/all.70473, PMID 42568319. Tier E. [V] Gut-on-a-chip, human colon organoids-on-a-chip and murine models. Soy lecithin and DATEM both produced dose-dependent cytotoxicity and barrier disruption at daily-exposure doses (reduced TEER, ZO-1 disorganisation), activated TNF, NF-kB and unfolded-protein-response pathways, raised serum cytokines, and soy lecithin induced IgE in mice while DATEM enhanced IL-4-induced IgE production in human PBMCs. In vitro and murine only.

1.6.5 Carrageenan: the rubric’s summary of Wagner 2024 is incomplete

Wagner R, Buettner J, Heni M, … Birkenfeld AL. “Carrageenan and insulin resistance in humans: a randomised double-blind cross-over trial.” BMC Medicine 2024; 10.1186/s12916-024-03771-8, PMID 39593091. Tier B. [V, abstract read in full]

n=20 males, 250 mg carrageenan twice daily for 2 weeks vs placebo, double-blind crossover, primary outcome insulin sensitivity by OGTT and hyperinsulinaemic-euglycaemic clamp.

The primary outcome was null. “Overall insulin sensitivity did not show significant differences between the treatments.”

[Corrected on review: an earlier version said “every positive finding is a BMI-interaction subgroup result”. Withdrawn.] The split is:

• Insulin sensitivity results are the subgroup ones. OGTT-ISI p=0.04, fasting IR p=0.01, hepatic ISI p=0.04, all BMI-by-treatment interactions present in overweight participants only, along with raised CRP and IL-6 and a trend toward brain inflammation.
• The permeability result is whole-sample. The lactulose-mannitol ratio was elevated in the whole sample (N=19, P=0.03), not in a subgroup. Zonulin was P=0.05 and post hoc. Ex vivo NK-cell activation was also reported. https://pmc.ncbi.nlm.nih.gov/articles/PMC11590543/

Two 2026 items add caution: a Clinical and Experimental Allergy letter, “Reframing Mechanistic Inference in Carrageenan-Induced Epithelial Injury” (10.1111/cea.70330, PMID 42065219) [V title only, no abstract], and a Biochem Biophys Rep 2026 study (10.1016/j.bbrep.2026.102645, PMID 42292694) showing that under simulated gastric conditions (pH 1.2, 37 C) iota-carrageenan hydrolyses only moderately and fragments stay above the 20 kDa poligeenan threshold; sub-20 kDa fragments appear only with combined acid and >100 C heat [V]. That supports the rubric’s existing point that degraded carrageenan is a different molecule from the food-grade additive.

1.6.6 Phosphates: exposure is broader than assumed, causality no better supported

Dunford EK, Calvo MS. “Phosphate-based additives in processed foods: is excess exposure a cause for concern? A cross-sectional examination of the United States packaged food supply.” Am J Clin Nutr 2025; 10.1016/j.ajcnut.2025.01.009, PMID 40180501. Tier D. [V]

Ingredient lists for 39,937 products from the top 25 US manufacturers, representing >$120 bn in 2020 purchases. Phosphate additives present in 56% of products. Top entries: lecithin 32%, sodium phosphate 13%, calcium phosphate 11%, modified starches 10%, sodium acid pyrophosphate 6%. 21% of foods contain more than one.

Two things to note. First, the paper repeats the “more rapid/efficient absorption” claim that the rubric already flags as an in-vitro digestibility ceiling rather than an absorption measurement, and Calvo is the author of the 2013 review the rubric attributes it to, so this is the same claim recycled, not independent corroboration. Second, its own top category is lecithin, which is an organic phospholipid and precisely the compound the rubric says must never enter an inorganic-phosphate detector. A 56% prevalence figure that is mostly lecithin is not a measure of inorganic phosphate exposure.

---

1.7 Personalized postprandial response: does person swamp food?

What changed. The variance decomposition is now public and quotable, and the answer is no. Meanwhile the measurement layer underneath consumer personalization has been shown to be far noisier than the modelling layer built on top of it.

1.7.1 The actual variance numbers

Berry SE, Valdes AM, … Spector TD. “Human postprandial responses to food and potential for precision nutrition.” Nature Medicine 2020;26:964-973, 10.1038/s41591-020-0934-0, PMID 32528151. n=1,002 UK (TwinsUK plus unrelated), 1 clinical day plus 13 home days, 8 meals, plus a US validation set of n=100. Tier C. [V, full text]

Two-way ANOVA on glycaemic iAUC (n=483). [Not verified on review: these three percentages come from the paper’s supplementary ANOVA, which was not re-fetched. The abstract-level figures below are verified.]

Source	% of variance (95% CI)
Meal macronutrient composition	16.73 (15.37-18.92)
Individual glucose scaling (global high/low responder)	18.74 (17.96-19.46)
Person-by-meal interaction (the personalization signal)	7.63 (6.11-8.96)

Regression-model variance explained: for lipaemia, person-specific factors including microbiome 7.1% vs meal macronutrients 3.6%; for glycaemia, person-specific 6.0% vs meal macronutrients 15.4%. Genetics explained 9.5% of glucose and 0.8% of triglyceride variance. Test-retest ICCs: glucose iAUC 0.74, triglyceride 0.46, C-peptide 0.62. Twin heritability: glucose iAUC 48%, triglyceride 6-hour rise 0%.

The answer to the framing question. Inter-individual variation does not swamp food-level differences. The person’s global offset is comparable to the meal effect, but the person-by-meal interaction, which is the only thing a personalized food score can exploit beyond “you are a high responder” plus “eat fewer refined carbohydrates,” is about 8% of iAUC variance.

Funding: supported by ZOE Global. Spector, Berry, Valdes, Asnicar, Franks, Delahanty and Segata are ZOE consultants; ten co-authors are or were ZOE employees.

1.7.2 The measurement substrate is worse than the model

Study	Design	Finding
Hengist A, Ong JA, McNeel K, Guo J, Hall KD. AJCN 2025;121:74-82, 10.1016/j.ajcnut.2024.10.007 [V]	n=30 non-diabetic inpatients, duplicate meals ~1 week apart, 1,189 responses	ICC 0.28 (Libre Pro), 0.17 (Dexcom G4). Variability to duplicate meals was similar to variability across different meals
Howard R, Guo J, Hall KD. AJCN 2020;112:1114-1119 [V]	n=16, 28 inpatient days, two CGMs concurrently, 27,489 paired readings	Bias 12.9 mg/dL, r=0.57, meal-ranking Kendall tau 0.43
Merino J, … Berry SE. AJCN 2022 (ZOE PREDICT secondary) [V]	n=394	Two identical Libre sensors r=0.97; Libre vs Dexcom G6 r=0.56-0.61. Probability a meal flips between top and bottom quintile: 5% intra-brand, 19% inter-brand. Intra-individual CV to repeat identical meals 29.5%
Hutchins KM, … Gonzalez JT. AJCN 2025;121:1025-1034 [V]	n=15 healthy, crossover	Libre 2 read ~16 mg/dL higher than capillary; time above 7.8 mmol/L overestimated 3.8-fold; the same smoothie classified GI 69 by CGM vs 53 by capillary, i.e. medium vs low
Spartano NL et al. JCEM 2025;110:1128 [V]	Framingham, n=1,175 (560 normoglycaemic), blinded Dexcom G6 Pro	Normoglycaemic adults already spend ~12.3% of the day (~3 h) above 140 mg/dL

An ICC of 0.17-0.28 to the same meal means most of what a consumer CGM app labels “your response to this food” is within-person noise. Regulatory context: FDA cleared Dexcom Stelo on 2024-03-05 as the first over-the-counter CGM explicitly for people without diabetes, and Abbott Lingo on 2024-06-10 as a wellness product with no diabetes indication [V]. Tier G.

The counterweight, and its own counterweight. Wu Y, … Snyder MP. Nature Medicine 2025, 10.1038/s41591-025-03719-2 [V]: n=55, seven carbohydrate meals in replicate, and the responder classes map onto measured insulin resistance. Real, but it depends on replicate supervised meals, not single-sensor consumer data. Against it: Della Corte KA, Brand-Miller J, Wolever TM, Della Corte D, AJCN 2026, 10.1016/j.ajcnut.2026.101363, PMID 42177949 [V]: 382 healthy adults, 1,022 reference and 1,116 food tests. Individual responses are predicted by scaling a person’s own reference curve by the food’s population mean glycaemic index, with prediction error (RMSD 0.78 mmol/L) below that person’s own test-retest noise (1.02). Their conclusion is that most apparent personal food effects are day-to-day glucose-tolerance drift.

1.7.3 What the personalization RCTs actually show

Trial	n / duration	Comparator	Result
Food4Me, Celis-Morales et al., IJE 2017;46:578-588 [V]	1,607 randomised, 6 mo	Conventional guideline advice	Red meat -5.48 g/d, salt -0.65 g/d, SFA -1.14% E, HEI +1.27. Adding phenotype or genotype added nothing over diet-only personalization. No clinical endpoint benefit
Ben-Yacov et al., Diabetes Care 2021;44:1980-1991 [V]	225 prediabetes, 6 mo	Active: Mediterranean diet	HbA1c -0.16% vs -0.08%. Statistically significant, clinically trivial. Funded by DayTwo and Janssen; Segal and Elinav developed the DayTwo technology
Popp CJ et al., JAMA Netw Open 2022;5:e2233760, NCT03336411 [V]	204, 6 mo	Active: standardised low-fat diet	Null. Weight -3.26% personalized vs -4.31% standard; between-group 1.05% (95% CI -0.40 to 2.50, p=0.16), numerically favouring the non-personalized arm
ZOE METHOD, Bermingham KM, … Berry SE, Nature Medicine 2024;30:1888-1897, NCT05273268 [V]	347 randomised, 18 weeks	Leaflet, video and email check-ins	Co-primary triglycerides -0.13 mmol/L (p=0.016); co-primary LDL-C null. Weight -2.46 kg, waist -2.35 cm, HbA1c -0.05%. Null for BP, insulin, glucose, C-peptide, apoA1, apoB, postprandial TG and alpha diversity. Per-protocol 225/347

The METHOD caveats are in the paper’s own text [V]: “The interventions were not matched for contact or intensity.” The 8- and 12-month follow-ups were intervention-arm only, so there is no randomised durability estimate. Funding: “funded by ZOE Ltd; the study funder contributed, as part of the scientific advisory board, to study design, data collection and analysis, and the writing of the manuscript.” Spector, Wolf and Hadjigeorgiou are ZOE co-founders; 11 authors are or were ZOE employees; 16 authors hold ZOE share options.

The pattern across the field, verified in a systematic review. Robertson et al., Nutrients 2024;16:1479, 10.3390/nu16101479 [V], 7 RCTs, n=873: trials with generic-information comparators showed larger benefits than trials whose comparator involved dietitian guidance. Martinsen T, Brennan L, Food Funct 2026;17(2):646-658, 10.1039/d5fo02969d, PMID 41493072 [V], 24 RCTs 2000-2025: diet quality, HbA1c, triglycerides and insulin sensitivity improve, but few trials show between-group weight-loss differences and most show no blood-pressure difference.

1.7.4 Personalization is orthogonal to processing, and its own authors say so

The METHOD paper concedes, verbatim [V]: “we also acknowledge that the composition of foods is more nuanced than their nutrient composition, such that the matrix and processing level of foods can have major effects on health.” The ZOE 2022 algorithm scores macronutrients plus food metadata, not processing. Worse for interpretation, the personalized arm also received generalized advice in weeks 2-6 to diversify plants, increase fibre, replace refined carbohydrates with wholegrains and eat fermented foods. METHOD therefore cannot separate the personalized score from a conventional whole-food message delivered through an engagement app.

Nothing in the personalization literature contradicts processing-reduction advice, and nothing shows personalization substitutes for it.

---

1.8 Microbiome: what actually moved from association to intervention

What changed. It mostly moved the other way. The two most-cited intervention results in this space each acquired a serious problem, and the largest trials are null.

1.8.1 Wastyk 2021 missed its own primary endpoint

Wastyk HC, … Gardner CD, Sonnenburg JL. Cell 2021;184:4137-4153.e14, 10.1016/j.cell.2021.06.019, PMID 34256014, NCT03275662 [V]. 37 randomised (19 fibre, 18 fermented), 17 weeks.

The registered primary outcome, change in Cytokine Response Score at week 10, was unchanged in both arms. Everything that made the paper famous (fermented arm: rising alpha diversity and 19 serum inflammatory proteins down including IL-6, IL-10 and IL-12b; fibre arm: flat diversity, CAZymes up, three baseline-diversity-stratified immune trajectories) is secondary or exploratory. No results were posted to ClinicalTrials.gov. Tier B, with a failed primary.

1.8.2 The nearest replication points the other way

van den Belt M, … Kort R. medRxiv 2025.08.05.25332853, posted 2025-08-07, preprint, not peer reviewed, NCT05900609 (n=146 actual) [V]. n=147 healthy adults, 8 weeks, high-fibre (chicory root) vs high-fermented-food vs control, 21-week follow-up.

• Alpha diversity rose in the fermented arm and in the control arm, with no fermented-versus-control difference.
• The high-fibre arm showed a significant decrease in diversity.
• Fermented foods increased immune markers (CD5, CD6, CD8A, IL-18R1, SIRT2; Q<0.05), the opposite direction to Wastyk.

Registry lists industry collaborators (WholeFiber, Cidrani, Keep Food Simple).

Other fermented-food intervention data [V]: a sauerkraut crossover (DRKS00027007, n=87, 100 g/day fresh vs pasteurised, 4 weeks each; Microbiome 2025 PMID 39940045 and Eur J Clin Nutr 2026 PMID 42321414) found a small systolic BP drop with both fresh and pasteurised sauerkraut, so not live-microbe dependent, no gut-barrier change, and “not linked to an appreciable systemic health benefit.” A kombucha trial (Sci Rep 2024, PMID 39738315, n=16 vs 8, 4 weeks) found no inflammation change, with fasting insulin and HOMA-IR rising within the kombucha arm. FeFiFo-MOMS (Stanford, NCT05123612, n=135, 4 arms) has a design paper only and an estimated completion of 2029.

As of 2026-09 no larger fermented-food RCT with a positive pre-registered inflammatory primary endpoint exists.

1.8.3 Fibre intervention trials: the largest is null and one shows harm

• Largest, and null. Nature Communications 2025, PMID 41390484, 10.1038/s41467-025-66498-x [V]. n=802 prediabetic adults, 6 months, fibre (409) vs usual care (393), open-label. Primary (percent change in HbA1c) and all secondary outcomes null. Its “personalization” is post-hoc clustering plus a LightGBM score and is hypothesis-generating only. Tier A, null.
• Positive: Cell Rep Med 2025, PMID 40669445, NCT04714944 [V]. 12 weeks, placebo-controlled intrinsic chicory plant-cell fibre: whole-body insulin sensitivity improved (p=0.032), triglycerides down (p=0.049).
• Harm signal: Gut Microbes 2026, PMID 41459804 [V]. 12 weeks, potato fibre / sugar beet pectin (n=19) vs maltodextrin (n=21) on a high-protein background: whole-body insulin sensitivity decreased (p=0.034), colonic permeability up (p=0.046), plasma IL-6 up (p=0.025).
• Fermentability may not be the mechanism: medRxiv 2025.11.20.25340625, PMID 41332831, NCT02322112, preprint [V]. 6 weeks, acacia gum vs RS4 vs cellulose: inflammation, barrier and satiety improved in all arms including the non-fermentable cellulose control.

Note that none of this touches the rubric’s fibre position, which rests on Reynolds 2019 (185 prospective studies plus 58 RCTs, hard outcomes). It touches the microbiome-mediated account of why fibre works.

1.8.4 FMT and defined consortia for metabolic endpoints: null across the board

Target	Trial	Primary result
MASLD	Gut Microbes 2025, PMID 40755230, NCT04465032, n=20, 3 FMTs [V]	MRI-PDFF null (p=0.50). Unrestricted Vedanta Biosciences grant, Vedanta-employed authors
T2D	Nutrients 2024, PMID 39458486, n=21, FMT vs probiotic vs placebo [V]	No improvement; HbA1c rose in the FMT arm (+0.25%, p=0.041)
Metabolic syndrome	J Obes Metab Syndr 2025, TCTR20240805004, n=8 vs 10 [V]	HOMA-IR better at 6 weeks (adj diff -1.63, p=0.001), lost by 12 weeks (p=0.388)
Hypertension	Microbiome 2025, PMID 40410854, NCT04406129, n=124 [V]	SBP at day 30 null (p=0.62); a transient -4.34 mmHg at 1 week did not persist
Adolescent obesity	Nat Commun 2025, PMID 40877311, 4-year follow-up, 55/87 retained [V]	BMI no difference (p=0.095); unblinded secondary outcomes favoured FMT
Akkermansia muciniphila	Gut Microbes 2026, PMID 42343233, n=142, 4 months [V]	Matsuda index null
AKK-WST01	Cell Metab 2025, PMID 39879980, NCT04797442, n=58, 12 weeks [V]	No between-group difference; benefit only in a low-baseline-Akkermansia subgroup

1.8.5 What remains association-only

Fermented food lowering systemic inflammation in humans; microbiome alpha diversity as a causal target rather than a marker; fibre acting on HbA1c through a microbiome mechanism (the n=802 trial is null); FMT for obesity, T2D, MASLD or blood pressure; gut-brain and autoimmune diet-microbiome claims; and “your microbiome tells you which fibre to eat,” where every signal to date is post hoc or ML-derived and none has been prospectively validated in a randomised allocate-by-microbiome design.

The one genuinely large new descriptive resource is Asnicar F, … Segata N, “Gut micro-organisms associated with health, nutrition and dietary interventions,” Nature 2026, 10.1038/s41586-025-09854-7, PMID 41372407 [V], with >34,000 participants and an intervention component of n=746 from two dietary trials. It is a mapping exercise, tier C.

---

1.9 GLP-1 drugs and food choice: a pharmacological probe of the food environment

What changed. There is now two-year randomised craving data, a category-level food-preference substudy, a 150,000-household purchase panel, and three randomised alcohol trials. Together these make GLP-1 agonists an unplanned natural experiment on what UPF-heavy eating actually is. The result is less flattering to the “UPF hijacks satiety” story than the popular framing suggests.

1.9.1 Intake falls hard, and it is not a gastric-emptying effect

Study	Design / n	Result
Blundell J et al., Diabetes Obes Metab 2017;19:1242-1251, 10.1111/dom.12932, PMID 28266779 [V]	RCT, double-blind, placebo-controlled crossover, n=30 obesity, 12 wk, semaglutide 1.0 mg	Total daily ad libitum intake -24% (-3,036 kJ, p<0.0001); lunch -1,255 kJ. “Lower relative preference for fatty, energy-dense foods.” Weight -5.0 kg. RMR adjusted for lean mass unchanged. Four authors are Novo Nordisk employees
Gibbons C et al., Diabetes Obes Metab 2021;23:581-588, 10.1111/dom.14255, PMID 33184979, NCT02773381 [V]	Crossover, n=15 T2D (13 evaluable), 12 wk, oral semaglutide to 14 mg	Total ad libitum intake -38.9% (ETD -5,096 kJ, 95% CI -7,000 to -3,192). Appetite differences reached significance only after a fat-rich breakfast, not a standard one
Friedrichsen M et al., Diabetes Obes Metab 2021;23:754-762, 10.1111/dom.14280, PMID 33269530 [V]	Double-blind parallel, n=72, 20 wk, semaglutide 2.4 mg	Intake -35% (ETD -940 kJ, p<0.0001); weight -9.9% vs -0.4%. No evidence of delayed gastric emptying at week 20 (paracetamol AUC0-5h +8%, p=0.12 after body-weight correction)
Knop FK et al., Diabetes Obes Metab 2024, 10.1111/dom.15802, PMID 39082206 [V]	RCT, n=61, 20 wk, oral semaglutide 50 mg	Intake -39.2 percentage points; weight -9.8% vs -1.5%. No significant gastric-emptying difference
Tronieri/DeRouen, J Nutr 2026, PMID 42323166, NCT05548647 [V]	RCT, n=120, 3:2, semaglutide 2.4 mg vs placebo, 60 weeks, all with intensive behavioural therapy	Ad libitum lunch intake lower at wk 20, 40, 60 (-291.9, -240.2, -269.5 kcal). Subjective appetite differences vanished after wk 20; Power of Food reward responsiveness lower at wk 20 and 40 but not 60

The gastric-emptying nulls matter: the effect is central, not mechanical. The Penn 60-week trial matters more: objective intake suppression persists while the felt experience fades, so “food noise lifting” and the caloric effect are different variables.

1.9.2 The best category-level data do not detect UPF-selectivity

Kennedy SF, Knights A, Ravussin E, … Martin CK. “Impact of tirzepatide treatment on participant-reported food craving and food preference: secondary analyses of a phase 1 randomised controlled trial.” Diabetes Obes Metab 2025;27:6784-6789, 10.1111/dom.70063, PMID 40874370, NCT04081337. Tier B. [V] n=55 randomised (27 tirzepatide 15 mg, 28 placebo), 18 weeks, all on a low-calorie diet. Weight -16.7 kg vs -8.3 kg.

Week-18 estimated treatment differences:

Measure	ETD	p
FPQ high fat	-1.07	0.0072
FPQ low fat	-0.75	0.0234
FPQ high simple sugar	-1.06	0.0075
FPQ high-fat/high-sugar	-1.34	0.0037
FPQ low-carbohydrate / high-protein	-0.84	0.0262
FPQ fat preference ratio	-8.28	0.0594 (NS)
FCI overall	-0.35	0.0098
FCI sweets	-0.49	0.0041
FCI carbohydrates/starches	-0.43	0.0068
FCI fast-food fats	-0.39	0.0384
FCI high fat	-0.14	0.31 (NS)
FCI fruits and vegetables	-0.29	0.064 (NS)

Tirzepatide reduced 10 of 12 preference scores and 4 of 6 craving scores. Liking fell for high-fat, low-fat, high-sugar, high-complex-carbohydrate and low-carb/high-protein foods alike.

[Corrected on review: this was previously summarised as “the dampening is global, not UPF-selective”. Withdrawn. The correct statement is that selectivity was not detected.] The two contrasts that would show selectivity are underpowered, not null: the fat-preference ratio is ETD -8.28 (-16.91, 0.34), P=0.0594, and fruit and vegetables are -0.29 (-0.61, 0.02), P=0.064. Both intervals are wide and both point in the selectivity-consistent direction. Two further limits: all participants were on a low-calorie diet, so preference change is confounded with prescribed restriction, and the FPQ does not index ultra-processing at all, so this instrument cannot answer a NOVA question even in principle (https://pmc.ncbi.nlm.nih.gov/articles/PMC12515769/). Effects attenuated between weeks 8 and 18. Lilly sponsored; 7 authors are employees or shareholders.

The two-year randomised craving data are asymmetric in a way no account predicts. STEP 5 CoEQ substudy, Wharton S et al., Obesity 2023;31:703-715, 10.1002/oby.23673, PMID 36655300 [V]. 104 weeks, semaglutide n=88 vs placebo n=86, weight -14.8% vs -2.4%. P values not multiplicity-controlled.

• Craving Control and Craving for Savoury: significant at weeks 20, 52 and 104 (p<0.01).
• Positive Mood and Craving for Sweet: significant at weeks 20 and 52 only.
• Hunger and fullness differed only at week 20.
• At 104 weeks: reduced desire for salty and spicy food, cravings for dairy and starchy foods, difficulty resisting cravings.

Instrument warning. Only Blundell and Gibbons used the Leeds Food Preference Questionnaire; the tirzepatide work used FPQ plus FCI; STEP and Friedrichsen used CoEQ-19. These are not interchangeable, so cross-drug comparisons are weak. The systematic review of RCTs on appetite, taste and food preference (Aldawsari 2023, Diabetes Metab Syndr Obes 16:575-595, PMID 36890965) [V] found only 12 RCTs totalling n=445. That is the entire randomised food-preference evidence base.

1.9.3 Rodent work isolates the receptor and the macronutrient

Samms RJ et al., Diabetes Obes Metab 2023, 10.1111/dom.14843, PMID 36054312. Tier E. [V] In mice and rats, tirzepatide suppressed total intake while shifting choice toward chow. Two clean dissections: GIPR agonism alone did not alter food choice, and the shift was completely absent in GLP-1R knockout mice; and in a macronutrient-choice paradigm the suppression was specific to lipid and spared sucrose solution. The macronutrient specificity that is clean in rodents is the thing that does not replicate cleanly in the human FPQ data.

1.9.4 The purchase panel, and the single most informative result in this section

Hristakeva S, Liaukonyte J, Feler L. “The No-Hunger Games: How GLP-1 Medication Adoption Is Changing Consumer Food Demand.” Journal of Marketing Research, 10.1177/00222437251412834, online Dec 2025, issue Apr 2026 (SSRN preprint 10.2139/ssrn.5073929). Tier D. [V] Numerator panel, ~150,000 US households, survey-linked adoption plus transaction records, matched-control design.

• Grocery spend -5.3% within 6 months of adoption; -8.2% among higher-income households.
• Savoury snacks -10.1%; similarly large declines in sweets, baked goods and cookies. Limited-service restaurants, fast food and coffee -8.0%.
• Most categories declined, including bread, meat and eggs. Only a handful rose: yogurt largest, then fresh fruit, nutrition bars, meat snacks.
• Effects persist through year 1 with attenuation after 6 months.
• Discontinuers revert toward pre-adoption spend and shift to baskets slightly less healthy than their own baseline. [Qualitative wording verified against the SSRN preprint, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5073929. Effect size UNSOURCEABLE: no magnitude for the discontinuer shift could be extracted from the accessible versions.]

That last finding is the most theoretically loaded result in this document, and it is also the softest. It is consistent with pharmacological suppression of a persistent underlying drive and less consistent with a learned or durably corrected preference. Caveats: household spend rather than individual intake; adoption is not randomly assigned; and discontinuers are self-selected, so the people who stop are plausibly the ones for whom the drug worked least well or who were least engaged, which produces reversion without any statement about the underlying drive.

1.9.5 Diet quality does not improve, and that is the awkward part

• The registration trials essentially did not measure what people ate. Babazadeh A et al., scoping review, Advances in Nutrition 2025;16:100491, 10.1016/j.advnut.2025.100491, PMID 40812508 [V]. 129 RCTs of liraglutide, semaglutide and tirzepatide. Only 36 recorded diet quality or food intake and only 10 reported those outcomes; half of those used a single-timepoint ad libitum meal.
• CRAVE, Babazadeh, Obesity Pillars 2026;19:100292, 10.1016/j.obpill.2026.100292, PMID 42440974, NCT06467604 [V]. Prospective observational, n=28 analytic, 24 weeks, real-world, no structured nutrition support. Weight, adiposity and BIA skeletal muscle all fell (p<0.001), with roughly one quarter of weight loss from estimated skeletal muscle. HEI-2020 diet quality did not improve. Food cravings were unchanged overall. Significant declines across multiple micronutrients. Small and exploratory, but it is the only prospective real-world diet-quality study located, and it does not replicate the trial craving findings.
• Baseline context: J Nutr 2026;156:101753, PMID 42520970 [V], NHANES 2005-2020, n=16,143: 43.5% of US adults meet GLP-1 eligibility criteria, and diet quality and micronutrient adequacy are already poor in that group before any drug.
• “Food noise” was only formally operationalised in 2025 (Dhurandhar et al., Nutrition & Diabetes 2025;15:30, 10.1038/s41387-025-00382-x, PMID 40628707) [V], so there is no validated pre-2024 instrument and every before/after claim rests on retrospective recall. The INFORM survey (Adv Ther 2026, PMID 42217114, n=550) reports median Food Noise Questionnaire 13 before vs 6 after, but it is retrospective recall of the pre-treatment state, Novo Nordisk employees are authors, and Numerator was paid by Novo Nordisk. A cross-sectional chemosensory study (Rutigliani & Mattes, Physiol Behav 2026;316:115486, PMID 42641845, n=79) [V] found 6-month users lower on food noise, cue responsivity, craving and intake, with chemosensory change limited to elevated sweet-taste detection thresholds (p=0.001) and protein intake not differing across groups.

1.9.6 Cross-substance evidence: it is a general consummatory-reward effect

Trial	Design / n	Result
Hendershot CS et al., JAMA Psychiatry 2025;82:395-405, 10.1001/jamapsychiatry.2024.4789, PMID 39937469, NCT05520775 [V]	Phase 2, n=48 non-treatment-seeking AUD, 9 wk, low-dose semaglutide	Reduced lab self-administration (grams beta -0.48, peak BrAC -0.46), drinks per drinking day (-0.41), weekly craving (-0.39). No effect on drinks per calendar day or number of drinking days. Also reduced cigarettes/day in smokers
Klausen/Fink-Jensen, The Lancet 2026;407:1687-1698, 10.1016/S0140-6736(26)00305-3, PMID 42070571, NCT05895643 [V]	n=108, 26 wk, semaglutide 2.4 mg plus CBT, AUD with comorbid obesity	Heavy drinking days -41.1 pp vs -26.4 pp; ETD -13.7 pp (95% CI -22.0 to -5.4, p=0.0015)
Schacht, Am J Psychiatry 2026;183:636-645, 10.1176/appi.ajp.20260003, PMID 42522065 [V]	Phase 2, n=50 treatment-seeking, 8 wk oral semaglutide	Missed its primary endpoint (cue-elicited craving) and drinks/day. Reduced heavy drinking days (-0.58), drinks per drinking day (-1.18), naturalistic craving (-2.20) and cannabis use days (-1.43)

The pattern is consistent across substances: an effect on intensity per occasion, inconsistent effects on frequency and on cue-elicited craving. That is the same shape as the food data. A drug that reduces alcohol, cannabis and savoury-fat craving is acting on a general consummatory-reward pathway.

1.9.7 What this does and does not license

Defensible: GLP-1 agonists show that a large share of energy-dense processed food intake is appetitive rather than habitual or purely cue-driven, and that this appetite is pharmacologically compressible. The discontinuation-reversion result shows the drive is persistent and environmentally re-supplied, not unlearned.

Not licensed by the data:

1. That UPF specifically hijacks satiety signalling. The tirzepatide preference table shows dampening across protein and low-fat foods too, and the two selectivity contrasts (fat-preference ratio P=0.0594, fruit and vegetables P=0.064) are underpowered rather than null. The honest claim is that selectivity was not detected, not that it was excluded, and the FPQ does not index ultra-processing in any case.
2. That people become disproportionately averse to UPF versus whole foods. The purchase data are more selective than the lab data, but purchases reflect budget reallocation under a fixed appetite constraint and are confounded by the intention-to-improve-health that accompanies drug initiation.
3. That residual diet quality improves. It does not (CRAVE), and a quarter of the weight lost is muscle.
4. No study has measured NOVA-classified UPF intake as an outcome on GLP-1 therapy. Every “GLP-1 reduces ultra-processed food” claim traces to category-level spending data or to non-NOVA craving instruments. This is a clean, cheap, unclaimed study.

A case series in Obesity Pillars 2026 (PMID 42518361) [V] reports anhedonia-like symptoms, emotional flatness and loss of interest in exercise in three patients on 15 mg/week tirzepatide, resolving on dose reduction. If the drug blunts a general reward pathway, the blunting is not confined to food.

---

1.10 Ultra-processed plant-based meat and dairy alternatives

What changed. The field acquired a meta-analysis of RCTs, a properly independent trial that contradicts the famous one, and cohort analyses that separate “healthful plant-based UPF” from “unhealthful.” The funding structure turns out to explain most of the disagreement.

1.10.1 The meta-analysis, and why its headline is narrower than it reads

Fernandez-Rodriguez R, Bizzozero-Peroni B, Diaz-Goni V, Garrido-Miguel M, Bertotti G, Roldan-Ruiz A, Lopez-Moreno M. Am J Clin Nutr 2025;121:274-283, 10.1016/j.ajcnut.2024.12.002, PMID 39653176, PROSPERO CRD42024556191. Tier A (pooled), small. [V] 8 publications from 7 RCTs, n=369, all <=8 weeks.

Outcome	Mean difference (95% CI)	I²	Trials
LDL cholesterol	-0.25 mmol/L (-0.42, -0.08)	65.8%	7
Total cholesterol	-0.29 mmol/L (-0.52, -0.06)	64.8%	6
Body weight	-0.72 kg (-1.02, -0.42)	0%	5
HDL, triglycerides, blood pressure, fasting glucose	null

The lipid signal is driven by mycoprotein products (LDL-C -0.37 mmol/L across the four mycoprotein trials, I²=0 for total cholesterol), not by soy or pea burgers. Generalising “plant-based meat lowers LDL” from this pool to a Beyond Burger is not supported. A second review (Del Bo’ et al., Nutrients 2024;16:2498, PMID 39125378) [V] covering 19 intervention studies found evidence sparse and heterogeneous, with lower protein bioavailability for plant-based alternatives and no effect on muscle protein synthesis.

1.10.2 SWAP-MEAT versus the independent replication

SWAP-MEAT. Crimarco A, … Gardner CD. Am J Clin Nutr 2020;112:1188-1199, 10.1093/ajcn/nqaa203, PMID 32780794, NCT03718988. Tier B. [V] Single-site randomised crossover, no washout, n=36 completers, 8 weeks per phase, >=2 servings/day Beyond Meat vs animal meat.

• TMAO 2.7 vs 4.7 umol/L, difference -2.0 (95% CI -3.6, -0.3), p=0.012, but with a significant order effect (p=0.023): present only in the 18 who received plant second (2.9 vs 6.4, p=0.007), absent in the plant-first group (2.5 vs 3.0, p=0.23).
• LDL-C -10.8 mg/dL (-17.3, -4.3), p=0.002. Weight -1.0 kg.
• Null on IGF-1, insulin, glucose, HDL-C, triglycerides, systolic and diastolic BP.
• Funded by a research gift from Beyond Meat. Mitigations were a pre-submitted statistical analysis plan and blinded third-party analysis. The comparator was 80/20 beef at 9 g saturated fat per burger versus 6 g for Beyond, which mechanically favours the plant arm on LDL-C. Sample size was set by available resources, not a power calculation.
• Secondaries: inflammation null (J Nutr Sci 2022;11:e82, PMID 36304815; 4 of 92 biomarkers significant and in the contrary direction) [V]. Urinary and kidney measures favourable (CJASN 2024;19:1417-1425, PMID 39514692) [V].

The independent contradiction. Toh DWK, Fu AS, Mehta KA, Lam NYL, Haldar S, Henry CJ. Am J Clin Nutr 2024;119:1405-1416, 10.1016/j.ajcnut.2024.04.006, PMID 38599522, NCT05446753. Tier A, and the largest PBMA trial. [V] Parallel-arm, 8 weeks, n=89 randomised / 82 analysed, Singaporean adults at elevated diabetes risk, 2.5 servings/day.

• Primary outcome LDL cholesterol: null. No lipid-lipoprotein effects at all.
• Diastolic BP lower on the plant arm (p-interaction 0.041); nocturnal DBP dip diverged (+3.2% animal vs -2.6% plant, p-interaction 0.017), though this was a 40-person subsample.
• Glycaemia favoured the meat arm: CGM time-in-range 94.1% vs 86.5%, p=0.041.
• Fibre, sodium and potassium rose on the plant arm; dietary trans fat rose on the animal arm.

Other trials, all null or trivial [V]: Weschenfelder et al., J Nutr 2025;155:4365-4372, PMID 41138758, randomised double-blind parallel, n=52 iron-deficient women, 8 weeks, 4 oz/day beef vs Beyond Meat, null on body composition, BP, glycaemia, lipids and lipoproteins; part-funded by the National Cattlemen’s Beef Association. Cloward et al., Am J Clin Nutr 2026, PMID 41519303, n=36 three-group crossover, no difference in CRP, IL-6 or TNF-alpha (p=0.64). FOOD-1 (Sci Rep 2026, PMID 41813762, n=41 but only 6 days per arm): TMAO -0.61 log units, LDL-C -6.0 mg/dL, but weight +0.6 kg and NT-proBNP +0.19 log units. SWAP-MEAT Active adult (Nutr J 2022;21:69, PMID 36384651, n=22) and the MCURC multi-site follow-up (Nutr J 2026;25:42, PMID 41715155, NCT06014307, n=36 across four sites): null on 12-minute run and strength.

The pattern is a funding pattern. The one trial showing an LDL benefit was Beyond Meat funded with a fattier comparator. The largest independent trial was null on LDL and mildly unfavourable on CGM glycaemia. The beef-industry-funded trial was null on everything. The meta-analytic lipid effect is mycoprotein-driven.

1.10.3 Micronutrient status: fortification wins, which is itself the finding

Fu AS, Osman F, Cameron-Smith D, … Toh DWK. Clinical Nutrition 2026, 10.1016/j.clnu.2026.106610, PMID 41785660. Tier A. [V] Same n=89 cohort, 8 weeks.

The plant-based arm came out ahead on intake of thiamine, folate, B6, sodium, calcium, potassium, magnesium and iron, and on plasma B12 (382.56 vs 357.76 pmol/L, p=0.004), folate (32.13 vs 23.62 nmol/L, p=0.003) and selenium (p=0.006). Marginal bone mineral density increase in the plant arm (0.06 g/cm², 95% CI 0.01-0.12, p=0.030) with no change in beta-crosslaps or P1NP, though a sex-specific signal suggested higher beta-crosslaps in females.

This is a fortification result, not a whole-food result, and it is the cleanest demonstration in this document that “ultra-processed” and “nutritionally inadequate” are separable properties.

1.10.4 Cohorts: no PBMA exposure exists, but the NOVA subgroups are informative

No cohort has measured plant-based meat or dairy alternative consumption as an exposure against hard outcomes. The cohort literature is entirely NOVA-classified plant-sourced UPF, a much broader category that includes packaged breads, sweetened beverages and snacks. Within that limit:

Study	Cohort / n	Result
Rauber F et al., Lancet Reg Health Eur 2024;43:100948, PMID 39210945 [V]	UK Biobank n=126,842, median 9 y	Per 10 percentage points of energy: plant-sourced non-UPF CVD HR 0.93 (0.91-0.95), CVD mortality 0.87 (0.80-0.94); plant-sourced UPF CVD 1.05 (1.03-1.07), CVD mortality 1.12 (1.05-1.20)
Thompson et al., Lancet Reg Health Eur 2026;67:101736, PMID 42294356 [V]	UK Biobank n=124,836, 8.3-10.5 y, 5,780 deaths	Opposite conclusion. High-UPF healthful plant-based diet index: mortality HR 0.92 (0.85-1.00) vs low-UPF hPDI 0.91 (0.84-0.98); T2D 0.89 vs 0.72; CVD 0.89 (0.82-0.96) for high-UPF hPDI. Reading: diet quality dominates processing level
Prioux et al., Lancet Reg Health Eur 2025;59:101470, PMID 41466603 [V]	NutriNet-Sante n=63,835, 9 y	Healthful-plant + unprocessed CHD HR 0.56 (0.42-0.75); unhealthful-plant + UPF 1.46 (1.11-1.93); healthful-plant + UPF null
Cordova et al., Lancet Reg Health Eur 2023;35:100771, PMID 38115963 [V]	EPIC n=266,666, 11.2 y	The “plant-based alternatives” UPF subgroup was null for cancer and cardiometabolic multimorbidity: HR 0.97 (0.91-1.02), versus animal-based UPF 1.09 (1.05-1.12)
de Fragas Hinnig, EClinicalMedicine 2026;97:104050, PMID 42433276 [V]	UK Biobank n=17,374, 5.3 y	Plant-sourced UPF HR 1.10 for >=5% BMI gain; plant non-UPF 0.89

The healthful-versus-unhealthful separation exists and is reproducible across three cohorts, and the harm consistently attaches to the unhealthful axis rather than to processing per se. That is exactly the subgroup analysis the Lancet Series authors reject as “conceptually and methodologically flawed.” EPIC-Oxford has published no analysis of meat substitutes against CVD, T2D or mortality [V, searched].

1.10.5 Plant-based milk: soy has outcome data, nothing else does, and iodine is the real problem

The only outcome-adjacent synthesis is soy-specific. Erlich MN et al., BMC Medicine 2024;22:336, 10.1186/s12916-024-03524-7, PMID 39169353. Tier A (pooled). [V] 17 RCTs, n=504, >=3 weeks, median 500 mL/day soymilk substituted for cow’s milk: non-HDL-C -0.26 mmol/L (-0.43, -0.10), LDL-C -0.19 mmol/L, CRP -0.82 mg/L, SBP -8.00 mmHg (-14.89, -1.11), DBP -4.74 mmHg (-9.17, -0.31). GRADE high for LDL-C and non-HDL-C, moderate for BP. No effect modification by added sugars. Heavy COI: Sievenpiper discloses Protein Industries Canada, the United Soybean Board, a University of Toronto “Plant Milk Fund,” Danone and Dairy Farmers of Canada; co-author Messina is at Soy Nutrition Institute Global. The very wide blood-pressure confidence intervals indicate small heterogeneous trials.

Iodine is the adequacy failure that nobody prices in [all V]:

• Dineva M et al., Br J Nutr 2021;126:28-36, PMID 32993817. UK NDNS 2014-2017, n=3,976 diet / n=2,845 urine. Exclusive milk-alternative consumers: iodine intake 94 vs 129 ug/day, median urinary iodine 79 vs 132 ug/L, i.e. below the WHO 100 ug/L sufficiency threshold while cow’s-milk consumers were sufficient.
• Nicol K et al., Eur J Nutr 2024;63:599-611, PMID 38212424. Modelling: a 58% iodine drop in children 1.5-3 y with unfortified substitution; prevalence below the LRNI rising from 20% to 48% in girls 11-18 and from 13% to 33% in women of reproductive age. At least 22.5 ug/100 mL fortification needed. (Author Bath discloses honoraria from Oatly UK and Dairy UK.)
• Nicol K et al., Br J Nutr 2026, online ahead of print, 10.1017/S0007114526106758, PMID 41772783. UK supermarket survey 2020-2024, 466 products in 2024. Iodine fortification in 48% of non-organic milk alternatives, 5% of yogurts, 4% of cheeses and 0% of seafood analogues, versus calcium 88% and B12 71%. Even fortified products deliver only 75-83% of cow’s-milk iodine.

1.10.6 What is still nutrient-comparison only

• All plant-based dairy other than soymilk (oat, almond, pea; every yogurt and cheese analogue): zero RCTs with cardiometabolic endpoints, zero cohorts. Composition surveys and dietary modelling only.
• All hard endpoints (MI, stroke, incident T2D, mortality) for both PBMA and plant dairy: zero. Every trial is <=8 weeks, maximum n=89, surrogates only.
• Blood pressure: no trial is powered for it.
• Long-term safety: animal work only (e.g. J Agric Food Chem 2025;73:22698-22713, PMID 40671669, 21 Wistar rats) [V].
• Pipeline is thin [V, CT.gov API]: no large PBMA cardiometabolic RCT is registered. NCT07766928 (Saskatchewan, n=10, not yet recruiting) and NCT06874400 (Toronto, Anderson, n=160 randomised crossover, dairy vs plant-based alternatives, primary outcome blood glucose, primary completion April 2027) are the whole near-term pipeline.

---

2. Where this overturns or sharpens the rubric

Stated as “rubric says X, new evidence says Y.” Ordered by how much they should change the rubric.

Verdicts after adversarial review (2026-09-06). The table below is the adjudicated status of each conflict. Where a verdict changed, the subsection below carries the correction inline.

#	Subject	Verdict on review
2.1	Decomposition and eating rate	Does not overturn. Add the texture-as-candidate note: the flat g/min in NCT05290064 is an imprecise contrast measured while energy density doubled, and Forde 2026 is confounded with fat, sugar and protein
2.2	Eucaloric trials moving lipids and glycaemia	Does not overturn. Preston gained 1.3-1.4 kg on UPF in both arms and Capra’s arms differ at baseline. Watch item; cite the three Cell Metab letters (PMIDs 41500195, 41500197, 41500199)
2.3	Protein fortification does not fix the matrix	Sharpens. Valid as written
2.4	The subgroup dispute	Sharpens. Keep the position, keep the BMJ rebuttal caveat
2.5	Carrageenan	Partly wrong. State the primary null; the P=0.03 permeability result is a whole-sample secondary, not a subgroup finding; weight unchanged
2.6	Titanium dioxide	Not supported as written (wording only). Human intervention exists, signal is transcriptome-only, inflammation markers null, tier and weight unchanged
2.7	Erythritol	Does not follow. The new evidence bears on the cohort signal; the weight rests on pharmacodynamics. Hold 0.5 until NCT05967741
2.8	Inorganic phosphates	Sharpens. Absence of new supporting evidence is not evidence against the weight
2.9	Nitrite and “uncured”	Does not overturn. Nitrite remains unresolved, not exonerated
2.10	Fibre	Sharpens
2.11	NOVA instruments	Sharpens, with a caveat: PFI is the Forde group’s own instrument, which creates a circularity risk when it is used to adjudicate T7
2.12	Tier 3 watch items	Agree, watch. No weight change
2.13	Emulsifier confound	Wrong as stated. ADDapt’s placebo-controlled re-supplementation design escapes it
2.14	Personalization ceiling	Sharpens. The 7.63% figure is not verified; the abstract-level figures are

2.1 The decomposition is incomplete, and texture is a candidate for the missing term

Rubric says (headline 1): the ~950 kcal/day UPF overeating effect decomposes into energy density (+662), hyperpalatability (+158) and processing itself (+128, p=0.065, not significant). “Processing is real as a warning sign, not as a cause.”

New evidence says: every number is exactly right and now verified from the registry results section [V], and the decomposition is incomplete, though less dramatically than an earlier version of this section claimed.

1. In that same trial, measured eating rate was similar across all four arms (39.21, 39.02, 39.04, 38.43 g/min, every pairwise p=0.99) and palatability VAS was similar (65.24, 62.41, 66.67, 64.18, all p>=0.6). [Corrected on review: this was previously reported as the mediators “not moving at all”. Withdrawn.] The eating-rate contrasts are 0.19, -0.02, 0.61 and 0.78 g/min with intervals of about ±2.8 g/min, and the palatability contrasts span roughly ±7 VAS points, so these are imprecise contrasts, not demonstrated nulls. More importantly, g/min was measured while energy density roughly doubled, so kcal/min roughly doubled at equal g/min. A flat mass eating rate is what the energy-density account predicts.
2. Forde 2026 (AJCN, n=41, 14-day crossover, NCT06113146) produced 369 kcal/day between two ultra-processed diets whose intended contrast was texture-derived eating rate. But the arms also differed in fat (22 vs 33 EN% as served, 9 points as consumed, F=1288, P<0.0001), sugar, protein and water intake, and the “not explained by macronutrients” claim is a post hoc covariate model in which fat EN% (P=0.082) is collinear with arm [see 1.2.2].

[Withdrawn on review: the earlier headline that “the correct partition is energy density + oral-processing resistance, with NOVA category contributing little once both are set”.] Nothing in these two trials establishes that. The defensible statement is: energy density remains the dominant measured term; texture-driven eating rate is a candidate lever within ultra-processed food (clean evidence: Lasschuijt 2023, n=18; larger but confounded: Forde 2026); and “NOVA contributes little once both are set” is untested. The rubric’s conclusion (“processing is a warning sign, not a cause”) survives unchanged.

Also correct these provenance details: NCT05290064 has actual enrolment 38 (n=36 analysed) and actual completion 2025-08-15, not 2026. It is COMPLETED with registry results posted and remains not peer-reviewed as of 2026-09-06.

2.2 Two trials are read as breaking the pure-energy-density account, and neither does

Rubric says: “Match energy density and the effect on intake collapses from SMD 0.71 to 0.02.”

New evidence says: true for intake, and Rego 2026 (Obesity, n=27, eucaloric, matched for energy density) independently confirms a null on ad libitum intake in the full sample. But intake is not the only outcome.

Preston 2025 (Cell Metabolism, NCT05368194, n=43 men, 3 weeks per diet, 2 x 2 crossing processing with caloric load) reported increased body weight and increased LDL:HDL ratio on UPF, decreased GDF15 and FSH, and a downward trend in sperm motility. Capra 2026 (J Nutr, 18 participants total in a parallel pilot, 6 weeks) found glucose AUC and MAGE trending worse on high-UPF (p=0.054, p=0.055).

[Withdrawn on review: “two 2025-2026 eucaloric trials say lipids and glycaemia move when calories do not”.] Neither trial supports that sentence.

• Preston: weight rose 1.3 to 1.4 kg on UPF in both arms, so calories or absorbed energy did move. Its registered primary outcome is sperm DNA methylation, not lipids. The LDL:HDL rise is in the adequate arm only, the FSH fall in the excess arm only, and motility was not significant after correction. Three Cell Metab letters (Ludwig PMID 41500195; Matthiessen, Rosane, Mosig, Ahrné, Magkos, Bügel PMID 41500197; reply Barrès, Simpson, Nóbrega, Preston PMID 41500199) make overlapping objections and were missing from the earlier version of this document.
• Capra: the arms started 1,918 units apart on the primary glucose AUC (13,431 high-UPF vs 15,349 non-UPF). The arm that “improved” is the arm that started worse, with about 9 people per arm. Regression to the mean is a plausible complete explanation.

Verdict: does not overturn the rubric’s model. Both remain watch items: the question of whether processing acts on lipids or glycaemia at fixed energy is live and unanswered, and the trial that answers it (NCT06538831, 2 x 2 factorial, n=120, fasting LDL-C and HOMA-IR, 2027) is in section 4.2.

2.3 Protein-fortifying an ultra-processed matrix does not fix it

Rubric says: protein is the second-highest-weighted nutrient and carries 24 of 110 points for the adult.

New evidence says: Hagele 2025 (Nature Metabolism, n=21, whole-room calorimeter) fed two ultra-processed diets at 30% vs 13% protein, ad libitum, matched palatability. The high-protein arm ate 196 kcal/day less and expended 128 kcal/day more, and still ran +18% energy balance against +32% [V]. At above 3 g/kg body weight protein, participants still overate. The accompanying commentary is titled “More protein in ultra-processed foods: no shortcut to eating less.”

This does not change the protein weight, which is justified on lean-mass grounds (Morton 2018). It does mean a high-protein score cannot be read as a proxy for a satiating food when the matrix is ultra-processed. That is exactly the inference a protein-heavy meal-scoring rubric invites.

2.4 The rubric is on the contested side of a live methodological dispute, and should say so

Rubric says (headline 4): “Remove processed meat and sugary drinks from the UPF variable and the cardiovascular association goes from 1.11 to 1.00. The entire cohort signal was those two categories.”

New evidence says: the subgroup pattern replicates well. The 2026 digestive cancer meta-analysis gives overall HR 1.12 (1.05-1.20) but ultra-processed meat/protein products 1.33 (1.15-1.53) [V]. EPIC (n=266,666) gives the “plant-based alternatives” UPF subgroup HR 0.97 (0.91-1.02), null, against animal-based UPF 1.09 (1.05-1.12) [V]. Three cohorts separating healthful from unhealthful plant-based UPF (Rauber 2024, Prioux 2025, Thompson 2026) all put the harm on the unhealthful axis, and Thompson 2026 finds a high-UPF healthful plant-based index performs essentially as well as a low-UPF one on mortality, T2D and CVD [V].

But the Lancet Series authors now formally reject this entire class of analysis as “conceptually and methodologically flawed,” in a BMJ Analysis titled “Misleading narrative of ‘healthy’ ultraprocessed foods” (2026;392:e087538) [V]. The rubric asserts the subgroup finding as settled. It is not settled; it is the central live dispute in the field. The rubric should keep the position and add the caveat.

2.5 Carrageenan: the rubric leads with the wrong result

Rubric says: “Wagner 2024 BMC Medicine: RCT crossover, n=20 healthy young men, 250 mg 2x/day for 2 weeks, increased small-intestinal permeability (p=0.03).”

New evidence says: the primary outcome was insulin sensitivity, and it was null. Verbatim: “Overall insulin sensitivity did not show significant differences between the treatments.”

[Corrected on review: an earlier version of this section claimed every positive finding was a BMI-interaction subgroup result. That is wrong.] The insulin-sensitivity positives are the subgroup ones (OGTT-ISI p=0.04, fasting IR p=0.01, hepatic ISI p=0.04, overweight participants only). The permeability positive that the rubric actually cites is a whole-sample secondary outcome: lactulose-mannitol ratio elevated across the sample (N=19, P=0.03), with zonulin at P=0.05 post hoc [V].

So the rubric is partly wrong, not wrong: it should state the primary null, and it should label the permeability result as a secondary outcome in a 20-man trial. The 0.25 weight is unchanged. Two 2026 items add caution in opposite directions: a Clin Exp Allergy letter reframing mechanistic inference from carrageenan epithelial-injury models, and a degradation study showing food-grade iota-carrageenan does not produce sub-20 kDa poligeenan fragments under gastric conditions [V], which supports the rubric’s existing poligeenan caveat.

2.6 Titanium dioxide now has human intervention data, and the tier does not move

Rubric says: “EFSA’s 2021 ‘no longer safe’ verdict rested on unresolved genotoxicity rather than on that study.” Weight 0.25, grouped with BHA, BHT and food colours as a marker.

New evidence says: Bischoff 2026 (Mol Nutr Food Res, PMID 42613907) ran a randomised crossover human dietary intervention, n=31, 2 mg/kg bw/day E171 for 2 weeks, with faecal titanium confirming exposure (6.3 to 377 mg/kg) [V]. GSEA of colon biopsy transcriptomes gave 73 upregulated KEGG pathways (BH FDR < 0.05), one of whose labels is “colorectal cancer”.

[Corrected on review: an earlier version of this section was headed “titanium dioxide should move up”. Withdrawn.] The right statement is narrower: a human intervention now exists; the signal is transcriptome-only; the inflammation markers are null; tier and weight are unchanged. Specifically there is no comet assay, no micronucleus assay, no 8-oxo-dG and no histology; hs-CRP is P=0.076, eight cytokines are null, faecal calprotectin is null, and superoxide is a geometric mean ratio of 1.175 (0.995-1.389), P=0.057. Blinding is not described. A KEGG pathway label named after a disease is a gene-set name, not evidence of that disease.

What does change: the dose sits within EFSA high-consumer estimates, and this is the only additive in the Tier 2 table with a human randomised exposure and a target-tissue molecular readout, so grouping it with “rodent forestomach tumours in an organ humans lack” is no longer the right comparison. That is an argument about kind of evidence, not about weight.

2.7 Erythritol: the reverse-causation case strengthened, and the weight holds anyway

Rubric says: “The observational signal is likely reverse causation; the pharmacodynamics survive that critique.” Weight 0.5, joint-highest in the table.

New evidence says the first clause is now well supported:

• ATBC (Nutrients 2024, n=4,468 Finnish male smokers, serum banked 1985-1993, 19.1 y): total mortality HR 1.50 (1.17-1.92), CVD 1.86, cancer 1.54 [V]. Caveat added on review: commercial erythritol use began in the 1990s and natural dietary sources exist, so “endogenous only” is an inference.
• ARIC (JACC Advances 2025, n=4,006): total mortality erythritol 1.18 (1.10-1.26) versus erythronate 1.61; CVD death 1.19 versus 1.72 (1.43-2.06). Intake never measured [V].
• FDA’s own 2023 memo states that “detected erythritol levels are likely a reflection of endogenous production” [V], and in the same memo says the mechanistic studies “suggest that erythritol can augment platelet activation and enhance thrombosis potential” and recommends coagulation testing after ingestion.
• Mendelian randomisation is uninformative in both directions. Khafagy 2024 is null (b = -0.033 ± 0.02, P = 0.14) but draws instruments from the same small GWAS as the two positive MRs, so it shares their weak-instrument problem [V].

[Withdrawn on review: the earlier conclusion that the 0.5 weight “should move down”.] The reason is a category error in the earlier draft. Every item above addresses whether the cohort association is causal. The rubric’s 0.5 does not rest on the cohort association; it rests on pharmacodynamics, and there the only human interventional data (Witkowski 2024, n=10, within-subject, unrandomised, uncontrolled) are positive, with FDA itself asking for coagulation testing. Weak positive evidence on the load-bearing question is not refuted by strong evidence on a different question.

Hold the 0.5 until NCT05967741 reads out, and note its dose when it does: 1 g/kg/day, about 70 g/day, well above ordinary dietary exposure, so a positive result will need a dose-relevance argument and a null will be a null at a high dose. Registry status is RECRUITING, estimated completion 2026-05-31, last updated 2025-12-05 [V]. That is still the single highest-value pending readout for this rubric.

2.8 Inorganic phosphates have the highest weight and the thinnest new support

Rubric says: weight 0.6, the largest in the additive block, kept “on dose/mechanism alignment, not on a demonstrated outcome.”

New evidence says: nothing new supports it. The only relevant 2024-2026 paper is Dunford & Calvo 2025 (AJCN, PMID 40180501), a cross-sectional label survey of 39,937 US products [V]. Two problems with using it:

1. It repeats the “more rapid/efficient absorption” claim, and Calvo is the author of the 2013 review the rubric already flags as an in-vitro digestibility ceiling rather than an absorption measurement. Same claim, same author, not independent corroboration.
2. Its headline “56% of products contain phosphate additives” is driven by lecithin at 32%, an organic phospholipid that the rubric explicitly forbids from entering the inorganic-phosphate detector.

The rubric’s own reasoning already notes Mendelian randomisation finds no causal effect of serum phosphate on CHD, heart failure, atrial fibrillation or hypertension, and that controlled feeding leaves fasting serum phosphate unchanged. The 0.6 weight is now the least evidence-supported number in the table. Its defence is the “bonus signal” argument that phosphates track cured meat, which is a sodium and processed-meat proxy, and which the processing block already scores at 0.7.

2.9 Uncured is still marketing, but the mechanism underneath is wrong

Rubric says: “‘Uncured’ and ‘cultured celery powder’ are chemically identical to added nitrite.” Processed meat weighted 0.7.

New evidence says: the label claim is confirmed and strengthened. Gueraud 2023 (npj Sci Food, rat AOM model, six diets) found vegetable-stock replacement behaved like full nitrite for N-nitroso compound formation [V].

The implied causal story is weakened, not refuted. EPIC (n=367,463, 15-year median, 5,115 incident CRC) measured nitrosyl-heme and found HR 1.01 (0.93-1.09) for incident colorectal cancer, null overall and in every subsite and subgroup [V]. But nitrosyl-heme is one route of the nitrite hypothesis rather than the whole of it, and the exposure was assayed in 52 Spanish products and extrapolated to other countries’ FFQ items.

[Withdrawn on review: the earlier conclusion “nitrite is probably not the operative agent” is not supported by these data.] The same rat experiment cuts the other way on the point that matters most: cutting nitrite from 120 to 90 mg/kg reduced mucin-depleted foci, which is a dose-response result in favour of a nitrite mechanism. What failed was complete removal, which “induced a strong increase of lipid peroxidation” and so substituted one mechanism for another [V].

Keep the 0.7 weight. Treat nitrite as unresolved, with heme iron, salt and cooking chemistry as live co-candidates rather than replacements. The NutriNet 2026 result where potassium sorbate (HR 2.15), sodium ascorbate (1.41) and sodium erythorbate (1.43) scored as high as or higher than nitrite [V] is the tell: sodium ascorbate and erythorbate are the cure accelerators added specifically to block nitrosamine formation. When the inhibitor scores like the agent, the model is reading “cured meat,” not chemistry.

2.10 Fibre: the hard-outcome case is intact, the supplement case is not

Rubric says (headline 7): “Fiber has the strongest outcome evidence of anything on this list.”

New evidence says: unchanged for dietary fibre and hard outcomes (Reynolds 2019 stands). But do not let that license fibre supplementation:

• The largest fibre RCT to date (Nature Communications 2025, n=802 prediabetic adults, 6 months) was null on its primary HbA1c endpoint and on every secondary outcome [V].
• One 12-week trial of potato fibre and sugar beet pectin on a high-protein background found whole-body insulin sensitivity decreased (p=0.034), colonic permeability up (p=0.046) and IL-6 up (p=0.025) [V].
• A three-arm trial found benefit in all arms including the non-fermentable cellulose control, so fermentability may not be the mechanism [V].

2.11 NOVA: the rubric’s advice is right, the recommended instrument is now superseded

Rubric says (headline 10 and the NOVA section): “NOVA tells you almost nothing about a specific food … Use a continuous marker count instead of a 1-4 label.”

New evidence says: correct in direction, and two things now exist that the rubric should adopt.

1. An objective biomarker of UPF intake. Poly-metabolite scores (PLoS Medicine 2025, IDATA n=718, validated in the n=20 Hall crossover): r >= 0.47 continuous, AUC 0.66-0.72 in held-out test sets, and within-person separation between 80% and 0% UPF phases at p<0.001 [V]. Discrimination is modest, but this is the instrument that answers the measurement objection which the Lancet Series authors conceded in print.
2. A published continuous processing metric. Roos & Forde 2026 (Curr Res Food Sci) propose Food Processing Levels and Processed Food Intake descriptors and report that, expressed quantitatively, there is substantial overlap between the UPF and non-UPF arms of the feeding-study corpus [V]. A “marker count” is a proxy for this; FPL/PFI is the thing itself. Caveat added on review: PFI is the Forde group’s own instrument, and Forde is also an author of the largest trial (1.2.2) whose effect T7 would be used to explain. Using PFI to adjudicate T7 therefore carries a circularity risk, and any such analysis needs an independently constructed processing metric as a sensitivity check.

The rubric’s disclosure-depth trap is unaffected and remains, in my view, the most durable methodological point in the document.

2.12 Two Tier 3 entries acquired signals worth watching, not weighting

Rubric says: lecithin “is an organic phospholipid and must never enter that detector”; potassium sorbate and sodium benzoate have “no human outcome evidence at food doses.”

New evidence says, at tiers too low to justify a weight change:

• Allergy 2026 (gut-on-a-chip, human colon organoids, murine): soy lecithin and DATEM produced dose-dependent barrier disruption at daily-exposure doses, activated TNF/NF-kB/UPR, and soy lecithin induced IgE in mice while DATEM enhanced IL-4-induced IgE in human PBMCs [V]. Tier E. The rubric’s detector rule (keep lecithin out of the phosphate list) is about chemistry and is still correct; the assumption that lecithin is therefore inert is now contested at the bench.
• NutriNet-Sante 2026 gives potassium sorbate HR 2.15 for T2D, higher than nitrite [V]. Tier C, almost certainly confounded by cured meat (see 2.9), but it is a data point the Tier 3 justification does not currently have.

2.13 The emulsifier confound is real for subtractive trials, and ADDapt escapes it

Rubric says: “the larger ADDapt trial of emulsifier restriction found benefit (49% vs 31% response).”

New evidence says: the numbers are right (n=154, 49.4% vs 30.7%, multicentre, randomised, double-blind, placebo-controlled re-supplementation, NCT04046913), but two qualifications matter [V]:

1. ADDapt is still abstract-only. The results are published as ECCO 2025 abstracts in J Crohns Colitis 19(Suppl 1):i262 and i1460. I found no full peer-reviewed primary paper as of 2026-09-06.
2. The FOAM sub-analysis shows the confound for subtractive designs (Clin Nutr ESPEN 2026, n=60 healthy volunteers, 6 weeks): removing emulsifiers cut emulsifier intake 96.6% and cut UPF intake 37.1% as an inadvertent side effect [V]. [Corrected on review: an earlier version generalised this to “no emulsifier-restriction trial, ADDapt included, can attribute its effect to emulsifiers rather than to processing”. That is wrong, and it also contradicted this document’s own description of ADDapt in section 1.6.4.] ADDapt is placebo-controlled re-supplementation on a fixed low-emulsifier diet: both arms eat the same diet and the randomisation adds emulsifier or placebo. Processing level is held constant by construction, so ADDapt is precisely the design that separates emulsifiers from ultra-processing. The FOAM confound applies to subtractive free-living trials, not to this one.

Genuinely new on the other side: ENIGMA (Gut 2026, 1,461 biosamples from 487 subjects) is the first human quantification of polysorbate-80 metabolism, found native P-80 undetectable with distinct degradation routes in Crohn’s versus controls, and separated active from inactive disease with AUC 0.86-0.94 [V]. Cross-sectional, so tier D, but biomarker-based rather than FFQ-based, and therefore immune to the dietary-instrument critique.

2.14 Personalization does not displace any of this, and its ceiling is now quantified

Rubric says: nothing about personalization. Worth adding a line, because the question comes up whenever a scoring rubric is proposed.

New evidence says: in PREDICT 1’s own ANOVA the person-by-meal interaction is 7.63% (95% CI 6.11-8.96) of glycaemic iAUC variance, against meal composition 16.73% and a global person offset 18.74% [not verified: supplementary ANOVA, not re-fetched; the abstract figures are verified]. Underneath that, duplicate-meal CGM ICC is 0.17-0.28 [V, Hengist 2025], inter-brand CGM correlation is r=0.56-0.61, and consumer CGMs overestimate time above range by 3.8-fold in healthy people [V]. Every personalization RCT with an active, intensity-matched comparator was null or trivially positive; the one clearly positive trial (ZOE METHOD) compared an app-plus-coaching program against a leaflet, was ZOE-funded and ZOE-analysed, missed one of two co-primaries, and has no randomised data beyond 18 weeks [V].

A fixed rubric that scores food composition is not obviously worse than a personalized score, and the METHOD authors themselves concede that “the matrix and processing level of foods can have major effects on health” [V].

---

3. Novel theory candidates

Each is stated so it could be wrong, with the observation that would kill it. None of these is established. They are the hypotheses the 2024-2026 evidence makes newly worth testing.

Reordered and re-labelled after adversarial review (2026-09-06). The original numbering (T1 through T9) is kept so that cross-references elsewhere in this document still resolve, but the sections are now ordered by confidence rather than by novelty, and each carries an explicit confidence line plus a prior-art line where the idea is not actually new. Order: T7, T9, T1, T8, T4, T2, T3, T5, T6.

T7. Overlap invalidates the categorical trial literature

Confidence: moderate, and the highest-value item on this list. Cheap to run from published data, and decisive in one direction: a flat slope kills it outright. Risk: PFI is the Forde group’s own instrument and Forde is an author of one of the trials whose effect size would enter the regression, so the analysis needs an independently constructed processing metric as a sensitivity check (see 2.11).

Claim. Because UPF and non-UPF study diets overlap substantially in cumulative processing exposure when measured continuously, the between-arm contrast in most published trials is much smaller than the NOVA labels imply, and effect-size heterogeneity across trials is largely explained by how big the real processing gap was.

Supporting. Roos & Forde 2026 report exactly this overlap in the feeding study corpus.

Falsifier. Compute PFI for every arm of every published processing trial and meta-regress effect size on the PFI gap. T7 predicts a strong positive slope and that trials with small PFI gaps (Rego 2026, the Brazilian acute trials) are the null ones while trials with large gaps (Hall 2019, Dicken 2025) are the positive ones. A flat slope refutes it. This analysis requires no new data collection.

T9. There is a hard ceiling on personalized glycaemic advice

Confidence: high. Novelty: low. This is close to an arithmetic consequence of variance decomposition plus device reliability, and the intensity-matched RCTs already agree with it. It belongs here because it is well supported, not because it is new.

Claim. The maximum achievable advantage of CGM-personalized dietary advice over good generic advice is bounded by the person-by-meal interaction variance and by device noise, and is smaller than the effect of an intensity-matched behavioural program.

Supporting. Person-by-meal interaction 7.63% of iAUC variance [not verified]; duplicate-meal ICC 0.17-0.28; inter-brand meal-quintile flip probability 19%; Della Corte 2026 showing that scaling a person’s own reference curve by the food’s population mean GI predicts better than that person’s own test-retest noise; and every intensity-matched RCT null or trivial.

Falsifier. An RCT of CGM-personalized advice against a contact-matched, intensity-matched generic-advice arm, powered for HbA1c. T9 predicts a between-group HbA1c difference below 0.2%. A difference above 0.3% would refute it.

T1. Oral-processing resistance, not NOVA category, is the causal variable

Confidence: moderate that texture is a lever; low that it is the variable. The strong form (“NOVA is a lossy proxy for a mechanical property”) outruns the evidence, because the largest supporting trial is confounded with fat, sugar and protein (1.2.2).

Prior art, so this is not a new theory. Eating rate as an intake determinant is de Graaf and Forde’s own long-running programme; Robinson et al. 2014 meta-analysed slower-eating interventions; Hall 2019 already reported a kcal/min difference; Lasschuijt 2023 (n=18) is the clean texture manipulation. What is new in 2026 is duration and size, not the idea.

Claim. The energy-intake effect attributed to “ultra-processing” is substantially mediated by texture-derived eating rate, which is correlated with but separable from NOVA class. NOVA is a lossy proxy for a mechanical property.

Supporting, with the corrections from 1.2.2 and 2.1 applied. Forde 2026: 369 kcal/day between two UPF arms whose intended contrast was texture, matched on palatability, portion, energy served and non-beverage energy density and sustained across 14 days with no diet-by-time interaction, but also differing in fat, sugar, protein and water intake. Lasschuijt 2023 (n=18) is the clean but small version. NCT05290064: eating rate similar across arms (imprecise contrasts, measured in g/min while energy density doubled), and processing alone contributed a non-significant 128 kcal. Brunstrom 2026: unprocessed meals were 57% larger by mass at 1.08 vs 1.96 kcal/g.

Falsifier. A randomised trial of a high-texture-resistance, high-NOVA-4 diet against a minimally processed diet matched on energy density and eating rate. If the minimally processed arm still eats less, texture is not the mediator. Cheaper first pass: meta-regress the published processing-trial effect sizes on measured or estimated eating rate. If the slope is flat, T1 is dead.

Corollary that is testable now. Roos & Forde’s PFI descriptors let you compute a continuous processing gap for every published feeding trial. T1 predicts trial effect sizes correlate with the eating-rate gap and not with the NOVA contrast. That regression can be run today from published data.

T8. Additive epidemiology is measuring co-occurrence with cured meat, not toxicology

Confidence: moderate. The ascorbate and erythorbate hazard ratios are hard to explain any other way, and the falsifier is runnable from published supplementary tables.

Claim. In additive-exposure cohorts, an additive’s hazard ratio is predicted by its co-occurrence rate with processed meat and other incriminated categories, not by any independent toxicological property.

Supporting. NutriNet-Sante 2026: potassium sorbate HR 2.15, sodium ascorbate 1.41, sodium erythorbate 1.43, sodium nitrite 1.50. Ascorbate and erythorbate are cure accelerators whose function is to prevent nitrosation. There is no toxicological account on which they should score like nitrite.

Falsifier. Compute, from a food-composition database, each additive’s co-occurrence frequency with processed meat and with sugar-sweetened beverages. Regress the published NutriNet hazard ratios on those co-occurrence rates. T8 predicts R² is high and that residual variance does not track any toxicological ranking. If some additives deviate strongly from the co-occurrence line, those are the real candidates and T8 is refuted for them.

This is a direct generalisation of the rubric’s own disclosure-depth trap from label-scanning to cohort epidemiology, and it is runnable from published supplementary tables.

T4. Plasma erythritol is a pentose-phosphate flux marker, and dietary erythritol is close to inert

Confidence: split, and the two halves should not be graded together. The marker half (plasma erythritol largely reflects endogenous pentose-phosphate flux) is well supported. The “dietary erythritol is close to inert” half is contradicted by the only human interventional data there is (Witkowski 2024, n=10, within-subject, uncontrolled, positive), and FDA’s own memo asks for coagulation testing.

Prior art: endogenous erythritol synthesis from glucose via the pentose phosphate pathway was shown by Hootman et al. 2017 (PNAS) and the human labelling work of Ortiz et al. 2020. The marker interpretation is theirs, not new here.

Claim. The erythritol-CVD association is entirely reverse causation via endogenous production under oxidative and hyperglycaemic stress. Dietary erythritol at food doses has no meaningful thrombotic effect.

Supporting the marker half. ATBC serum banked 1985-1993, largely before erythritol was in wide commercial use, still predicts mortality (HR 1.50). ARIC’s erythronate, a pure PPP metabolite, outperforms erythritol (CVD death 1.72 vs 1.19). Age, not adiposity or glycaemia, predicts fasting erythritol. FDA’s own memo says the cohort levels reflect endogenous production.

Against the inertness half. The bidirectional MR is null but weakly instrumented from the same small GWAS as the positive MRs, so it supports nothing. The only human interventional data (Witkowski 2024, n=10) are positive, and the FDA memo that rejects the cohort exposure interpretation also says the mechanistic studies “suggest that erythritol can augment platelet activation and enhance thrombosis potential” and asks for coagulation testing.

Falsifier, already running. NCT05967741 (UC Davis, randomised triple-blind crossover, n=24, erythritol vs aspartame beverages, 2 weeks each, 1 g/kg/day, about 70 g/day, primary outcomes P-selectin, PAC-1, annexin V, agonist-stimulated aggregation). Status RECRUITING, estimated completion 2026-05-31, registry last updated 2025-12-05. If it shows a clear platelet activation signal against an active comparator, the second half of T4 is falsified and the rubric’s 0.5 weight is vindicated. If null, note that the null is at a supra-dietary dose, which weakens rather than settles the case for lowering the weight.

Second falsifier. Any cohort that measures dietary erythritol intake alongside plasma erythritol. T4 predicts plasma predicts outcomes and intake does not.

T2. Micronutrient deleveraging, the inverse of protein leverage

Confidence: low. The entire empirical basis is a post hoc re-analysis of n=20 with r=0.78 and two constructed predictors, which is a specification search on a small sample.

Prior art: nutrient leverage as a framework is Simpson and Raubenheimer’s, and Brunstrom has argued a related “nutritional intelligence” line for years. The 2026 paper is a new application, not a new theory.

Claim. Unprocessed diets suppress energy intake because meeting micronutrient requirements forces consumption of low-energy-density items. Fortification of ultra-processed food co-locates micronutrients with calories and removes the constraint.

Supporting. Brunstrom 2026 (AJCN): a carbohydrate-fat blend index plus a fruit/vegetable score accounted for 87.6% of the 330 kcal per-meal difference (r=0.78) in Hall 2019.

Falsifier, and it is a clean one. Run the Hall 2019 design with the unprocessed arm micronutrient-supplemented (a multivitamin and mineral pill, so micronutrient adequacy is met without eating vegetables). T2 predicts intake on the unprocessed arm rises toward the UPF arm. If it does not move, T2 is wrong. The mirror test: strip fortification from the UPF arm and predict its intake falls.

Why it matters here. If T2 is right, a meal-scoring rubric should score micronutrient density per calorie, which is close to what the rubric’s potassium term is already doing as an “instrument for intact food matrix.” T2 says that instrument is measuring the causal variable, not a proxy for it.

T3. GLP-1 discontinuation shows the drive is environmental, not learned

Confidence: low to moderate. Directionally supported but weakly evidenced: the load-bearing purchase-panel result has no extractable effect size and self-selected discontinuers. Note also that the reversion phenomenon is already established in randomised data (the STEP 1 extension shows weight and cardiometabolic variables reverting after withdrawal), so the novel content here is the basket composition claim, which is the least well measured part.

Claim. UPF-heavy eating is maintained by a persistent appetitive drive that the environment continuously re-supplies, not by habit, learning or corrupted satiety signalling. Pharmacological suppression holds it down without changing it.

Supporting. Hristakeva et al. (JMR 2026, ~150,000 households): discontinuers revert toward pre-adoption spending and end with baskets slightly less healthy than their own baseline. Plus: subjective appetite differences vanish after week 20 while objective intake suppression persists to week 60 (Penn trial); the tirzepatide preference dampening showed no detected UPF-selectivity (which is not the same as demonstrated globality, see 1.9.2); and the same drug reduces alcohol and cannabis.

Falsifier. A randomised discontinuation trial with measured food intake and NOVA-classified UPF share as outcomes, with and without a structured behavioural transition. T3 predicts reversion is near-complete and largely unaffected by the behavioural arm. If a behavioural transition prevents reversion, the drive was learnable after all.

Second, cheaper test. No study has yet measured NOVA-classified UPF intake as an outcome on GLP-1 therapy. That is an unclaimed, straightforward study, and T3 predicts the UPF share of energy barely changes even as absolute intake falls.

T5. Processed meat harms through heme, salt and cooking chemistry, not nitrite

Confidence: low. The strong form is not supported. See 2.9: the EPIC null is for one nitrosation product against one outcome with exposure extrapolated from 52 Spanish products, and the rat experiment’s 120 to 90 mg/kg reduction did help, which is evidence for a nitrite mechanism. The right status for nitrite is unresolved, with heme, salt and cooking chemistry as co-candidates.

Claim, restated more weakly on review. The IARC processed-meat signal is not fully nitrite-mediated, and part of what additive epidemiology attributes to nitrite is category co-occurrence. [Withdrawn on review: the earlier form, “nitrite is a correlate of the food category, not the operative agent.” The dose-response step in the rat model (120 to 90 mg/kg reduced mucin-depleted foci) is evidence against it.]

Supporting. EPIC nitrosyl-heme, n=367,463, 5,115 CRC: HR 1.01 (0.93-1.09), null at every subsite. Complete nitrite removal in the rat AOM model gave no benefit over 90 mg/kg because lipid peroxidation rose. In NutriNet 2026, sodium ascorbate and erythorbate, the additives used to block nitrosation, scored HRs as high as nitrite.

Falsifier. A randomised human feeding trial of nitrite-cured versus nitrite-free otherwise-identical processed meat, with prespecified faecal apparent total N-nitroso compounds, urinary DHN-MA (lipid peroxidation) and faecal water genotoxicity. T5 predicts ATNC falls, DHN-MA rises, and net genotoxicity does not improve. This trial does not exist and is cheap.

Practical implication if true. “Nitrite-free” reformulation is a null-effect intervention, and regulatory effort spent on E249-E252 limits is misallocated relative to heme and cooking.

T6. Food-contact chemical migration is a measurable mediator of processing effects

Confidence: low. This is the weakest item on the list, and it was previously presented as the most interesting, which conflated novelty with support. Its two human supports are a pilot with no mediation analysis and no effect size for the chemical readouts (Capra) and a trend in a trial where weight also moved (Preston).

Claim. Part of the eucaloric processing effect on lipids and glycaemia is carried by non-nutrient chemical exposure (plasticisers, packaging migrants, thermal by-products) rather than by macronutrients or energy.

Supporting, and this is the newest idea in the document. Capra 2026: the 0% UPF arm showed reductions in the food-contact chemical 2,4-di-tert-butylphenol and in N6-carboxymethyllysine. Preston 2025: the UPF arm showed a trend toward increased serum cxMINP phthalate and decreased plasma lithium, alongside LDL:HDL and hormone changes independent of caloric load. Geueke 2025 documents 3,601 food-contact chemicals with evidence of presence in humans, 80 of them of high concern, and 59 prioritised chemicals with no hazard data at all.

Falsifier. A eucaloric, macronutrient-matched processing trial powered for mediation, with prespecified plasma phthalate metabolites, bisphenols, 2,4-DTBP and advanced glycation end-products as mediators of an LDL or insulin sensitivity endpoint. T6 predicts a non-zero indirect effect. A negative mediation analysis kills it. A cheaper version: repackage a minimally processed diet in the same materials as the UPF diet and see whether the biomarker gap closes without the food changing.

Why it is worth keeping despite the low confidence. It is the only mechanism on this list that would make “processing” causal in a way that is not reducible to energy density, texture or nutrient profile, and it is now measurable in a feeding trial. That is an argument about value of information, not about current support. It is also the mechanism the Lancet Series asserts (“increased intake of toxic compounds, endocrine disruptors”) without this level of evidence.

---

4. What to watch in the next 12 to 24 months

All entries verified from the ClinicalTrials.gov API v2 on 2026-09-06 unless marked otherwise. Ordered by how much the readout would change the picture.

4.1 Readouts due within roughly 12 months

Trial	Design	n	Primary outcome	Sponsor	Primary completion	Why it matters
NCT05967741	Randomised triple-blind crossover, erythritol vs aspartame beverages, 2 wk each, 1 g/kg/day (about 70 g/day)	24	P-selectin, PAC-1, annexin V, agonist-stimulated aggregation, platelet-leukocyte aggregates	UC Davis	est. 2026-05-31; status RECRUITING, last updated 2025-12-05	The single highest-value readout for this rubric. Bears on T4 and the erythritol weight, at a supra-dietary dose
NCT06853288	Randomised crossover, 80% vs 20% energy from UPF	20	Faecal energy loss	Columbia	2026-12	Tests intestinal energy harvest as a UPF mechanism. No other trial is measuring this
NCT04966299	Daily erythritol vs sucrose, 5 wk, adolescents	30	Insulin resistance	(see registry)	2026-07	Second erythritol readout, different endpoint and population
NCT07175701	UPF-reducing intervention with CGM	50	Change in postprandial glucose responses	(see registry)	2026-08-31	First trial combining a UPF intervention with CGM outcomes
NCT06310603	Randomised parallel, 1 year, high vs low UPF feeding	60	Body weight, body composition, fat mass %	Oklahoma State	2025-05-31, active not recruiting, readout overdue	Longest controlled-feeding processing trial registered
NCT06356220 GF-NOURISH	Randomised parallel, coeliac/gluten-free, children from age 2	120	Fat-free mass over 6 months	Boston Children’s	2026-08	One of very few processing trials in children
ADDapt full publication	Already complete, abstract-only since ECCO Feb 2025	154	CDAI response >=70	King’s College London	overdue	The 49.4% vs 30.7% result needs its methods in print before it can be weighted properly
NCT05290064 peer-reviewed paper	Complete, registry results posted 2025	36	Energy intake	NIH/NIDDK	overdue	The rubric’s headline finding is still only a registry posting

4.2 Readouts due in 2027 and 2028, worth tracking now

Trial	Design	n	Primary outcome	Sponsor	Primary completion
NCT06538831	Randomised 2 x 2 factorial controlled feeding, 6-week diets at 5% vs 75% energy from UPF crossed with nutrient density	120	Fasting LDL-C, HOMA-IR, daytime ambulatory SBP	Laval / CIHR	2027-06-30
NCT06907862 SWITCH	Randomised parallel, 12 wk, non-inferiority framing: high-UPF vs low-UPF soy protein foods inside a guideline-based diet, hypertension and obesity, 50% with T2D	300	Systolic blood pressure	Toronto	2027-06-30
NCT06749327	Randomised parallel, 4 months	330	HbA1c, total cholesterol	Queen’s Belfast	2028-12-01
NCT07436312	Randomised factorial, adults at high cancer risk	170	>=1-point WCRF score improvement at wk 12	Gustave Roussy	2027-09
NCT06920914	Randomised crossover, minimally vs ultra-processed diets in CKD	48	Fasting serum potassium	Alberta / CIHR	2027-12-01
NCT06044285 Ultra Crave	Randomised parallel, double-blind	210	UPF withdrawal symptoms, craving, low-UP intake	Michigan / NIDA	2028-08-05
NCT07717970	Metabolomics-driven UPF crossover	40	Within-person metabolomic response per UPF vs paired non-UPF diet	(registry)	2028-03
NCT05743374	Micronutrient and additive modification	70	Faecal calprotectin	(registry)	2029-06-26
NCT07361406 SweetSpot	Non-nutritive sweeteners, 6 wk	60	2 h iAUC glucose	(registry)	2027-03-26
NCT06548828	NNS reduction in pregnancy and lactation	324	Infant adiposity	(registry)	2028-02-29
NCT06874400	Randomised crossover, dairy vs plant-based alternatives	160	Blood glucose	Toronto (Anderson)	2027-04
NCT07225881 Eatwell	Randomised, R01	472	Total dietary adherence score	(registry)	2029-10-31
NCT06165952	Observational, processed foods and brain reward circuitry	162	Body fat	(registry)	2029-02-01

SWITCH is the sharpest single test in the pipeline. It asks directly whether nutrient-dense, high-UPF soy protein foods are non-inferior to low-UPF soy protein foods inside a guideline-compliant diet, on blood pressure. That is precisely the question the rubric’s headline 4 and the Lancet Series’ BMJ rebuttal are fighting about, in a randomised design with an intermediate hard outcome.

4.3 Regulatory and publication events to watch

• FDA BHT reassessment, docket FDA-2026-N-2526, comments reopened to 2026-08-31 [V], and azodicarbonamide, docket FDA-2026-N-4126. First real outputs of the post-market chemical review programme finalised 2026-05-12. The rubric currently weights BHT at 0.25 as a marker only.
• EPA PFAS rescission proposals published 2026-05-20 (docs 2026-10085 and 2026-10086) [V]. Comment periods and final rules will determine whether the 2024 MCLs for PFHxS, PFNA, GenX and the mixture survive.
• California AB 418 takes effect 2027-01-01 for brominated vegetable oil, potassium bromate, propylparaben and Red 3 [V]. FDA’s Red 3 revocation bites the same month.
• EU nitrite/nitrate limits: Commission Regulation (EU) 2023/2108 phase 1 applied 2025-10-09 with a second phase in October 2026 [R]. Watch for reformulation data and, if T5 is right, for no measurable change in anything.
• EFSA carrageenan E407: a data call reportedly closed 2025-11-25 with no new opinion published [R]. The 2018 temporary group ADI of 75 mg/kg bw/day expires on data. Same route that removed titanium dioxide.
• Voluntary US synthetic dye phase-out target is end-2026 [R]. No authorisation has been revoked, so the observable is market share, not law.

4.4 What is not coming, and should be

• No hard-outcome (mortality, incident CVD or cancer) randomised trial of whole-diet processing level exists or is registered. Every processing trial in the pipeline uses weight, lipids, glycaemia or adherence.
• No trial of whole-diet processing level in a low-income population was found in any registry. The two large UPF trials at Johns Hopkins (NCT07214805 n=14,720; NCT07533877 n=7,000) are warning-label perception experiments, not diet trials [V].
• DGA-UP is dead. Its protocol was peer-reviewed and published (Br J Nutr 2026, 10.1017/S0007114526108034, PMID 42458767) but the registration NCT07252037 (USDA Grand Forks) is WITHDRAWN, whyStopped “Change in administrative priorities. Protocol has been published, but recruitment never started,” actual enrolment 0, last updated 2026-08-07 [V, date corrected on review]. Do not cite DGA-UP as forthcoming.
• No large plant-based meat alternative cardiometabolic RCT is registered. The near-term pipeline is NCT07766928 (n=10, not yet recruiting) and NCT06874400 (n=160).
• No study measuring NOVA-classified UPF intake as an outcome on GLP-1 therapy. See T3.
• No independent replication of Marfella 2024 using an analytical method that survives the Rauert critique. Given that Py-GC-MS is declared unsuitable for exactly the two polymers Marfella reported, this is the most important missing study in the microplastics literature.

---

Hardest open questions (from review)

The twelve questions below are the ones the adversarial review of this document could not resolve. Each is stated so that a specific document, dataset or readout would answer it. They are the highest-value next reads.

1. Forde 2026, fat versus texture. Fat EN% is collinear with arm (22 vs 33 as served, 9 points as consumed). In the post hoc covariate model fat EN% has P=0.082. Is there any specification, or any subgroup of matched meals, in which texture and fat are separable in this trial? If not, how much of the 369 kcal can be attributed to texture at all?
2. NCT05290064, flat g/min. Given that energy density roughly doubled while g/min stayed near 39, what would the eating-rate mediator predict in kcal/min, and does the registry contain enough to compute it? An eating-rate account is only tested once the rate is expressed in energy units.
3. Preston 2025, the weight gain. If weight rose 1.3 to 1.4 kg on UPF in both arms, what exactly is the “independent of caloric load” claim asserting? Do the three Cell Metab letters (41500195, 41500197, 41500199) and the reply (41500199) agree on whether energy balance was matched, and did any of them obtain per-arm p-values?
4. Capra 2026, regression to the mean. With baseline glucose AUC 13,431 vs 15,349 and about 9 per arm, is there an ANCOVA adjusting for baseline in the paper, and does the p = 0.054 survive it?
5. Bischoff 2026, the KEGG label. Does the paper anywhere claim more than a gene-set label, given null calprotectin, eight null cytokines, hs-CRP P=0.076, and the absence of comet, micronucleus, 8-oxo-dG or histology data? And was the trial blinded?
6. Erythritol, which datum addresses platelets. Every 2024-2026 addition (ATBC, ARIC, IJMS, the MRs) addresses whether the cohort association is causal. Which datum, if any, addresses the pharmacodynamic claim that the 0.5 weight actually rests on, other than the n=10 within-subject study and the FDA memo’s own recommendation for coagulation testing?
7. Khafagy shares the weakness. If the positive MRs are discounted for weak instruments from a small erythritol GWAS, on what basis is the null MR exempt, given it uses the same instrument set?
8. Nitrosyl-heme extrapolation. Exposure was assayed in 52 Spanish products and mapped onto FFQ items in six other countries. How much exposure misclassification does that introduce, and is the null HR 1.01 distinguishable from an attenuated true effect?
9. Kennedy 2025, the interval. The fat-preference ratio is -8.28 (-16.91, 0.34). What effect size would that trial have been powered to detect, and does the interval exclude a UPF-selective effect of the size anyone has proposed?
10. Hristakeva, size and selection. Is there any published magnitude for the discontinuer basket shift, and can discontinuers be characterised well enough to rule out that reversion reflects who stops rather than what the drug does?
11. ADDapt’s design. Does the full paper, when it appears, confirm that the fixed background diet was genuinely low-emulsifier and identical across arms? That single fact determines whether the emulsifier literature has one clean trial or none.
12. PFI independence. Can T7’s meta-regression be run with a processing metric not authored by a group that also authored the trials being regressed, and does the slope survive?

---

Related

• /research/health/rubric/ the baseline this document updates
• [meal-service comparison removed; it is a separate private evaluation] the implementation the rubric describes