# HSEF Review: Ultra-processed Food and Excess Energy Intake

**HSEF v5.4 evidence review · original edition 2026-09-29 · red-team synthesis edition 2026-09-30**

> **Adversarial review:** Generated by Sol 6.1; adversarially reviewed by Claude Opus 5.5 (Anthropic) on 2026-10-14. 2 corrections applied (see ADVERSARIAL REVIEW AND APPROVAL). Status: **Approved as revised.**

## LAY SUMMARY

The claim is that industrial processing has an effect of its own. On this view, ultra-processed foods lead people to eat more and gain weight because of what processing does to food, not simply because of the nutrients the food contains. Examples of what processing does would be softer textures that are eaten faster, or broken-down food structure.

This review assigns **Γ — Mechanistically plausible, partial system support**, with **Moderate confidence**. Controlled feeding studies show that some ultra-processed diets make people eat more, and gain weight or fat relative to a minimally processed diet, over one to eight weeks [1-3,18]. That narrower result is causal in the settings tested. The stronger claim, that the effect does not depend on nutrient composition, has not been shown:
- A recent inpatient trial kept energy density low and limited foods that combine fat, sugar, or salt in especially appealing ways. Under those conditions the ultra-processed diet did not produce a statistically clear increase in intake over an unprocessed diet. This result is posted in a trial registry and is not yet published in a journal [18].
- Changing the nutrient make-up of an ultra-processed diet changes how much people eat [21].

Food texture and eating speed do change how much people eat, but they do so in both processed and minimally processed foods, and some ultra-processed foods are eaten slowly and are filling [4-6,23,24]. The evidence supports particular food properties as contributors: energy density, nutrient combinations, and texture, which processing often but not always creates. It has not identified a property that all ultra-processed foods share and that works regardless of what they contain. The trials are small and short, which is why confidence is moderate rather than high.

## AT A GLANCE

| HSEF Level | Confidence | Role of Mechanism | Momentum | Stability | Divergence |
|---|---|---|---|---|---|
| **Γ — Mechanistically plausible, partial system support** | **Moderate** | **reinforcing / stabilizing factor** | Mechanistic only | Stable | Contested |

| Item | Adjudication |
|---|---|
| Submitted wording | Ultra-processed food causes excess energy intake and weight gain through properties of processing itself, independent of nutrient composition. |
| Gradeable restatement (revised in this edition) | In adults given ad libitum access to a defined diet, assigning foods classified as ultra-processed (NOVA 4) causes greater energy intake and subsequent weight gain (body weight and fat mass) than a minimally processed comparator. The effect persists when nutrient composition as consumed is matched: macronutrient and micronutrient amounts, their combinations, and the compositional determinants of energy density. The effect is carried instead by properties that industrial processing confers on foods, such as texture and eating rate, food-matrix disruption, and additives, which are characteristic of the category. |
| Primary rubric | `06-metabolism-aging.md`: the closest supplied rubric, with an explicit nutrition/obesity fit limitation |
| Aggregation | Weakest-link, preregistered. Scope-matched and primary-subclaim (SC4 designated) also return Γ. |
| Decisive distinction | **Whole-diet package causality reaches Ε. Independence from nutrient composition, and category-wide processing-conferred properties, remain Γ.** |
| Most direct test to date | NCT05290064 [18]. UPF high in energy density and hyperpalatable foods vs unprocessed: +948 kcal/d. UPF with both low vs unprocessed: +128 kcal/d (95% CI −8 to 265; P = 0.065). Registry-posted, not yet peer reviewed. |

### Subclaim matrix

| ID | Subclaim | Scope | Level | Confidence | Binding reason |
|---|---|---|---|---|---|
| SC1 | UPF-rich whole-diet assignment can increase ad libitum intake | Defined controlled human menus, one week to eight weeks | **Ε** | Moderate | Replicated at three independent sites; boundary mapped [1-3,18] |
| SC2 | The higher intake produces relative short-term weight or fat gain | Relative randomized condition effects; absolute change reported separately | **Ε** | Moderate | Human functional outcome in bounded, short contexts |
| SC3 | The effect survives selected, as-offered aggregate nutrient matching | Only the variables and menus actually matched | **Ε** | Moderate | Hall rules out several offered aggregate variables, not composition as consumed |
| SC4 | The effect persists when nutrient composition as consumed (including compositional determinants of energy density and hyperpalatable nutrient combinations) is matched | UPF vs minimally processed diets at matched composition | **Γ** | Moderate | Most direct contrast is non-significant [18]; matched designs null [8,11,22]; composition moves intake within UPF [21] |
| SC5 | Processing-conferred properties that raise intake are characteristic of NOVA 4, so category status predicts excess intake beyond composition | Foods, populations, durations | **Γ** | Moderate | Texture/eating-rate effects operate in both directions within NOVA 4 [4,6,23,24]; category assignment unreliable [27] |

**Aggregation derivation:** SC1 = Ε; SC2 = Ε; SC3 = Ε; SC4 = Γ; SC5 = Γ. The submitted sentence requires all five clauses, so weakest-link aggregation returns **Γ**. Scope-matched aggregation scores the claim by the subclaims whose scope matches it (SC4, SC5) and returns **Γ**. Primary-subclaim aggregation with the distinctive clause (SC4) designated returns **Γ**; it would return Ε only if SC1 were designated, which would grade the sentence by its least distinctive clause. A narrower claim limited to SC1 would score **Ε**.

## PROVENANCE

- Scientific evidence cutoff: **2026-09-30** (red-team pass; original cutoff 2026-09-29).
- Sources: **29** ledger rows. **25 Tier A** primary records, including one registry-posted results record [18] that is not peer reviewed. **4 Tier B** definition or synthesis records [15,17,28,29], used for framing only; they carry no numbers.
- Claim register: **35 verified**, 0 unverified, 0 contradicted.
- Every PMID, DOI, and PMCID was re-resolved in the red-team pass. None was corrected or removed, and 12 sources were added.
- Access depth is stated on each ledger row. Full text was read for the load-bearing trials [1,3,4,6,7,8], and the complete posted-results module was read for [18]. Newly added journal records other than [27] were read at abstract depth and carry no claims beyond their abstracts.
- Shared cohorts:
  - Hall / Sciarrillo / Brunstrom = **NCT03407053**
  - Forde / van Bruinessen = **NCT06113146**
  - Gama / Barros = **RBR-10gfr3fb**
  - NCT05290064 is a separate NIH cohort from the same research group as Hall 2019.
- Corrections: both Hall errata and the UPDATE supplementary correction are attached to their parent records. No retraction or expression of concern was found.
- Scite citation-context counts were unavailable in both passes and were not inferred.
- Rubric packet: SHA256 `7E097D3D50D2068F68BDB1CD022DF4A045DAC46A9F25E51262EC7568E9DB11EB`; used as methodology, not as biological evidence.

## COMPETING MECHANISTIC MODELS

### 1. Proximal food-property model (energy density, hyperpalatable nutrient combinations, texture and eating rate) — Ε

This model holds that energy density, nutrient combinations that are especially rewarding, and texture-driven eating rate determine how quickly and how much energy is consumed. Industrial processing often produces these properties, but it can also produce ultra-processed foods with low energy density and slow eating rates.

The evidence:
- **Energy density and hyperpalatability, within UPF.** In the NIH four-arm inpatient trial, both properties changed intake with processing category held constant. High versus low non-beverage energy density added 661.6 kcal/d. Adding hyperpalatable foods at high energy density added a further 158.4 kcal/d [18].
- **Energy density × eating rate, within UPF-ingredient meals.** A factorial of energy density and eating rate produced a twofold intake range (570 vs 1143 kcal) in sandwiches made largely of ultra-processed ingredients [23].
- **Eating rate, within UPF diets.** Slowing texture-derived eating rate within two approximately 95% UPF diets lowered intake by 369 kcal/d over 14 days [6].
- **Texture across categories.** Hard textures lowered intake across both processing categories [4,5].
- **Meal-level analysis.** Across 2,733 inpatient meals, energy density, eating rate, and hyperpalatable foods each tracked meal energy intake [20].

Limits:
- The texture manipulations in [4] and [6] were not composition-pure: soft and fast diets were higher in fat (RT-04, RT-05).
- Energy density and hyperpalatability are partly compositional properties, so this model overlaps model 3.
- Its quantitative mediation of durable weight change is untested.

### 2. Whole-diet ultra-processed package model — Ε

Some UPF-rich menu packages increase intake and relative weight through the combined effect of formulation, energy density, food form, eating behaviour, and food choice, without requiring one residual processing essence. The evidence:
- **Hall.** Hall's inpatient crossover produced 508 kcal/day greater intake and opposite two-week weight changes [1].
- **Hamano.** A one-week crossover in nine men replicated the direction [2].
- **UPDATE.** The trial found greater relative weight and fat loss on an eight-week minimally processed diet, within net weight loss on both guideline-conforming diets [3].
- **Second NIH trial.** A UPF diet high in energy density and hyperpalatable foods raised intake by 948 kcal/d over an unprocessed diet [18].
- **Prescribed-intake trial.** A trial with prescribed rather than self-selected intake reported greater body weight on UPF regardless of caloric load [19].

The model fits these experiments without claiming that every UPF shares one causal property.

### 3. Nutrient-form, matrix, formulation, and protein-leverage model — Γ

Matching aggregate macronutrients as offered can leave behaviourally and physiologically important compositional differences:
- **Hall's diets** differed in added-sugar fraction, insoluble-fibre fraction, saturated-fat fraction, n-6:n-3 ratio, protein share, and non-beverage energy density [1]. The authors estimated that protein leverage could explain at most about half of the intake difference [1].
- **A post hoc analysis of the same meals** implicated low-energy-density food mass and the carbohydrate–fat blend [13].
- **Changing composition within UPF changes intake.** A high-protein, lower-carbohydrate UPF diet lowered ad libitum intake and raised energy expenditure relative to a normal-protein UPF diet matched for calories, fat, and fibre [21].

These findings are plausible and human-relevant, but no single compositional feature has yet explained the replicated package effect across trials.

### 4. Nutrient-independent processing model (nominal) — Γ

The nominal model predicts a UPF effect that persists when composition is matched and is carried by processing-conferred properties.

**Evidence in its favour:**
- Hall's contrast survived matching on several offered aggregate variables [1].
- Prospective associations remain after adjustment for nutritional quality [14].
- In a one-day factorial trial, soft (ultra-)processed menus produced significantly higher intake than soft unprocessed menus at equal energy density [4]. Sodium, fibre, and saturated fat were not matched.
- A prescribed-intake trial reported UPF-associated weight gain independent of caloric load [19].
- The composition-matched NIH contrast has a positive point estimate (+128 kcal/d) [18].

**Its defining test is not met:**
- That same NIH contrast is not significant.
- Matched-exposure, matched-meal, and matched-preload designs are null for later or concurrent intake [8,11,22].
- Composition changes intake within UPF [21].
- The only direct human test found for a reward-based "processing itself" mechanism detected no post-ingestive striatal dopamine response [26].

**Tournament result:** the proximal food-property model ranks first, the whole-diet package model second, the nutrient-form model third, and the nominal model fourth. The most direct separator between the first-ranked and nominal models is NCT05290064 [18]. With processing category fixed, energy density and hyperpalatability moved intake by hundreds of kcal/d. With both low, the UPF − unprocessed difference was small and non-significant. Because that result is registry-posted and not yet peer reviewed, it informs the ranking but the level does not rest on it. Lasschuijt 2023 [4] and Forde 2026 [6] are complementary separators, each with the composition caveats noted above.

## HYPOTHESIS STATE EVALUATION FRAMEWORK (HSEF v5.4)

<!-- consensus-consistency-2026-10-09 -->
### Consensus provenance reconciliation — 9 October 2026

Case-specific Consensus discovery records were located and linked to this final edition. This supplements the review's provenance; it leaves the original assessed claim, grades and source counts intact. Provider summaries and full-text badges do not establish examination of original primary passages by this reviewer. Source access and incorporation retain the limits stated in the original methods and any dated research update.

Recorded evidence artifacts:

- `deliverables/prepublication-enrichment-2026-10-05/website-series/ultra-processed-food/consensus-browser-links.json`
- `deliverables/prepublication-enrichment-2026-10-05/website-series/ultra-processed-food/consensus-browser.txt`
- `deliverables/prepublication-enrichment-2026-10-05/website-series/ultra-processed-food/consensus-targeted-check.txt`

The located artifacts do not state a reliable source-comparison count. No count is inferred.

**Existing source reconciliation:** Yes; 9 primary identifiers; selected Results passage reconciled; timeline rejects generated CRISPR caption and marks shared registry/dependent reports. No consumed-density matching inferred; shared registry reports not counted as independent cohorts; service body access is distinct from selected external body check.

Matched primary identifiers in the preserved dated ledger: PMID 31105044, PMID 35285882, PMID 37452167, PMID 39267249, PMID 40043691, PMID 40760353, PMID 41314613, PMID 42588167, PMID 42633852. These are historical checks, not newly credited complete-paper examinations.

Shared template wording about an amyloid-specific Deep follow-up is not credited as a case-specific run for this topic. The actual raw case records above establish provider use.

The 2026-10-06 adversarial approval attached to this case covers the dated 2026-10-05 research addendum, not this review edition, which carried no earlier adversarial approval. This provenance supplement is not a new independent or blinded adjudication. No score or confidence change is made by this reconciliation.
<!-- end-consensus-consistency -->


### 1. Hypothesis parsing and falsifiability

The submitted sentence is a conjunction of five claims:
1. an exposure effect on intake;
2. a downstream effect on weight;
3. independence from nutrient composition;
4. mechanistic attribution to properties of processing;
5. category-wide scope.

A positive randomized comparison of two menu packages can establish the first two clauses without establishing the last three.

**How this edition reads the two key phrases.** "Independent of nutrient composition" is read as its words say. The effect must persist when composition as consumed is matched: amounts and combinations of nutrients, including the compositional determinants of energy density and of hyperpalatable nutrient combinations. "Properties of processing itself" is read as the proponents' own framing, which attributes overeating to energy density, hyper-palatability, soft texture, and disrupted food matrices [28]. Texture, eating rate, matrix disruption, and additives are therefore treated as the hypothesis's *candidate mediators*, not as confounders to be removed.

**Correction from the original edition.** The original restatement required the effect to survive control of texture, eating rate, and matrix as well. That grades a residual "processing essence" no proponent defends. To be gradeable, the processing property must still be named before outcomes are measured. It cannot be defined as whatever remains unexplained after a UPF menu wins.

### 2. Rubric routing and fit

The packet contains no nutrition- or obesity-specific rubric. All seven rubrics were checked. `04-microbiome-systemic` covers microbial effects on obesity, not diet processing. `07-psychiatry-neuromodulation` covers psychiatric phenotypes; the food-reward mechanism is one candidate mediator of an intake outcome, not a psychiatric claim. `06-metabolism-aging.md` is closest, because it separates mechanistic biomarkers from organism-level functional outcomes. Its field-valid Δ wording is aging-specific:

> "Genetic or pharmacologic intervention producing reproducible lifespan/healthspan extension"
> "Must separate developmental effects from true aging effects"
> "Biomarker change alone does NOT satisfy this threshold"

That wording cannot literally classify a dietary-intake hypothesis. This review uses randomized human changes in measured energy intake and body weight or fat as the functional analogue, and it states the fit limitation openly. The adaptation is conservative for the broad claim. A positive menu contrast can cross the boundary for that menu package, but not for a processing property that was never isolated.

**The rubric's single-species hard rule.** Rubric 06 also lists "Single-species or single-strain evidence" among its hard rules that "trigger Γ cap regardless of other evidence". Read literally, it would cap every human-only subclaim at Γ. Under the framework's cap-precedence rule, a cap "governs the inference from that system to the hypothesis's claimed scope". This claim is scoped to humans and the evidence is human, so no cross-species inference occurs and the rule does not bind. The resolution is recorded rather than applied silently.

### 3. What the strongest positive experiments establish

**Hall et al. [1].** Twenty weight-stable adults were randomized to consecutive two-week inpatient diet periods. Participants ate **508 ± 106 kcal/day** more on the UPF diet. Weight changed by **+0.9 ± 0.3 kg** versus **−0.9 ± 0.3 kg**, and fat mass by **+0.4 ± 0.1 kg** versus **−0.3 ± 0.1 kg**. An order analysis found no significant diet-order effect for the UPF comparison [12]. Three further details bear on interpretation:
- Intake **declined linearly** during the UPF period (−25.5 ± 6.4 kcal/d per day) but not during the unprocessed period, so durability of the difference is uncertain.
- Offered non-beverage energy density of meals was **1.957 versus 1.057 kcal/g**.
- The authors estimated that protein leverage could explain **at most about half** of the difference.

This is strong evidence that the assigned whole-diet package caused the short-term intake and body-composition contrast.

**Hamano et al. [2].** This one-week crossover in nine Japanese men with overweight or obesity found **813.5 kcal/day** (95% CI 342.4 to 1284.7) greater intake and **1.1 kg** (0.2 to 2.0) relative weight change, alongside fewer chews per calorie. The small, all-male, open-label sample limits precision, but the trial is an independent replication.

**UPDATE (Dicken et al.) [3].** UPDATE extended the question to eight weeks, free-living, with home-delivered diets that met UK healthy-eating guidance (50 ITT participants, 90.9% female at baseline). Both conditions produced weight loss: **−2.06%** on minimally processed food and **−1.05%** on UPF. The relative difference favoured minimally processed food by **1.01 percentage points** (95% CI 0.14 to 1.87), with **0.98 kg** greater fat-mass reduction. Two points qualify the intake result:
- Self-reported intake was **327.3 kcal/day** lower on minimally processed food, but diaries came from only 32 and 35 of 50 participants.
- The body-composition-derived difference in energy imbalance, **−170.4 kcal/day**, is the less biased index.

The corrected energy densities were **1.25 versus 1.60 kcal/g**, so UPDATE supports a package effect but not independence from energy density. UPDATE shares an author with [1], which is relevant to independence (§10).

**NCT05290064 [18].** The NIH Clinical Center four-arm inpatient crossover (36 analysed participants; four 1-week ad libitum diets; unmasked) posted its results to ClinicalTrials.gov on 2026-08-12. They are not yet peer reviewed. A UPF diet high in both non-beverage energy density and hyperpalatable foods produced **948.4 kcal/day** (95% CI 811.9 to 1084.8) greater intake than an unprocessed diet low in both. This reproduces the direction of [1] within the same research group in a separate cohort.

**A prescribed-intake trial [19].** A controlled 2 × 2 crossover in young men assigned UPF or unprocessed diets at prescribed, not self-selected, caloric levels. It reported greater body weight on UPF "independent of caloric load". The verified record is abstract-level. Magnitudes and composition matching could not be confirmed from the primary text, so none are used here.

Together these studies cross Ε for SC1 and SC2. Causal effects exist in constrained human settings, and the boundaries are visible (§6). They do not establish that every ultra-processed diet produces absolute weight gain: UPDATE is a direct counterexample to that reading. The trials are small (9 to 50 participants) and short (one to eight weeks), which bears on confidence rather than on the level (§11).

### 4. Why as-offered nutrient matching is narrower than nutrient independence

Hall's design rules out several simple explanations. The offered menus did not differ merely in total carbohydrate, fat, protein, sugar, sodium, or total fibre. That justifies Ε for SC3, as explicitly scoped.

The menus nevertheless differed in composition and physical delivery:
- The added fraction of sugar was about 54% versus 1%.
- The insoluble fraction of fibre was about 16% versus 77%.
- The saturated fraction of fat was about 34% versus 19%.
- The n-6:n-3 ratio was about 11:1 versus 5:1 [1].
- The offered energy-density match relied on beverages. Consumed energy density was **1.36 versus 1.09 kcal/g**, and non-beverage energy density and eating rate also differed [1,16].

Once people eat different quantities, their absolute nutrient intakes differ by design. "Matched as offered on selected aggregate variables" therefore cannot be translated into "independent of nutrient composition" without qualification. A 2026 synthesis of the randomized trials likewise concluded that the diets compared have not been matched for nutrient profile [15].

This does not erase the randomized effect. Energy density and texture can be downstream consequences of processing and therefore genuine mediators. What the hypothesis additionally requires is that the effect survive when *composition* is matched, and that the carrier be a processing-conferred property.

### 5. Direct separator experiments

**NCT05290064 [18]: the most direct test of the independence clause.** The trial crossed processing with energy density and hyperpalatability. With processing held at UPF:
- High versus low non-beverage energy density raised intake by **661.6 kcal/day** (95% CI 524.9 to 798.3).
- Adding hyperpalatable foods raised it by a further **158.4 kcal/day** (21.9 to 294.9; P = 0.023).

With both properties low, the UPF diet exceeded the unprocessed diet by **128.4 kcal/day (95% CI −8.2 to 265.0; P = 0.065)**. Meal eating rate (about 38–39 g/min) and palatability did not differ between any arms. The composition-matched contrast is therefore not statistically significant. Its interval does not exclude a modest residual effect of up to about 265 kcal/day, and the result is registry-posted and unpublished.

**Lasschuijt et al. [4].** This trial crossed industrial processing with hard versus soft texture at equal energy density (1 kcal/g) in 18 adults over one day per condition. Hard menus reduced daily intake by **571 ± 135 kcal (33%)**, and the processing main effect on intake including snacks was null (**P = 0.91**). The interaction was significant (**P = 0.002**), so the processing effect changed sign with texture:
- **Soft menus.** Soft (ultra-)processed menus gave the highest intake (**2608 kcal**), significantly above soft unprocessed menus (**2205 kcal**).
- **Hard menus.** Hard (ultra-)processed menus gave the lowest intake (1706 kcal). Excluding snacks, they were significantly below hard unprocessed menus.

The arms were not composition-pure. Soft menus were about 40–43% fat energy versus 24–27% for hard, fibre ranged from 43 to 110 g, and the processed arms carried about 7–14 times more sodium. This is a small one-day pilot, and it does not isolate either texture or processing.

**Teo et al. [5].** In a lunch crossover in 50 adults, hard texture reduced food mass consumed by **21%** and energy by **26%**. Intake ranged from about **483 kcal** (hard minimally processed) to **789 kcal** (soft UPF), and meal energy density varied by up to about ±0.2 kcal/g. In fixed-portion lunches in 43 men, the hard UPF meal was eaten about half as fast as the soft UPF meal and produced the greatest fullness and the lowest later snack intake [24]. Classification as ultra-processed did not by itself imply lower satiety.

**Forde et al. [6].** This trial held NOVA category nearly constant, with both 14-day diets approximately **95% UPF**. The slower texture-derived eating-rate diet lowered intake by **369 kcal/day** (95% CI 221 to 517). Fat mass fell by **0.43 kg** on that diet, with no difference in body weight. The served diets were matched for energy density, palatability, portion, and variety but not for macronutrients as served. The slow diet was 22% fat and 22% protein energy; the fast diet was 34% and 18%. The authors report that the intake effect was not attributable to macronutrient intake in adjusted models.

A secondary analysis [7] found no difference in postprandial GLP-1, ghrelin, pancreatic polypeptide, or PYY responses between baseline (habitual low-UPF diet) and after either high-UPF diet, or between the fast and slow diets. Fasting PYY was 17% higher after the fast diet than at baseline.

**Other within-UPF contrasts.** Energy density and eating rate together varied intake twofold within ultra-processed sandwiches [23]. Changing macronutrient composition within UPF lowered intake by 196 ± 396 kcal/day and raised expenditure by 128 ± 98 kcal/day over 54 hours in a calorimeter [21].

These studies support the first-ranked model and hold SC4 and SC5 at Γ. Processing can create causally relevant properties, and those properties go in both directions within NOVA 4. No positive experiment shows the UPF effect persisting once composition is matched.

### 6. Null and boundary-mapping studies

**Rego et al. [8].** Twenty-seven emerging adults ate two 14-day eucaloric diets (81% vs 0% UPF) matched on macronutrients, fibre, added sugar, diet quality, and energy density, followed by an ad libitum buffet. Overall energy intake, food mass, UPF selection, and eating rate did not differ. An exploratory analysis found that late adolescents (18–21 years) ate more after the UPF diet (p = 0.03, d = 0.79) while young adults did not. That subgroup result is hypothesis-generating only. The design tests acquired appetite or reward adaptation, not intake during continuous ad libitum exposure, so it narrows the mechanism without contradicting Hall's design.

**Larcom et al. [22].** Forty young adults ate single meals matched for energy, macronutrients, flavour profile, and texture. Intake was similar for the non-UPF meal (**585.7 kcal**) and the UPF meal of equal nutritional quality (**609.7 kcal**), although energy intake rate was higher for the UPF meal. Intake was lower (**503.0 kcal**) for a less palatable UPF meal of low nutritional quality.

**Galdino-Silva et al. [9].** In a parallel trial of 42 adults with obesity, a matched UPF breakfast was eaten in **7:52** versus **11:07** minutes, with less chewing, but there was no difference in metabolic outcomes and only a marginal difference in capacity to eat. This supports oral processing as a proximal mediator but stops short of later intake or weight.

**The Gama/Barros 2 × 2 preload trial.** This trial shows why endpoints must remain separate. An exploratory secondary report found higher postprandial insulin after the UPF preloads (**P = 0.016**) [10]. The registered primary report found no processing effect on later dinner intake (**−24.5 kcal; 95% CI −87.0 to 38.0**) [11]. Both reports come from one trial, and the hormone result cannot substitute for the null functional endpoint.

**Longer-term free-living evidence.** A 12-month randomized trial in 148 adults with obesity added UPF-restriction advice to energy restriction. The difference in UPF share achieved was small (13.86% vs 20.02%, p = 0.08), and final weight was 82.9 versus 86.3 kg (p = 0.01), a difference the authors judged not clinically significant [25].

These nulls are not evidence that all UPF diet effects are false. They show that the category effect depends on the exposure design, the proximal properties, and the timing of the outcome. That boundary supports Ε for specific packages and Γ for independence from composition and for category-wide properties.

### 7. Observational relevance and its limit

The NutriNet-Santé cohort (110,260 adults; 78.2% women) linked each 10-percentage-point higher UPF share to incident overweight (**HR 1.11**, 95% CI 1.08–1.14) and obesity (**HR 1.09**, 1.05–1.13). The associations persisted after adjustment for sociodemographic and lifestyle factors, energy intake, and nutritional quality of the diet [14]. The authors themselves list residual confounding, selection, and classification limitations. Prospective ordering makes reverse causation less likely than in cross-sectional data, but it cannot isolate processing from composition, food choice, access, or other correlated exposures.

Exposure classification is itself imprecise. When food and nutrition specialists assigned foods to NOVA groups, agreement was low (Fleiss κ **0.32** for marketed foods with ingredient lists, **0.34** for generic foods) [27]. This bears on both the observational estimates and the transfer of trial menus to "the category".

A 2026 meta-analysis of randomized trials [15] reached four conclusions:
- Whether intake was higher on UPF arms depended on the analytic model.
- Weight gain and eating rate were significantly greater on UPF arms.
- Intake effects appeared confined to trials in which the UPF arm had higher energy density.
- Certainty was low.

Because that source is Tier B and was read at abstract depth, it informs heterogeneity and divergence only. Primary experiments determine the score.

### 8. Cap and Β/Γ/Δ boundary adjudication

No universal pure-association or biomarker-only cap binds SC1–SC3, because randomized human functional outcomes exist. The rubric's single-species hard rule does not bind a human-scoped claim (§2). The cap-4 biomarker rule does bind the insulin finding [10]: it cannot stand in for the null intake endpoint [11].

**SC4.** The barrier is not a generic low-evidence cap but a failed Δ boundary. No positive field-valid perturbation shows the UPF effect persisting when composition as consumed is matched. The one direct composition-controlled whole-diet contrast is non-significant [18], and the matched single-meal, exposure, and preload designs are null [8,11,22].

**SC5.** Multiple human interventions exist, but processing-conferred properties act in both directions within NOVA 4 [4,6,23,24]. A category label with κ about 0.3 [27] cannot carry a category-wide causal prediction.

The whole-diet perturbations are phenotype-relevant. Hall, Hamano, UPDATE, and NCT05290064 change intake and relative weight or fat, so SC1–SC3 cross Δ and reach Ε in their bounded contexts. A null field-valid perturbation does not cross Δ for a broader claim (pitfall 17). Equally, an unmatched positive contrast is not a composition-independent one.

### 9. Threshold testing upward

The metabolism rubric defines Ε as: **"Requires causal effect in constrained contexts, backgrounds, or species."** SC1, SC2, and the deliberately narrow SC3 satisfy this requirement. Their boundaries are mapped, not merely unexamined (pitfall 16): the effect disappears when energy density and hyperpalatability are low [18], and it becomes a relative difference within net weight loss under healthy-guideline menus [3].

Ζ requires:

> "Reproducible intervention effects"
> "Mechanistic narrowing"
> "Evidence not reducible to general sickness, caloric reduction, or developmental artifacts"

The whole-diet effect is reproducible in direction across independent sites. The mechanism has narrowed, but toward energy density, nutrient combinations, and texture rather than toward a composition-independent processing property. The only replicated lines are one evidence class (short controlled feeding), and several of their positive results share a research group. No subclaim reaches Ζ.

### 10. Falsification and pitfall audit

**Three falsification patterns are met:**
1. A hormonal change without a change in later intake appears in RBR-10gfr3fb [10,11], so a biomarker cannot carry the phenotype.
2. Effects vary across menus and disappear, or reverse sign, in matched or factorial contexts [4,8,11,18,22].
3. Weight or fat differences track realized intake rather than demonstrating a storage effect of process history. The prescribed-intake result [19] is the one possible exception, and it could not be verified beyond its abstract.

**Interpretive pitfalls flagged:**
- a category label treated as a mechanism (pitfall 15);
- selected matching treated as exhaustive independence;
- relative and absolute weight outcomes conflated;
- shared cohorts counted as replication;
- post-outcome definition of "processing properties";
- a biomarker taken as confirmation (pitfall 8).

**Replication within research networks (pitfall 9).** Positive whole-diet evidence and several analyses of it come substantially from one inpatient research group and its collaborators [1,12,13,18,20]; that group also shares an author with [3]. Most texture and eating-rate separator evidence comes from one sensory-science network [5-7,16,23,24], and one of its members co-authored [1]. Three matched-meal and preload studies and the 12-month trial come from one Brazilian group [9-11,25]. Independent groups contribute [2], [8], [19], [21], and [22]. Replication inside a network raises confidence less than replication across networks. This is recorded as a reason for Moderate confidence, not as a criticism of any group.

The score card records each pitfall and the shared study identifiers.

### 11. Steelman and confidence

**The case for Δ on the full claim (written first).** Randomized whole-diet trials at three independent sites cause large intake and relative weight effects [1-3], and a second NIH cohort reproduces the direction [18]. Hall matched several nutrients as offered [1]. Cohort associations persist after adjustment for nutritional quality [14]. At equal energy density, soft (ultra-)processed menus beat soft unprocessed ones by about 400 kcal in a day [4]. The composition-matched NIH contrast is positive (+128 kcal/day) with an interval reaching +265 [18]. A prescribed-intake trial reports UPF-associated weight gain regardless of caloric load [19].

**Why the case fails for SC4 and SC5.** None of these contrasts is composition-matched and significant. The soft-texture contrast carried about 7 times the sodium, less fibre, and more saturated fat. The NIH contrast is non-significant and unpublished. The prescribed-intake trial's composition matching is unverified. Composition-matched designs are null [8,11,22], and composition changes intake within UPF [21].

**The case for Β.** "Processing itself" is loosely specified, and the category is unreliably assigned [27]. Β is rejected because randomized whole-diet and property perturbations show real causal effects of process-linked properties in partial systems. Γ is therefore the stable landing under either reading of the claim (§1).

**Confidence is Moderate, not High.** The framework reserves High for "extensive independent replication across labs, models, and assay modalities". Here the relevant trials are small (9 to 69 participants), short (one meal to eight weeks), in part composition-confounded, and concentrated in a few research networks (§10). One key contrast is registry-only [18], and several nulls cannot exclude modest effects. Confidence measures evidence density and reproducibility, not certainty that the strong claim fails.

### 12. Role of mechanism

The role is **reinforcing / stabilizing factor**. The framework's steps were applied in order, stopping at the first match:

1. **Is the mechanism produced by the phenotype?** No. Randomized diet assignment precedes the intake change.
2. **Is there no causal dependency in either direction, only co-occurrence?** No. Randomized UPF packages [1-3,18] and processing-modifiable properties such as texture and energy density [4,6,18,23] causally change intake, so the process-linked mechanism is not a mere co-occurrence.
3. **Primary driver: does removing it prevent the phenotype, does engaging it initiate the phenotype, and is it rate-limiting?** No. Excess intake occurs without UPF, UPF with low energy density and low hyperpalatability does not clearly raise intake [18], and rate-limiting status is not shown.
4. **Permissive factor: is it required for the phenotype?** No. UPF exposure is not required for excess intake.
5. **Otherwise** → **reinforcing / stabilizing factor.** Processing-linked properties amplify intake in some formulations.

The classification applies to the processing mechanism as claimed. It does not deny that particular processed food designs can be causal.

### 13. Momentum, stability, and divergence

**Momentum: Mechanistic only.** The NIH four-arm inpatient trial NCT05290064 did prespecify the UPF-versus-unprocessed contrast at low energy density and low hyperpalatability, which is SC4's boundary test, but that trial completed in 2025 and its posted primary contrast is non-significant (+128.4 kcal/day, P = 0.065), so no experiment that could cross the next threshold is being run: the pending item is a journal report of a result already in hand, and by the reading in the Stability paragraph below it would not carry SC4 across Δ. The decisive contrast therefore belongs to the evidence narrative above rather than to this dimension. What remains active is mechanistic work on texture, energy density, composition and satiety [6,7,21-24]. Corrected from "Decisive work underway" in the 2026-10-14 adversarial review; Momentum modifies no letter, and the level and confidence are unchanged.

**Stability: Stable.** The named pending trigger is the peer-reviewed report of NCT05290064. Its posted primary contrast (+128.4 kcal/day, P = 0.065) would not carry SC4 across Δ. Only a materially different peer-reviewed analysis could, and none is signalled.

**Divergence: Contested.** The field presents this hypothesis in two incompatible ways:
- A major proponents' synthesis concludes that "the totality of the evidence supports" UPF-based dietary patterns as "a key driver" of diet-related chronic disease [28]. That implies Ζ or above: Ζ − Γ = **+3**.
- Critical syntheses conclude that the experimental evidence "does not establish a causal role of UPF consumption in obesity" [29] and that "evidential certainty is low" [15]. That implies Β–Γ: **−1 to 0**.

No single implied level applies, which the framework defines as Contested. The original edition's value, Severe overstatement, was computed from the wording of the submitted sentence rather than from the field's presentation. The strong wording does imply roughly Ζ, and against that wording the gap would be +3.

## PROPOSED DECISIVE EXPERIMENTS

1. **Composition-matched equivalence trial with a processing-route factor.**
   - **Design:** a residential randomized crossover at a site independent of prior trials, with at least four weeks per condition. Compare UPF and minimally processed diets matched *as consumed* on macronutrient and micronutrient amounts, hyperpalatable nutrient combinations, and total and non-beverage energy density.
   - **Processing factor:** cross this with a prespecified processing route, for example industrial versus home-style preparation of ingredient-identical foods, so that processing-conferred texture and matrix vary by design.
   - **Pre-specification:** register a smallest effect of interest so that a null result is informative.
   - **Reading the outcome:** a replicated composition-independent effect on intake and fat mass would move SC4 to Δ. An equivalence result would falsify the independence clause at the tested scale.
2. **Hall-menu causal mediation reconstruction.** Rebuild the Hall menus. Then equalize non-beverage energy density, hyperpalatable combinations, texture and eating rate, protein share, fibre type, and carbohydrate–fat blend, one prospectively registered block at a time. Measure how much of the original intake contrast survives each block, then repeat at an independent site.
3. **Long-duration within-category negative controls.**
   - **Design:** compare three diets over 12–16 weeks: slow-eaten, low-energy-density UPF; fast-eaten, high-energy-density UPF; and matched minimally processed food.
   - **Composition:** match macronutrients across the texture arms, which earlier texture trials did not do.
   - **Test:** whether NOVA status adds out-of-sample prediction beyond measured food properties.

## KNOWLEDGE GAPS

- No peer-reviewed whole-diet trial yet matches composition as consumed while varying processing. The most direct contrast [18] is registry-posted and cannot exclude a residual effect of about 250 kcal/day.
- It remains unresolved what fraction of the Hall and UPDATE effects is mediated by each of these: non-beverage energy density, hyperpalatable combinations, texture and eating rate, protein share, matrix, and food choice.
- The texture and eating-rate manipulations to date also changed macronutrient or sodium content [4,6]. A composition-pure texture trial of whole diets is lacking.
- Most controlled trials last one meal to two weeks. The eight-week trial kept an energy-density difference and produced weight loss on both diets [3]. The 12-month trial achieved only a small difference in UPF exposure [25].
- It is unexplained whether UPF-associated weight gain at prescribed intake [19] reflects composition, metabolizable energy, water or glycogen, or adherence.
- Within-category negative controls are underused, NOVA assignment is unreliable between assessors [27], and the prevalence of high- and low-risk UPF designs is unknown.
- Children, adolescents (an exploratory signal exists [8]), older adults, diverse food systems, and people with metabolic disease are underrepresented in controlled feeding evidence.

## LIMITATIONS OF THIS REVIEW

- **Rubric fit.** The primary rubric was written for metabolism and aging, not nutrition and obesity. Its field-valid intervention wording is not a literal fit, and the adaptation is stated rather than hidden.
- **The trials themselves.** Most primary trials are small, short, and unblinded, and their menu contrasts cannot be generalized to every food in a heterogeneous category.
- **Access depth.**
  - Several records were available only at abstract depth, and each is labelled as such in the ledger. That includes the prescribed-intake trial [19], whose full text could not be retrieved.
  - One key trial [18] has registry-posted summary results only.
  - No claim is taken beyond what its read depth supports.
- **Recency.** Recent 2025–2026 publications have not yet accumulated independent replication.
- **Process.** Scite citation-context augmentation was unavailable in both passes. The red-team pass was carried out by a single model, and the literature searches, though broad, may have missed relevant work.

## CALIBRATION STATUS

This is **not** a calibration entry. It was graded as a standalone evaluation, and no independently authored sealed expected band was supplied. The claim, evidence ledger, and score were derived without a calibration target.

## RED-TEAM SYNTHESIS NOTE

This edition is the published synthesis of a cross-model red-team pass (Claude, Anthropic) on an original edition authored by a different model family.

**Citations.** All 17 original references were re-verified by identifier, and none was corrected or removed. Twelve were added, including the NIH four-arm trial NCT05290064 [18], whose registry-posted results are the most direct test of the independence clause.

**What changed:**
- The gradeable restatement no longer requires the effect to survive control of texture, eating rate, and matrix, which proponents treat as properties of processing. Independence is now tested against nutrient composition as consumed.
- Lasschuijt 2023 [4] and Forde 2026 [6] are reported with their significant interaction and their composition differences.
- Confidence moved from High to **Moderate**, to match the framework's definition.
- Momentum moved to **Decisive work underway**, and Divergence to **Contested**. (Momentum was returned to **Mechanistic only** in the 2026-10-14 adversarial review; Divergence stands.)
- **The level is unchanged at Γ** and holds under either reading of the claim.

The full audit trail is in `redteam/REDTEAM_REPORT.md` (25 findings with dispositions and a citation-verification table) and `redteam/SYNTHESIS_CHANGE_LOG.md` (before→after table and itemised changes).

## VALIDATION NOTE

The companion validator report records the HSEF structural checks for this edition. A pass confirms the following:
- schema presence and canonical descriptor use;
- aggregation arithmetic;
- source/claim links;
- access depth and status flags;
- study grouping and provenance counts.

It does not constitute independent scientific replication, inter-reviewer calibration, or external endorsement. The validator does not resolve identifiers; that was done by hand in the red-team pass.

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** Sol 6.1 (original edition), with a cross-model red-team synthesis already recorded in this document. **Adversarial reviewer for this pass:** Claude Opus 5.5 (Anthropic). **Date:** 2026-10-14.

**Scope of checks.** All 28 PMIDs in REFERENCES, plus the two errata attached to [1], were verified against PubMed metadata for title, first author, journal, year and DOI: 30/30 resolve to the stated record, 0 failures, no retraction or expression-of-concern flag, and both attached notices are genuine published errata of their parent record. Item [18] is a ClinicalTrials.gov results record and is described as such. Thirteen load-bearing claims were spot-checked against the primary abstracts and matched exactly, including the quantities the verdict turns on: 508 plus or minus 106 kcal/day with plus 0.9 and minus 0.9 kg weight change (PMID 31105044); 813.5 kcal/day with 95% CI 342.4 to 1284.7 and 1.1 kg, 0.2 to 2.0, in nine men (PMID 39267249); the UPDATE percentages minus 2.06 and minus 1.05 with a difference of minus 1.01, 95% CI minus 1.87 to minus 0.14, P = 0.024, in 50 intention-to-treat participants (PMID 40760353); 369 kcal/day, 95% CI 221 to 517, with a 0.43 kg fat-mass fall and no weight difference between two approximately 95% ultra-processed diets (PMID 41314613); the 27-participant eucaloric 81% versus 0% crossover with a null full-sample result and an exploratory late-adolescent effect at p = 0.03, d = 0.79 (PMID 41255123); 585.7, 609.7 and 503.0 kcal across the three matched meals in 40 adults (PMID 41115473); minus 196 plus or minus 396 kcal/day intake with plus 128 plus or minus 98 kcal/day expenditure over 54 calorimeter hours (PMID 40082711); the null preload contrast of minus 24.5 kcal, 95% CI minus 87 to 38 (PMID 42633852); hazard ratios of 1.11, 1.08 to 1.14, and 1.09, 1.05 to 1.13, per 10 percentage points in 110,260 adults (PMID 32853224); and Fleiss kappa of 0.32 and 0.34 for NOVA assignment (PMID 35314769). Scoring logic was re-derived independently of the recorded letter: the five-clause parsing of the submitted sentence, the preregistered weakest-link aggregation and both reported alternatives, the cap and boundary adjudication including the recorded single-species cap-precedence resolution, the Epsilon and Zeta threshold wording, the five-step role derivation, and the Momentum, Stability and Divergence artifacts. The validator was re-run in both available builds.

**Findings.**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| E1 | Minor | Consensus provenance reconciliation block, and `consensus-addendum.md` | "The earlier adversarial approval refers to its original edition" corrected to state what is actually attached to this case: the 2026-10-06 approval covers the dated 2026-10-05 research addendum, not this review edition | The 2026-10-05 prepublication-enrichment package carries an adversarial-review banner and approval dated 2026-10-06 for this case's research addendum, with 9 of 9 identifiers checked and 0 corrections. The review edition itself was never adversarially approved, so the clause as spliced attributed an approval to the wrong artifact. The clause was inserted unconditionally by the 2026-10-09 integration script, which makes the misattribution revision-introduced. Rescoped rather than deleted, per the brief. |
| E2 | Minor | AT A GLANCE table; section 13, Momentum; score card `extended_scoring_MSD.momentum`; the red-team change note | Momentum "Decisive work underway" corrected to "Mechanistic only", with the artifact rewritten to say why the named trial does not qualify | Lead-adjudicated after this reviewer flagged the call as borderline. The framework defines the value as "Someone is running an experiment that could cross the next threshold", and both halves fail here: NCT05290064 has completed, so nothing is being run, and its posted contrast is non-significant, which the review's own Stability paragraph says would not carry SC4 across Δ. Naming a real trial satisfies the artifact-specificity requirement, which is necessary but not sufficient. A completed trial whose result does not cross belongs in the evidence narrative, not in Momentum. The original 2026-09-29 edition had "Mechanistic only"; the red-team pass moved it, and this correction returns it. Momentum modifies no letter, so the level and confidence are unchanged. |

**Rejected findings (checked, no change).**

1. Confidence Moderate. The rationale here is built from evidence density and reproducibility - trial size, duration, composition confounding, concentration in a few research networks - and quotes the framework's own band text. It does not use the reviewer's access depth, which is the confidence defect found in other cases this round. The registry-only status of [18] is a peer-review point rather than an access point, and the score card records `rests_on_non_peer_reviewed_work: false` together with the statement that the level does not rest on that record, which is correct: SC4 and SC5 sit at Gamma for want of any positive composition-matched result, not because of [18]. Re-tested against the High text and the band holds.
2. The clause "timeline rejects generated CRISPR caption" in the spliced provenance block. On its face this reads like wording imported from another case, and a stray CRISPR reference in a food review would be a serious provenance error. Checked against this case's own dated records: the 2026-10-05 enrichment files for this case carry the editorial warning that "A generated Consensus timeline caption incorrectly referred to CRISPR. It was discarded." The clause is accurate for this case. No change.
3. The "Best-supported rival" renderer defect reported elsewhere in this round. This edition's AT A GLANCE block does not carry that line, and `competing_models.models_evaluated[0]` is M2 with `is_nominal: false` and `rank_in_tournament: 1`, with the nominal at rank 4. Nothing to correct.
4. Aggregation. Weakest-link is preregistered and applied as written across five clauses, and both alternatives are reported with their derivations. Scope-matched is the value flagged for this round, so it was checked specifically: it is scored on SC4 and SC5, the subclaims whose scope matches the submitted sentence, and returns Gamma. The review also states what a narrower SC1-only claim would score, which is the right way to report a scope-restricted success.
5. The reviewer-notes contradiction flagged for the 2026-10-09 revised editions. Not present here: this review nowhere asserts that no provider work was done. Its process limitation concerns Scite, and the score card's literature-tools note covers the 2026-09-29 and 2026-09-30 passes, while the Consensus artifacts the spliced block links are dated 2026-10-05, after grading.
6. Validator result. The package documents a registry-compatibility build of validator 2.4, adopted so that an official ClinicalTrials.gov results record with a matching NCT identifier is not rejected for lacking a PMID or DOI. Under that build this edition passes with 0 fail and 1 warn, which matches the counts in the package's own `validation-report.json`. The base 2.4 build fails on exactly that registry row, and fails identically on the pre-edit original, so it is not a regression.
7. The `v5.study_identity` warning that eight ledger rows carry no `study_id`. Pre-existing and already recorded in the package's own validator report; the shared cohorts are stated explicitly in the PROVENANCE block, which is where a reader would look. Not re-litigated here.
8. The placement of the spliced provenance block between the framework heading and section 1 is structurally untidy but changes no required content, so it was left in place.

**Residual limitations.** Several records are abstract-depth by the review's own labelling, including the prescribed-intake trial [19], whose magnitudes the review deliberately declines to use. Item [18] remains registry-posted rather than peer reviewed, so its interval and arm contrasts could not be checked against a journal article; this pass confirmed only that the review reports them as posted and unpublished. Body-level figures that the abstracts do not carry - the offered and consumed energy densities in [1], the added-sugar, fibre, saturated-fat and n-6:n-3 fractions, the corrected UPDATE energy densities, and the macronutrient splits in [4] and [6] - were checked for internal consistency across their appearances but not against the full texts. No independent blinded calibration or scientific peer review is claimed by this adversarial pass.

**Approval:** Following adversarial review and the corrections listed above, the revised document is approved. - Claude Opus 5.5 (Anthropic), 2026-10-14.

## REFERENCES

1. Hall KD, Ayuketah A, Brychta R, et al. Ultra-Processed Diets Cause Excess Calorie Intake and Weight Gain: An Inpatient Randomized Controlled Trial of Ad Libitum Food Intake. Cell Metabolism. 2019;30(1):67-77.e3. · [PMID 31105044](https://pubmed.ncbi.nlm.nih.gov/31105044/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC7946062/) · [DOI](https://doi.org/10.1016/j.cmet.2019.05.008) · Errata: PMID 31269427; PMID 33027677
2. Hamano S, Sawada M, Aihara M, et al. Ultra-processed foods cause weight gain and increased energy intake associated with reduced chewing frequency: A randomized, open-label, crossover study. Diabetes, Obesity and Metabolism. 2024;26(11):5431-5443. · [PMID 39267249](https://pubmed.ncbi.nlm.nih.gov/39267249/) · [DOI](https://doi.org/10.1111/dom.15922)
3. Dicken SJ, Jassil FC, Brown A, et al. Ultraprocessed or minimally processed diets following healthy dietary guidelines on weight and cardiometabolic health: a randomized, crossover trial. Nature Medicine. 2025;31(10):3297-3308. · [PMID 40760353](https://pubmed.ncbi.nlm.nih.gov/40760353/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC12532614/) · [DOI](https://doi.org/10.1038/s41591-025-03842-0) · Supplementary correction (energy density, 2025-12-15)
4. Lasschuijt M, Camps G, Mars M, et al. Speed limits: the effects of industrial food processing and food texture on daily energy intake and eating behaviour in healthy adults. European Journal of Nutrition. 2023;62(7):2949-2962. · [PMID 37452167](https://pubmed.ncbi.nlm.nih.gov/37452167/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC10469122/) · [DOI](https://doi.org/10.1007/s00394-023-03202-z)
5. Teo PS, Lim AJ, Goh AT, et al. Texture-based differences in eating rate influence energy intake for minimally processed and ultra-processed meals. American Journal of Clinical Nutrition. 2022;116(1):244-254. · [PMID 35285882](https://pubmed.ncbi.nlm.nih.gov/35285882/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC9257473/) · [DOI](https://doi.org/10.1093/ajcn/nqac068)
6. Forde CG, Heuven LA, van Bruinessen M, et al. Eating rate has sustained effects on energy intake from ultraprocessed diets: a 2-week ad libitum dietary randomized controlled crossover trial. American Journal of Clinical Nutrition. 2026;123(4):101122. · [PMID 41314613](https://pubmed.ncbi.nlm.nih.gov/41314613/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC13084570/) · [DOI](https://doi.org/10.1016/j.ajcnut.2025.11.012)
7. van Bruinessen M, Liu Z, Heuven LAJ, et al. Satiety responsiveness following dietary exposure to ultra-processed foods differing in texture-derived eating rate: a secondary analysis of a randomized controlled crossover diet intervention. European Journal of Nutrition. 2026;65(7). · [PMID 42782361](https://pubmed.ncbi.nlm.nih.gov/42782361/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC13612719/) · [DOI](https://doi.org/10.1007/s00394-026-04111-7)
8. Rego MLM, Leslie E, Schmall E, et al. The Influence of Ultraprocessed Food Consumption on Energy Intake in Emerging Adulthood: A Controlled Feeding Trial. Obesity. 2026;34(2):344-356. · [PMID 41255123](https://pubmed.ncbi.nlm.nih.gov/41255123/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC12850616/) · [DOI](https://doi.org/10.1002/oby.70086)
9. Galdino-Silva MB, Almeida KMM, Oliveira ADS, et al. A Meal with Ultra-Processed Foods Leads to a Faster Rate of Intake and to a Lesser Decrease in the Capacity to Eat When Compared to a Similar, Matched Meal Without Ultra-Processed Foods. Nutrients. 2024;16(24):4398. · [PMID 39771019](https://pubmed.ncbi.nlm.nih.gov/39771019/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC11676177/) · [DOI](https://doi.org/10.3390/nu16244398)
10. Gama IRS, Paula DTDC, Soares DB, et al. Effects of Ultra-Processed Meals with Differing Energy Density on Postprandial Glyco-Insulinemic and Appetite-Related Hormonal Responses: A Randomized Crossover Trial. Nutrients. 2026;18(15):2534. · [PMID 42588167](https://pubmed.ncbi.nlm.nih.gov/42588167/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC13468057/) · [DOI](https://doi.org/10.3390/nu18152534)
11. Barros LM, Pereira MR, Peixoto VA, et al. Effect of ultra-processed-rich versus ultra-processed-free preload meals at high and low energy density on ad libitum energy intake: a randomized crossover trial. Appetite. 2026;228:108761. · [PMID 42633852](https://pubmed.ncbi.nlm.nih.gov/42633852/) · [DOI](https://doi.org/10.1016/j.appet.2026.108761)
12. Sciarrillo CM, Guo J, Hengist A, et al. Diet order significantly affects energy balance for diets varying in macronutrients but not ultraprocessing in crossover studies without a washout period. American Journal of Clinical Nutrition. 2024;120(4):953-963. · [PMID 39163976](https://pubmed.ncbi.nlm.nih.gov/39163976/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC11473439/) · [DOI](https://doi.org/10.1016/j.ajcnut.2024.08.013)
13. Brunstrom JM, Schatzker M, Rogers PJ, et al. Consuming an unprocessed diet reduces energy intake: a post-hoc analysis of a randomized controlled trial reveals a role for human nutritional intelligence. American Journal of Clinical Nutrition. 2026;123(3):101183. · [PMID 41475551](https://pubmed.ncbi.nlm.nih.gov/41475551/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC12975374/) · [DOI](https://doi.org/10.1016/j.ajcnut.2025.101183)
14. Beslay M, Srour B, Méjean C, et al. Ultra-processed food intake in association with BMI change and risk of overweight and obesity: A prospective analysis of the French NutriNet-Santé cohort. PLOS Medicine. 2020;17(8):e1003256. · [PMID 32853224](https://pubmed.ncbi.nlm.nih.gov/32853224/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC7451582/) · [DOI](https://doi.org/10.1371/journal.pmed.1003256)
15. Robinson E, Jones A, Evans R, et al. A systematic review and meta-analysis of randomized controlled trials examining the effect of ultra-processed food on energy intake and weight gain. Critical Reviews in Food Science and Nutrition. 2026:1-16. · [PMID 42746678](https://pubmed.ncbi.nlm.nih.gov/42746678/) · [DOI](https://doi.org/10.1080/10408398.2026.2731838)
16. Forde CG, Mars M, de Graaf K. Ultra-Processing or Oral Processing? A Role for Energy Density and Eating Rate in Moderating Energy Intake from Processed Foods. Current Developments in Nutrition. 2020;4(3):nzaa019. · [PMID 32110771](https://pubmed.ncbi.nlm.nih.gov/32110771/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC7042610/) · [DOI](https://doi.org/10.1093/cdn/nzaa019)
17. Monteiro CA, Cannon G, Levy RB, et al. Ultra-processed foods: what they are and how to identify them. Public Health Nutrition. 2019;22(5):936-941. · [PMID 30744710](https://pubmed.ncbi.nlm.nih.gov/30744710/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC10260459/) · [DOI](https://doi.org/10.1017/S1368980018003762)
18. National Institutes of Health Clinical Center / NIDDK. Effect of Ultra-processed Versus Unprocessed Diets on Energy Metabolism (Protocol 22DK0002). ClinicalTrials.gov NCT05290064; summary results first posted 2026-08-12. Registry-posted results; not a peer-reviewed publication. · [ClinicalTrials.gov record](https://clinicaltrials.gov/study/NCT05290064)
19. Preston JM, Iversen J, Hufnagel A, et al. Effect of ultra-processed food consumption on male reproductive and metabolic health. Cell Metabolism. 2025;37(10):1950-1960.e2. · [PMID 40882621](https://pubmed.ncbi.nlm.nih.gov/40882621/) · [DOI](https://doi.org/10.1016/j.cmet.2025.08.004) · Registry: NCT05368194
20. Fazzino TL, Courville AB, Guo J, Hall KD. Ad libitum meal energy intake is positively influenced by energy density, eating rate and hyper-palatable food across four dietary patterns. Nature Food. 2023;4(2):144-147. · [PMID 37117850](https://pubmed.ncbi.nlm.nih.gov/37117850/) · [DOI](https://doi.org/10.1038/s43016-022-00688-4)
21. Hägele FA, Herpich C, Koop J, et al. Short-term effects of high-protein, lower-carbohydrate ultra-processed foods on human energy balance. Nature Metabolism. 2025;7(4):704-713. · [PMID 40082711](https://pubmed.ncbi.nlm.nih.gov/40082711/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC12021659/) · [DOI](https://doi.org/10.1038/s42255-025-01247-4)
22. Larcom AB, Lofgren IE, Delmonico MJ, et al. Effects of processing level and nutritional quality on energy intake, ingestive behaviors, palatability, and satiety in a young adult population. Appetite. 2026;217:108352. · [PMID 41115473](https://pubmed.ncbi.nlm.nih.gov/41115473/) · [DOI](https://doi.org/10.1016/j.appet.2025.108352)
23. Lasschuijt MP, Heuven LA, Gonzalez-Estanol K, et al. The Effect of Energy Density and Eating Rate on Ad Libitum Energy Intake in Healthy Adults—A Randomized Controlled Study. Journal of Nutrition. 2025;155(8):2602-2610. · [PMID 40516651](https://pubmed.ncbi.nlm.nih.gov/40516651/) · [DOI](https://doi.org/10.1016/j.tjnut.2025.06.006)
24. Teo PS, Lim AJ, Goh AT, et al. Impact of food texture and degree of food processing on post-meal satiety and later snack intake. Appetite. 2026;225:108604. · [PMID 42176807](https://pubmed.ncbi.nlm.nih.gov/42176807/) · [DOI](https://doi.org/10.1016/j.appet.2026.108604)
25. Macena ML, Pereira MR, Silva DR, et al. Effectiveness and metabolic impacts of restricting the consumption of ultra-processed foods in individuals with obesity submitted to energy restriction: a randomized clinical trial. Nutrition, Metabolism and Cardiovascular Diseases. 2026;36(5):104426. · [PMID 41381310](https://pubmed.ncbi.nlm.nih.gov/41381310/) · [DOI](https://doi.org/10.1016/j.numecd.2025.104426)
26. Darcey VL, Guo J, Chi M, et al. Brain dopamine responses to ultra-processed milkshakes are highly variable and not significantly related to adiposity in humans. Cell Metabolism. 2025;37(3):616-628.e5. · [PMID 40043691](https://pubmed.ncbi.nlm.nih.gov/40043691/) · [DOI](https://doi.org/10.1016/j.cmet.2025.02.002)
27. Braesco V, Souchon I, Sauvant P, et al. Ultra-processed foods: how functional is the NOVA system? European Journal of Clinical Nutrition. 2022;76(9):1245-1253. · [PMID 35314769](https://pubmed.ncbi.nlm.nih.gov/35314769/) · [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC9436773/) · [DOI](https://doi.org/10.1038/s41430-022-01099-1)
28. Monteiro CA, Louzada MLC, Steele-Martinez E, et al. Ultra-processed foods and human health: the main thesis and the evidence. Lancet. 2025;406(10520):2667-2684. · [PMID 41270766](https://pubmed.ncbi.nlm.nih.gov/41270766/) · [DOI](https://doi.org/10.1016/S0140-6736(25)01565-X)
29. Ludwig DS. Is the experimental evidence on ultra-processed food and obesity reliable? Critical Reviews in Food Science and Nutrition. 2026;66(28):5791-5795. · [PMID 42063358](https://pubmed.ncbi.nlm.nih.gov/42063358/) · [DOI](https://doi.org/10.1080/10408398.2026.2664043)