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.

Mechanisms beyond calorie density

What drives overeating and metabolic harm from food composition and structure, once you accept the settled finding in /research/health/rubric/ that energy density is the dominant lever and that “ultra-processed” is mostly a proxy for it.

Built: 2026-09-06. Method: primary-source pass over Europe PMC, ClinicalTrials.gov API and publisher full text. Nothing here is re-derived from the rubric.

Verification marks. [V] = I read the abstract, full text or registry record directly in this session and the numbers below are transcribed from it. [S] = the number comes from a search-engine synthesis of the source, not from the source itself; treat as a lead. [R] = recalled, not checked. Peer-review status is stated per item; two of the most load-bearing results are not peer reviewed.

One correction to the framing that prompted this document. There is no “2019 vs 2026 Hall discrepancy” in the sense of two conflicting effect sizes. The +508 kcal/d is the 2019 trial; the four-arm ladder (NCT05290064) is a decomposition, and it is consistent with 2019 once you notice that the 2019 diets were not matched on non-beverage energy density. The consumed non-beverage energy density in 2019 was 1.96 kcal/g on the ultra-processed arm and 1.08 kcal/g on the unprocessed arm. That single number dissolves most of the puzzle. What remains unexplained is smaller and more interesting, and it is the subject of section 3.

Attribution. The density explanation is not a 2026 discovery. Hall 2019 itself reported that the ultra-processed meals were about 85% higher in non-beverage energy density and stated that this likely contributed to the intake difference. Brunstrom 2026 is confirmatory and quantitative, not the origin of the idea: what it adds is the per-participant consumed densities, the 726 g/d mass gap, and the micronutrient-deleveraging reading. Wherever this document credits Brunstrom with the attribution, read it as credit for the measurement, not the hypothesis.

1. What is established

1a. Food matrix and calorie availability: labels are wrong, and the direction is predictable

The claim “a calorie is a calorie” is true thermodynamically and false at the level of the food label, because Atwater factors embed digestibility coefficients calibrated on mixed, largely refined diets. Every controlled-feeding metabolizable-energy (ME) study that has looked at an intact plant matrix has found the label too high, and the size of the error tracks how thoroughly the cell wall is broken.

Food Design Measured vs Atwater Tier
Almonds, four forms n=18, 5-period randomized crossover, 42 g/d, 9 d adaptation + 9 d total collection whole natural 4.42, whole roasted 4.86, chopped 5.04, almond butter 6.53 kcal/g. Whole/roasted/chopped all below Atwater prediction (P<0.001); butter not different (P=0.08) RCT [V]
Cashews n=18, randomized crossover, 42 g/d for 4 wk, final-week total collection 137 kcal per 28 g serving, 16% below label (P<0.0001). Energy digestibility 92.9% vs 94.9% control (P<0.0001) RCT [V]
Chickpeas n=18, randomized crossover, 10 d adaptation + 7 d collection 123 kcal/85.5 g serving, 10.4% below Atwater General (P=0.002) RCT [V]
Lentils same trial 119 kcal/98.5 g serving, 16.0% below Atwater General (P<0.0001) RCT [V]
  • Gebauer SK, Novotny JA, Bornhorst GM, Baer DJ. Food Funct 2016;7:4231. https://doi.org/10.1039/c6fo01076h (PMID 27713968)
  • Baer DJ, Novotny JA. Nutrients 2018;11:33. https://doi.org/10.3390/nu11010033 (PMID 30586843)
  • Novotny JA, Henderson T, Baer DJ. Nutrients 2025;17:2725. https://doi.org/10.3390/nu17172725 (PMID 40944115)

The mechanistic explanation is fracture mechanics, and it is measured rather than assumed in the almond paper: whole natural almonds are harder (345 N) than whole roasted (298 N, P<0.05) and fracture into fewer, larger particles, so lipid stays inside intact cells and passes through. Roasting softens the nut and raises its usable calories by about 0.44 kcal/g. [V, mechanistic + RCT]

The same lever is under voluntary control at the point of eating. In a randomized 3-arm crossover (n=13) where 55 g of almonds were chewed exactly 10, 25 or 40 times, fecal fat excretion was higher after 10 chews than after 25 or 40 (both P<0.05), and every participant had higher fecal energy loss after 10 and 25 chews than after 40 (P<0.005). Early postprandial GLP-1 was lower after 25 chews than after 40 (P<0.05). Cassady BA et al. AJCN 2009;89:794. https://doi.org/10.3945/ajcn.2008.26669 (PMID 19144727) [V, RCT, small n]

Scaled to a whole diet, the matrix effect is real but modest:

  • Whole vs refined grain, n=81, 6-wk randomized parallel controlled feeding (WG 207 g/d and 40 g fiber/d vs RG 0 g and 21 g fiber/d, weight held stable): stool energy +57 kcal/d (P=0.003), resting metabolic rate +43 kcal/d (P=0.04), net 92 kcal/d (95% CI 28, 156; P=0.005). Karl JP et al. AJCN 2017;105:589. https://doi.org/10.3945/ajcn.116.139683 (PMID 28179223) [V, RCT]
  • Microbiome Enhancer Diet, metabolic-ward randomized crossover (NCT02939703): +116 ± 56 kcal/d lost in feces (P<0.0001); host metabolizable energy 89.5% vs 95.4% on the Western control (P<0.0001), with no change in energy expenditure, hunger, satiety or food intake (all P>0.05). Corbin KD et al. Nat Commun 2023;14:3161. https://doi.org/10.1038/s41467-023-38778-x (PMID 37258525) [V, RCT]
  • Dairy calcium, n=11, 7-d randomized crossover, +1600 mg Ca/d from low-fat dairy: fecal fat 5.4 to 11.5 g/d (P<0.001). The roughly 55 kcal/d of extra fecal energy quoted here is author-derived (the 6.1 g/d fat increment at 9 kcal/g), not a figure reported in the paper. Bile acid excretion unchanged, so the mechanism is calcium soap formation, not bile binding. Bendsen NT et al. Int J Obes 2008;32:1816. https://doi.org/10.1038/ijo.2008.173 (PMID 18838979) [V, RCT, very small n]

Bottom line on matrix. Label error is 10 to 30% for intact nuts and pulses eaten as foods, and roughly 90 to 120 kcal/d at the level of a whole diet shifted toward intact, fermentable matrices. That is real but it is roughly one fifth the size of the energy density effect on intake. Matrix matters more for what the label says than for what you eat.

1b. Eating rate and oral processing: a candidate lever, confounded with composition in the only long trial

This is the most active area of the field and it is not in the rubric. It is also the area where this document originally overstated its case, so the caveat is stated before the result.

Tier: RCT, peer reviewed, 2026. 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.” AJCN 2026. https://doi.org/10.1016/j.ajcnut.2025.11.012 (PMID 41314613, NCT06113146). n=41 (21 male, age 27 ± 5, BMI 23.4 ± 1.9), single-blind block-randomized crossover, two 14-day all-UPF diets with a 2-wk washout. The two diets were matched for palatability, portion size served, total energy served, non-beverage energy density, and meal variety. Result: 369 kcal/d lower intake on the slow-eating-rate diet (95% CI 221, 517), F(1,1051)=23.98, P<0.001, with no diet by time interaction (P=0.486) across 14 days. No weight change, but fat mass fell 0.43 kg on the slow arm (P=0.0002). [V]

The texture manipulation was not composition-neutral, and this is the central caveat of the whole document. [Withdrawn on review: the earlier statement that the two diets “differed only in food texture” and that the effect was “not attributable to macronutrient intake” was wrong; the diets were matched on energy density and palatability, not on macronutrients.] Read from the full text (https://www.ebi.ac.uk/europepmc/webservices/rest/PMC13084570/fullTextXML), the slow versus fast diets as served differed in fat 22 vs 33 EN%, carbohydrate 53 vs 47 EN%, mono- and disaccharides 15 vs 19 EN%, and protein 21 vs 16 EN%; as consumed, fat was 24 vs 33% (P<0.0001). Softness and fat content are not separable in a real food supply, so texture and composition are confounded by design. The paper’s exploratory model reports eating rate as significant (P<0.0001) with fat EN% not significant (P=0.082), but that model cannot separate two factors that were varied together, and the slow arm’s +5 EN% protein would on its own predict lower intake under protein leverage (section 1d).

Consequence for everything downstream: eating rate is a candidate lever, confounded with composition in the only long trial, not a demonstrated independent lever of the same magnitude as energy density. What the trial does establish is that the 369 kcal/d is not a proxy for energy density, which was matched. Design paper: Lasschuijt MP et al. Nutr Bull 2025. https://doi.org/10.1111/nbu.70027 (PMID 40926558). [V]

Supporting layers:

  • Meta-analysis, 22 experimental studies: slower eating rate lowers concurrent energy intake, SMD 0.45 (95% CI 0.25, 0.65), P<0.0001, consistent across manipulation type, with large heterogeneity. No effect on hunger at end of meal or up to 3.5 h later. Robinson E et al. AJCN 2014;100:123. https://doi.org/10.3945/ajcn.113.081745 (PMID 24847856) [V, meta-analysis of RCTs]
  • Dose-response across 24 meals, crossover, n=15 breakfast + n=15 lunch, hedonically matched (liking P=0.44 and P=0.76): slow meals eaten 39 to 45% slower, producing 22 ± 5% less food and 13 ± 6% less energy (71 kcal per meal); a 20% reduction in eating rate yields an 11 ± 1% reduction in food intake. Br J Nutr 2024. https://doi.org/10.1017/S0007114524001478 (PMID 39279635) [V, RCT]
  • Texture beats processing in a 2x2: n=50 crossover, four ad libitum lunches (soft/hard x minimally/ultra processed), energy served matched, energy density varying only ±0.20 kcal/g. Least eaten: hard minimally processed 482.9 kcal (95% CI 431.9, 531.0). Most eaten: soft ultra-processed 789.4 kcal (95% CI 725.9, 852.8), a gap of about 300 kcal in one meal. Texture had a main effect on grams (texture x processing on mass P=0.376) but the effect on energy was moderated by processing (P=0.015). Whole-day intake tracked the test meal (Δ15%, P<0.001). Teo PS et al. AJCN 2022;116:244. https://doi.org/10.1093/ajcn/nqac068 (PMID 35285882, NCT04589221) [V, RCT]
  • Head-to-head, density vs rate, with an interaction: n=69, 2x2 plus a medium control, ad libitum sandwiches at low (1.9 kcal/g) or high (3.8 kcal/g) energy density and slow or fast eating rate. Significant ED x ER interaction on energy intake, F(1,272)=5.2, P=0.024. Slow/low = 570 kcal (95% CI 442, 698) vs fast/high = 1143 kcal (1015, 1271), a 573 kcal (50%) single-meal gap. Against the control, slow/low cut intake 394 kcal and fast/high added 179 kcal. Heuven et al. J Nutr 2025. https://doi.org/10.1016/j.tjnut.2025.06.006 (PMID 40516651, NCT05659771) [V, RCT]. Note that a critical letter exists: Pereira et al. (last author Bueno), J Nutr 2025;155(12):4556, https://doi.org/10.1016/j.tjnut.2025.10.023 (PMID 41130340). I could not read the letter (no abstract, paywalled), so treat this trial’s interaction term as contested.
  • The unifying quantity, energy intake rate (kcal/min), varies enormously across a real food supply: in a laboratory dataset of 240 foods representing the whole Dutch diet, eating rate ran from 2 g/min (rice waffle) to 641 g/min (apple juice), which is 2.5 orders of magnitude (not three), and energy intake rate from 0 kcal/min (water) to 422 kcal/min (chocolate milk), a ratio that is undefined because the floor is zero. Foods 2017;6:87. https://doi.org/10.3390/foods6100087 (PMID 28974054) [V, descriptive]

1c. What the Hall trials actually show

Hall 2019 (Cell Metab 30:67, https://doi.org/10.1016/j.cmet.2019.05.008, PMID 31105044, NCT03407053), n=20, inpatient randomized crossover, 2 wk per arm: +508 ± 106 kcal/d (p=0.0001) on the ultra-processed arm, carbohydrate +280 ± 54 and fat +230 ± 53, and protein −2 ± 12 kcal/d (p=0.85); weight +0.9 ± 0.3 kg (p=0.009). [V, RCT]

Three re-analyses of that dataset matter more than the headline:

  1. The diets were not matched on non-beverage energy density. Hall 2019 said so itself, reporting the ultra-processed meals as about 85% higher in non-beverage energy density and noting this likely contributed to the intake difference; the re-analyses below quantify that observation rather than discover it. Forde CG, Mars M, de Graaf K. Curr Dev Nutr 2020;4:nzaa019. https://doi.org/10.1093/cdn/nzaa019 (PMID 32110771): eating rate was 37 vs 30 g/min and energy intake rate 48 vs 31 kcal/min, more than 50% higher on the ultra-processed arm. Subjective satiety was equivalent between arms despite the 508 kcal/d gap. [V]

    The hormone data are fasting only, and they are not null. Fasting PYY was higher on the unprocessed arm than on the ultra-processed arm and than at baseline (p=0.047), fasting ghrelin was lower on unprocessed vs baseline (p=0.01), fasting insulin was lower after the unprocessed arm (p=0.03), and glucose was p=0.06 (https://pmc.ncbi.nlm.nih.gov/articles/PMC7946062/). Postprandial gut peptides were never measured, so the trial is silent on enteroendocrine mediation. [Withdrawn on review: the earlier reading of “negligible differences in satiety hormones” misdescribed a fasting-only dataset in which three of four markers moved in the satiety direction.] [V]

  2. The consumed energy densities. Brunstrom JM, Schatzker M, Rogers PJ, Courville AB, Hall KD, Flynn AN. AJCN 2026. https://doi.org/10.1016/j.ajcnut.2025.101183 (PMID 41475551, PMC12975374): participants consumed non-beverage components at 1.96 kcal/g (ultra-processed) vs 1.08 kcal/g (unprocessed). Non-beverage energy intake differed by 330 kcal/d (15.3%) while non-beverage mass was 726 g/d (57%) higher on the unprocessed arm, in every one of the 20 participants. [V, post hoc of an RCT]
  3. Micronutrient deleveraging, the same paper’s proposed mechanism. On the unprocessed diet, components below 1.0 kcal/g delivered 42% of micronutrients consumed; on the ultra-processed diet they delivered 5%. Total micronutrient intake was 36% higher on the unprocessed arm. Participants left 370.1 kcal of >1.0 kcal/g components uneaten on the unprocessed arm (vs 217.8 kcal of <1.0 kcal/g components), which the authors note would have been roughly enough to close the gap had it been eaten. A second construct, the carbohydrate-to-fat “blend index” (how evenly a meal’s energy splits between carbohydrate and fat), differed by 0.22 at lunch and 0.24 at dinner (both P<0.0001, d=0.71 to 0.76). A mixed model with blend index + fruit-and-vegetable score + meal type predicted intake across 1678 meals at r=0.78, mean absolute error 171.8 kcal. [V]

Hall four-arm ladder, NCT05290064, n=38 enrolled and 36 completing, 4 x 1-week inpatient diets, randomized crossover. This is not a factorial. [Withdrawn on review: calling it a four-arm factorial was wrong.] It is a nested, one-factor-at-a-time ladder: each arm removes one attribute from the arm above it (UPF high-ED high-HPF, then UPF high-ED low-HPF, then UPF low-ED low-HPF, then unprocessed low-ED low-HPF). No interaction is estimable, and no arm isolates hyperpalatability at low energy density. Results are posted to the registry and are not peer reviewed. Read directly from the ClinicalTrials.gov API on 2026-09-06 [V]:

Arm Energy intake (kcal/d, SEM 108.5) Eating rate (g/min) Palatability VAS
UPF high-ED, high-HPF 3424.3 39.21 65.24
UPF high-ED, low-HPF 3265.9 39.02 62.41
UPF low-ED, low-HPF 2604.4 39.04 66.67
Unprocessed low-ED, low-HPF 2476.0 38.43 64.18

The registry posts four pre-specified rung contrasts, not six pairwise comparisons: hyperpalatability +158.4 kcal/d (95% CI 21.9, 294.9; p=0.023); energy density +661.6 (524.9, 798.3; p<0.0001); processing per se +128.4 (−8.2, 265.0; p=0.065, not significant); total +948.4 (811.9, 1084.8; p<0.0001).

How to read the eating-rate and palatability nulls. The eating-rate contrasts are posted at p=0.99, but the arm means differ by 0.0 to 0.8 g/min with 95% CIs of roughly ±2.8 g/min (about ±7%). That is a wide null, not a demonstration of invariance. Likewise the high-HPF versus low-HPF palatability difference is 2.8 points (−3.8, 9.5) on a 0-100 VAS, so a 9-point palatability difference is not excluded. Hypothesis (unverified): identical p-values repeated across several distinct contrasts (0.99 for every eating-rate contrast, ≥0.60 for every palatability contrast) look like multiplicity-adjusted values rather than raw ones; I have not confirmed this from the registry’s analysis notes.

What “hyperpalatability” means in this design. The Fazzino HPF definition is a nutrient content rule (fat, sodium and sugar thresholds), so the HPF rung is a composition contrast at matched energy density, not a pure hedonic contrast. Reading its +158.4 kcal/d as evidence about palatability per se is a mislabelling of the manipulation.

Registry link: https://clinicaltrials.gov/study/NCT05290064

1d. Protein leverage

Tier: RCT. Gosby AK et al. PLoS One 2011;6:e25929. https://doi.org/10.1371/journal.pone.0025929 (PMID 22022472). n=22 lean adults, three 4-day in-house ad libitum periods on fixed menus of 28 foods matched for palatability, availability, variety and sensory quality at 10, 15 or 25% energy from protein. Dropping 15% to 10% raised total energy intake +12 ± 4.5% (p=0.02), mostly from savoury between-meal foods, and hunger 1 to 2 h after breakfast rose more on 10% than 25% (1.6 ± 0.4 vs 0.5 ± 0.3, p=0.005). Raising 15% to 25% did not change intake. The authors state explicitly that the compensation was insufficient to hold protein intake constant, so leverage is incomplete. [V]

Tier: RCT, mechanism. Gosby AK et al. PLoS One 2016;11:e0161003. https://doi.org/10.1371/journal.pone.0161003 (PMID 27536869), n=22: 25% to 10% protein raised intake 14% (p=0.02) with a 6-fold rise in fasting plasma FGF-21 (p<0.0001) and a 1.5-fold rise in triglycerides (p<0.0001). Ghrelin, GLP-1 and CCK were unchanged. [V]

Tier: pooled analysis of trials. Gosby AK, Conigrave AD, Raubenheimer D, Simpson SJ. Obes Rev 2014;15:183. https://doi.org/10.1111/obr.12131 (PMID 24588967). 38 published ad libitum trials spanning 8 to 54% protein: percent protein inversely associated with total energy intake (F=6.9, P<0.0001), and it did not matter whether carbohydrate (F=0, P=0.7) or fat (F=0, P=0.5) was the diluent. [V]

Tier: cross-sectional, hypothesis-consistent only. Martinez Steele E, Raubenheimer D, Simpson SJ, Baraldi LG, Monteiro CA. Public Health Nutr 2018;21:114. https://doi.org/10.1017/S1368980017001574 (PMID 29032787). NHANES 2009-2010, n=9042: mean dietary protein density falls from 18.2% to 13.3% across quintiles of ultra-processed share, while absolute protein intake stays roughly constant and total energy rises. [V]

Tier: RCT, the direct test inside a UPF matrix, and it mostly fails. Nat Metab 2025;7. https://doi.org/10.1038/s42255-025-01247-4 (PMID 40082711, NCT05337007). n=21, single-blind crossover, 54 h in a whole-room calorimeter, two ad libitum all-UPF diets matched for palatability, calories served, fat and fibre: 30% protein / 29% carbohydrate vs 13% / 46%, which is a 2.3-fold difference in protein density. Energy intake −196 ± 396 kcal/d (that is mean ± s.d. at n=21, so the standard error is about 86, and the intake reduction was P<0.05) and energy expenditure +128 ± 98 kcal/d, giving energy balance +18% vs +32%, i.e. still positive on the high-protein arm. Ghrelin lower, glucagon and PYY higher. The authors’ own conclusion: protein enrichment “did not prevent overeating”, it only improved partitioning. https://pmc.ncbi.nlm.nih.gov/articles/PMC12021659/ [V]

[Withdrawn on review: the earlier note that “the standard deviation on the intake effect is twice the effect” implied the reduction was not significant. An s.d. is not a standard error; at n=21 the SE is about 86 and the reduction was significant. Also withdrawn: describing the contrast as a “tripling” of protein, which is 2.3-fold.]

1e. Hyperpalatability

The definition. Fazzino TL, Rohde K, Sullivan DK. Obesity 2019;27:1761. https://doi.org/10.1002/oby.22639. Three nutrient clusters: >25% kcal fat and ≥0.30% sodium by weight; >20% kcal fat and >20% kcal sugar; >40% kcal carbohydrate and ≥0.20% sodium by weight. 62% of FNDDS foods qualify (4,795 of 7,757), including reduced-fat items and vegetables cooked in sauces. (PMID 31689013) [V]

It does not measure palatability. Rogers PJ, Vural Y, Flynn AN, Brunstrom JM. Appetite 2024;201:107596. https://doi.org/10.1016/j.appet.2024.107596 (PMID 38969105). 52 foods, each rated by 72 to 224 people familiar with it. No significant difference in measured palatability between hyper-palatable and non-hyper-palatable foods, nor between ultra-processed and non-ultra-processed foods (both p≥0.412). High fat/sugar/salt classification was weakly predictive (p=0.049). None of the three metrics predicted desire to eat. [V]

It nevertheless predicts intake. Two independent lines:

  • The Hall factorial’s +158.4 kcal/d (p=0.023) with palatability VAS matched (p=0.60) [V, registry, not peer reviewed].
  • Jun D, Girard JM, Martin CK, Fazzino TL. Eat Behav 2025;58:101983. https://doi.org/10.1016/j.eatbeh.2025.101983 (PMID 40288138). n=29, 345 free-living eating occasions by smartphone food photography, Bayesian multilevel model: within-person %kcal from hyper-palatable foods predicted greater total energy intake controlling for pre-meal hunger and %kcal from high-energy-dense foods (median β=0.09, 95% HDI 0.02 to 0.16). [V, observational, tiny n]
  • Bellitti JS, Fazzino TL. Appetite 2026;217:108450. https://doi.org/10.1016/j.appet.2026.108450 (PMID 41506572). N=339 online, randomized parallel: sodium-containing hyper-palatable stimuli produced steeper delay discounting (ηp²=0.19 to 0.37) and lower demand elasticity (ηp²=0.16) than matched no-sodium comparators. Behavioural economics on images, not intake. [V, weak design for the claim]

A plausible substrate. DiFeliceantonio AG et al. Cell Metab 2018;28:33. https://doi.org/10.1016/j.cmet.2018.05.018 (PMID 29909968). Auction task with fMRI: participants pay more for fat+carbohydrate foods than for equally familiar, equally liked, and equally caloric fat-only or carbohydrate-only foods, with dorsal striatum and mediodorsal thalamus recruitment. This is a reward-valuation effect that is explicitly dissociated from liking. [V, mechanistic, small imaging sample]

1f. Gut and gastric responses to the same macronutrients in different structures

Tier: RCT with MRI. Krishnasamy S et al. J Nutr 2020;150:2890. https://doi.org/10.1093/jn/nxaa191 (PMID 32805050, NCT03714464). n=18, open-label 3-way randomized crossover, isocaloric 178 kcal portions of whole apple, apple puree or apple juice, serial MRI to 270 min. Gastric emptying t50 65 ± 3.3 min (whole) vs 41 ± 2.8 (puree) vs 38 ± 2.9 (juice), P<0.0001, with puree and juice indistinguishable. Post-prandial small bowel water content AUC higher for whole vs puree (P=0.025) and juice (P=0.0004). Fullness and satiety AUC higher for whole vs juice (P=0.002 and 0.004). [V] The abstract specifies isocaloric only; the earlier description of the portions as isovolumetric is unverified and I could not confirm it from the source.

Tier: OBSERVATIONAL (a controlled meal test, but nothing randomized). [Withdrawn on review: this was previously labelled “Tier: RCT”, which is wrong.] Liu Z, van Bruinessen M, Heuven LAJ, Lasschuijt MP, Stieger M, Forde CG. Eur J Nutr 2026. https://doi.org/10.1007/s00394-026-04051-2 (PMID 42397469). n=33 split by natural eating rate into slow (n=17) and fast (n=16) eaters, all given one fixed meal with a fixed carbohydrate load. Eating rate is therefore confounded with everything that covaries with it between people. Slower eaters ate 53% slower with 1.4x more chews per gram and 91.2% longer oro-sensory exposure, and had significantly higher insulin and C-peptide responses, with insulin iAUC correlating with oro-sensory exposure time (R=0.41) and chews per bite (R=0.44).

Glucose was not identical. iAUC was +20% (0-30 min) and +55% (0-180 min) in the slower eaters. Those group differences were not significant, but eating rate as a continuous predictor had a significant main effect on glucose, F(1,47)=4.51, p=0.038. The authors attribute part of the insulin rise to oral starch hydrolysis by salivary amylase, i.e. to faster glucose delivery, not to a cephalic-phase signal. (https://www.ebi.ac.uk/europepmc/webservices/rest/PMC13331850/fullTextXML) [V]

[Withdrawn on review: the earlier reading that “total post-prandial glucose did not differ” and that the insulin rise therefore reflects a cephalic-phase response is not supported. Glucose moved in the same direction as insulin, the design is observational, and the authors offer a digestive rather than a cephalic explanation.]

Prior art. This is not a new observation. Zhu Y, Hollis JH. Br J Nutr 2013 (n=21, randomized crossover, 15 vs 40 chews per bite; PMID 23181989) found glucose, insulin and GIP all higher with more chewing, under an actual randomization.

Set against this, Hall 2019 is often read as a negative on gut signalling. It is not: fasting PYY, ghrelin and insulin all moved in the satiety direction on the unprocessed arm; postprandial gut peptides were never measured, so the trial is silent on enteroendocrine mediation. [Withdrawn on review: “Gut peptide signalling is not where the ultra-processing effect lives” presented a fasting-only dataset as a postprandial null.]

1g. Two small trials that did not detect an effect at matched energy density

  • Match energy density and the acute effect is not detected. Barros LM et al. Appetite 2026. https://doi.org/10.1016/j.appet.2026.108761 (PMID 42633852). n=19, 2x2 randomized crossover, fixed-energy preload meals at 0% vs ≥80% UPF and ≤1.2 vs ≥2.0 kcal/g, outcome = ad libitum dinner. No main or interaction effects (UPF −24.5 kcal, 95% CI −87 to 38; ED −41.6, −104 to 20.9). High energy density moved appetite VAS by 3 to 5 mm without moving intake. Note the width: the UPF contrast’s CI of −87 to +38 kcal at one dinner comfortably admits an effect of about 100 kcal in either direction. [V, RCT]
  • Match energy density over 2 weeks and it is also not detected. Obesity 2026. https://doi.org/10.1002/oby.70086 (PMID 41255123, NCT05550818). n=27 aged 18 to 25, randomized crossover, two 14-day eucaloric controlled feeding periods at 81% vs 0% UPF, matched for macronutrients, fiber, added sugar, diet quality and energy density, 99% compliance. No effect on ad libitum buffet intake in the full sample. An exploratory age interaction (P<0.001) showed an effect in 18 to 21 year olds (p=0.03, d=0.79) but not 22 to 25. [V, RCT, subgroup finding is exploratory]

[Withdrawn on review: these two trials were previously described as showing the effect “vanishes” once energy density is matched. The correct statement is that the effect was not detected in two small trials whose confidence intervals admit 100 kcal effects. Absence of detection at n=19 and n=27 is not evidence of absence.]

1h. The one trial claiming harm at matched calories

Preston JM et al. “Effect of ultra-processed food consumption on male reproductive and metabolic health.” Cell Metab 2025. https://doi.org/10.1016/j.cmet.2025.08.004 (PMID 40882621, NCT05368194). Design: 43 men, 2x2 controlled crossover, 3 weeks per arm with a 3-month washout. Unprocessed to ultra-processed raised body weight and LDL:HDL independent of caloric load, lowered GDF-15 and FSH, and shifted pollutant exposure (lower plasma lithium, a trend for higher cxMINP phthalate). This is contested on exactly the point that matters here, and by more than one party:

  • Ludwig DS, “Confusion about energy and energy density in a 3-week trial of ultra-processed food,” Cell Metab 2026. https://doi.org/10.1016/j.cmet.2025.12.001 (PMID 41500195).
  • A second correspondence by Matthiessen et al. raising the same accounting objection, and the Preston et al. reply to both. [R, citation details not verified in this session]

I was unable to read any of the letters (paywalled, no abstracts), so I can state the dispute exists but not adjudicate it. Treat the “harm independent of calories” claim as unsettled, and note that the authors have responded rather than conceded.

2. Surprising and counterintuitive findings

  1. Eating rate in grams per minute barely moved across diets that differed in intake by 948 kcal/d. In the Hall ladder trial the four arms ate at 39.21, 39.02, 39.04 and 38.43 g/min. [Withdrawn on review: the earlier claim that this was “constant to within 2%” with “every pairwise p=0.99” overstated it twice over. The registry posts four pre-specified rung contrasts, not six pairwise ones, and the eating-rate differences of 0.0 to 0.8 g/min carry 95% CIs of roughly ±2.8 g/min, so a 7% difference is not excluded.] What survives is a suggestion that people defend a mass flow rate more tightly than an energy flow rate, which would let energy density convert almost directly into kcal/min. It is a hypothesis this dataset is underpowered to establish, not the settled number I called it. Note also that g/min is demonstrably not invariant elsewhere: it differed 23% between arms in Hall 2019 and ran 30 to 65 g/min between people in Liu 2026.
  2. “Hyperpalatable” is a nutrient rule, and whether it tracks liking depends on how liking is measured. [Withdrawn on review: “hyperpalatable foods are not more palatable” was stated far more strongly than the evidence supports.] Bellitti 2026 found sodium-containing HPF stimuli scored higher on both liking and wanting (p<.001, ηp² 0.34 to 0.68), a direct contradiction. The contrary evidence is Rogers 2024, whose palatability ratings were collected online from memory of familiar foods, not from tasting them, plus the Hall ladder’s palatability match, a null whose CI spans 9 VAS points. So “not liking” rests on remembered ratings and one wide-CI registry null. What does survive: Fazzino 2023 (2,733 meals) shows HPF stays predictive after adjusting for eating rate, so it is not merely an oral-processing proxy.
  3. Protein leverage is not one-sided. [Withdrawn on review: this item asserted that intake responds to protein dilution but is flat above 15% EN. That is contradicted by prior evidence. Martens EA et al. AJCN 2013 (n=79, 12-day crossover at 5/15/30% protein) measured 7.21 MJ/d at 30% protein vs 9.62 at 15% and 9.33 at 5%, P=0.001 (PMID 23221572), a large negative slope above 15%. Weigle DS et al. 2005 (n=19, 15% to 30% protein, 12 weeks ad libitum) measured −441 ± 63 kcal/d and −4.9 kg (PMID 16002798).] Gosby 2011’s flat 15-to-25% arm is one 4-day result in n=22. The 2025 calorimeter trial raised protein 2.3-fold, not three-fold, and it did reduce intake (−196 kcal/d, P<0.05) while still ending in a +18% positive energy balance. Protein fortification is a weak fix inside a UPF matrix, not a no-op.
  4. Almond butter is honestly labelled and whole almonds are not. The measured ME of almond butter (6.53 kcal/g) matched Atwater (P=0.08) while every intact form did not. Roasting alone adds about 0.44 kcal/g of usable energy by softening the nut. Processing here does not “add calories” in any nutritional sense; it makes the calories that were always in the food accessible.
  5. Chewing count is a dose-controllable variable with measurable metabolic output. Ten versus forty chews of the same 55 g of almonds changed fecal energy loss in every participant and moved GLP-1.
  6. Slower eating raised insulin, and glucose moved with it. In n=33 natural slow versus fast eaters (observational, one fixed meal), slower eaters had higher insulin and C-peptide, correlated with oro-sensory exposure time. [Withdrawn on review: the earlier claim that glucose was identical, and that the insulin rise therefore reflects a cephalic-phase response, is wrong. Glucose iAUC was +20% (0-30 min) and +55% (0-180 min) in the slower eaters, and eating rate as a continuous predictor had a significant main effect on glucose, F(1,47)=4.51, p=0.038. The authors attribute part of the effect to oral starch hydrolysis.] Zhu & Hollis 2013 (n=21, randomized crossover, PMID 23181989) had already found glucose, insulin and GIP all higher with more chewing, so the direction is not new.
  7. [Withdrawn on review: this item framed the Restructure trial’s body-composition result as a paradox. There is no paradox. 369 kcal/d over 14 days is about 5,200 kcal, which is roughly 0.6 to 0.7 kg of fat-equivalent, against an observed DEXA fat-mass loss of 0.43 kg (CI 0.23 to 0.63). Fat-free mass (p=0.102) and body weight (p=0.316) were simply underpowered over 14 days. The arithmetic reconciles without any special mechanism.]
  8. [Withdrawn on review: “gut hormones did not explain the 508 kcal” rested on fasting-only data. In Hall 2019, fasting PYY (p=0.047), ghrelin (p=0.01) and insulin (p=0.03) all moved in the satiety direction on the unprocessed arm, with glucose at p=0.06, and postprandial gut peptides were never measured. The defensible statement is that the trial is silent on enteroendocrine mediation, not that it rules it out.]
  9. Fully matching energy density produced no detected effect in two small trials. A fixed-preload acute trial (n=19) and a 2-week eucaloric feeding trial (n=27) both failed to detect an effect once energy density was matched. [Withdrawn on review: “produces nulls, twice” implied the effect vanishes. Both trials have confidence intervals that admit 100 kcal effects; Barros’ UPF contrast is −24.5 kcal with a 95% CI of −87 to +38 kcal at a single dinner.] The residual “processing per se” term in the Hall ladder trial was +128 kcal/d at p=0.065.
  10. The unprocessed arm of Hall 2019 left 370 kcal of energy-dense food on the plate. People were not full in a volumetric sense; they were choosing low-density components while energy-dense options sat there uneaten. Brunstrom’s reading is that they were solving a micronutrient problem, and paying for it with volume.

3. Novel theory candidates

Six falsifiable hypotheses, each stated so that a specific existing dataset could kill it.

H1. Mass-rate invariance (weak form is prior art; strong sufficient-statistic form is contradicted)

[Review verdict: The weak form of this hypothesis, that people eat a roughly constant mass so that energy density converts almost directly into kcal/min, is prior art, not a new candidate. See Bell EA & Rolls BJ 1998 (PMID 9497184), Rolls BJ 2001 and 2009, Viskaal-van Dongen M et al. 2011 (PMID 21094194), de Graaf C & Kok FJ 2010, and Forde CG 2017 and 2020. The strong form, that kcal/min is a sufficient statistic and energy density and eating rate collapse to one variable, is contradicted by Fazzino TL, Courville AB, Guo J, Hall KD 2023, Nat Food (PMID 37117850), which regressed 2,733 meals from Hall’s inpatient diets and found energy density, eating rate and hyperpalatability each independently predictive. The claimed mass-rate invariance is itself weaker than stated: g/min differed 23% across arms in Hall 2019 and ran 30 to 65 g/min between people in Liu 2026, and the registry’s g/min “invariance” carries 95% CIs of roughly ±2.8 g/min, about ±7%. The refutation logic below also needs a correction: a product model kcal/min = kcal/g x g/min predicts an interaction on the raw kcal scale and additivity on the log scale, so Heuven’s raw-scale ED x ER interaction (F(1,272)=5.2, P=0.024) is consistent with H1 rather than against it; only a log-scale interaction would refute it. Confidence: weak form about 0.6, strong form about 0.15.]

Claim. Over the range of real foods, the volumetric flow rate of ingestion (g/min) is approximately conserved within an individual and largely insensitive to a food’s energy content. Energy density and texture are therefore not two mechanisms; they are two multiplicands of one quantity, kcal/min = kcal/g x g/min. Energy intake rate predicts ad libitum meal size; energy density and texture predict it only through that product.

Why it is worth proposing. The Hall factorial holds g/min constant at 38.4 to 39.2 across a 948 kcal/d intake range while energy density varies. The Restructure trial holds non-beverage energy density constant and moves g/min, getting 369 kcal/d. Forde’s re-analysis of Hall 2019 shows g/min differing 23% while kcal/min differed more than 50%, and the intake gap tracked the latter.

Confirms. A pooled regression of ΔEI on Δ(kcal/min) across the trials in section 1 falling on a single line with a slope indistinguishable across trials that manipulated density, texture or both, and a much worse fit for ΔEI on Δ(kcal/g) or Δ(g/min) alone.

Refutes. The Heuven 2025 ED x ER interaction (F(1,272)=5.2, P=0.024) is already prima facie evidence against strict sufficiency, because a pure product model predicts additivity on a log scale, not a crossover interaction. If that interaction replicates in the Restructure and Hall datasets, H1 is dead in its strong form and survives only as “kcal/min explains most of the between-trial variance.”

Existing data that could test it now. Hall 2019 and the factorial both recorded meal duration and mass (that is how g/min was computed), so kcal/min per meal is derivable. Restructure recorded video-coded eating behaviour. Heuven 2025 has individual-level 2x2 data. A three-dataset individual-participant meta-regression is feasible without new subjects.

Confidence. Weak form (mass-rate defense converting to kcal/min) about 0.6; it is real, verified and already established elsewhere. Strong form (kcal/min as sufficient statistic) about 0.15; contradicted by Fazzino 2023’s independent-predictor regression.

H2. Protein leverage asymmetry (falsified by Martens 2013 and Weigle 2005)

[Review verdict: This hypothesis is falsified by evidence already available at the time it was proposed. Martens EA et al. AJCN 2013 (n=79, 12-day crossover at 5/15/30% protein) measured 7.21 MJ/d at 30% protein vs 9.62 MJ/d at 15% and 9.33 MJ/d at 5%, P=0.001 (PMID 23221572), a large negative slope above 15%, not a flat one. Weigle DS et al. 2005 (n=19, 15% to 30% protein, 12 weeks ad libitum) measured −441 ± 63 kcal/d and −4.9 kg (PMID 16002798), again above the proposed breakpoint. Both trials show leverage above 15% energy and neither shows it below, which is the opposite asymmetry from the one claimed. Gosby 2011’s flat 15-to-25% result is a single 4-day trial in n=22. Raubenheimer D & Simpson SJ 2019, “Ten points of clarification” (PMID 31339001), already treats the nonlinearity of protein leverage, so a 15% breakpoint with a flat upper tail is not a novel candidate. The 2025 Nat Metab calorimeter trial raised protein from 13% to 30%, a 2.3-fold change, not a tripling, and intake fell 196 kcal/d (SD 396, n=21, P<0.05) while energy balance was still +18% (https://pmc.ncbi.nlm.nih.gov/articles/PMC12021659/). Confidence: about 0.05.]

Claim. The relationship between dietary protein percentage and ad libitum energy intake is piecewise, with a breakpoint near 15% energy. Below it, energy intake rises steeply as protein falls. Above it, the slope is approximately zero. The population-level relevance of protein leverage is therefore entirely about the lower tail of the protein distribution, and protein fortification of processed foods cannot recover the effect.

Why. Gosby 2011 found +12% intake going 15% to 10% and nothing going 15% to 25%, in the same n=22 within-subject design. The 2025 calorimeter trial pushed protein to 30% inside an all-UPF diet and got an intake reduction (−196 kcal/d) whose SD was twice its magnitude, and still ended in a +18% positive energy balance. Martinez Steele 2018 shows the US ultra-processed gradient spans 18.2% down to 13.3%, which straddles the proposed breakpoint.

Confirms. Refitting the Gosby 2014 dataset (38 trials, 8 to 54% protein) with a segmented regression and finding a breakpoint with a 95% CI excluding both endpoints, and an above-breakpoint slope whose CI includes zero, with better AIC than the linear fit.

Refutes. A linear model fits as well or better after accounting for study-level random effects, or the breakpoint CI spans the whole range.

Existing data. Gosby 2014’s compiled 38-trial dataset. The Nature Metabolism 2025 calorimeter data. Hall’s per-food intake records from both trials, which allow computing consumed protein percentage per meal against meal size.

Confidence. About 0.05. Falsified: Martens 2013 and Weigle 2005 both show leverage above 15% energy, the opposite region from the one this hypothesis needs to be true.

H3. Micronutrient deleveraging and protein leverage are rival second-currency accounts, and a fortification factorial separates them

[Review verdict: The proposed discriminating experiment, a blinded isocaloric micronutrient pill set against a vegetable arm, cannot discriminate between the two accounts as designed. Brunstrom’s micronutrient-deleveraging mechanism, on the component-level Hall 2019 evidence cited below, operates through learned associations between a food’s sensory properties and its postprandial nutrient consequences. It is a sensory-selection mechanism, not a post-ingestive micronutrient sensor that a swallowed, tasteless pill could satisfy. An “isocaloric pill” is a category error against this kind of mechanism, so a null result in the pill arm would not distinguish micronutrient deleveraging from the parsimonious volume account; it would only show that pills do not work, which nobody doubts. The test should instead use covert fortification of a low-micronutrient ultra-processed food, matched on taste, texture and appearance to its unfortified counterpart, and measure whether food selection or intake shifts toward the fortified version. That engages the same learned-association pathway the mechanism actually claims. Confidence: about 0.2.]

Claim. Brunstrom’s micronutrient deleveraging and Simpson’s protein leverage make the same qualitative prediction (a non-energy nutrient target forces energy intake up when the diet dilutes it) but different quantitative ones about which fortification rescues intake. If the operative currency is micronutrients, then adding a low-energy-density micronutrient source to an ultra-processed diet should collapse the intake gap, while protein fortification should not. The 2025 calorimeter trial is already a weak protein-fortification arm, and it mostly failed.

Why. In Hall 2019, components under 1.0 kcal/g supplied 42% of micronutrients on the unprocessed arm and 5% on the ultra-processed arm, and total micronutrient intake was 36% higher on the unprocessed arm. Participants left 370 kcal of energy-dense components uneaten while consuming 726 g/d more mass. The Hall factorial’s low-energy-density UPF arm reached within 128 kcal/d of the unprocessed arm (p=0.065), and low energy density in practice means adding water and vegetable mass, which co-delivers micronutrients. That arm is a partially confounded test of exactly this.

Confirms. A four-arm ad libitum trial of an ultra-processed base diet crossed with (a) 400 g/d of sub-1.0 kcal/g vegetables and (b) an isocaloric micronutrient pill matched to what those vegetables deliver. Micronutrient deleveraging predicts the pill arm reduces intake; a pure volume/density account predicts only the vegetable arm does.

Refutes. The pill arm changes nothing and the vegetable arm reproduces the whole effect. That would collapse micronutrient deleveraging back into energy density and volume, which is the parsimonious reading and, in my judgment, the more likely outcome.

Existing data. Hall 2019 component-level records already permit computing, per meal, whether participants stopped eating closer to a micronutrient target than to a volume target. The factorial’s UPF low-ED arm composition would show whether its density reduction was achieved by adding water or by adding produce.

Confidence. About 0.2, and contingent on redesigning the test as covert fortification rather than a blinded pill. High confidence that the redesigned experiment is worth running, low confidence that micronutrients survive it.

H4. Hyperpalatability acts through bite size and reinforcement, not liking, and is fully mediated by bite size

[Review verdict: The “not liking” half is contradicted by Bellitti M et al. 2026, where sodium-containing hyperpalatable-food stimuli scored higher on both liking and wanting (p<.001, ηp² 0.34 to 0.68), a direct positive relationship between hyperpalatability and liking. Rogers 2024’s contrary palatability ratings were collected online from memory of familiar foods, not from tasting them, which weakens how much weight they can carry. The Hall ladder’s HH vs HL palatability match is a null whose 95% CI runs roughly −3.8 to +9.5 points, so a 9-point palatability difference is not excluded by that data either. What does survive: Fazzino 2023 shows hyperpalatable-food status stays predictive after adjusting for eating rate, which supports a bite-size or reinforcement framing over a pure oral-processing proxy, but the bite-size mediation itself remains untested. Confidence: about 0.3.]

Claim. The Fazzino nutrient clusters raise intake by increasing bite size and reinforcing value per unit of oral exposure, with no effect on rated pleasantness. If true, the hyperpalatability effect should vanish when bite mass and kcal/min are entered as covariates, and the nutrient-cluster definition should be replaceable by a directly measured bite-size or demand-elasticity variable with better predictive power.

Why. Palatability VAS was matched at p=0.60 in the arm contrast that produced +158 kcal/d. Rogers 2024 finds the clusters do not predict measured palatability or desire to eat across 52 foods. Fazzino’s own strongest positive results are behavioural-economic (demand elasticity, delay discounting), not hedonic. DiFeliceantonio 2018 shows fat+carbohydrate combinations command higher willingness to pay at equal liking and equal calories.

Confirms. In the Hall factorial video records, bite mass differs between the high-HPF and low-HPF arms even though g/min does not (which requires bite frequency to fall as bite size rises), and adding bite mass to the model reduces the arm effect from 158 kcal/d to non-significance.

Refutes. Bite size is identical across those two arms. Then the 158 kcal/d is not an oral processing phenomenon at all and requires a post-ingestive or reinforcement-learning explanation.

Existing data. Hall factorial video annotation (eating rate was derived from meal duration and mass, so bite-level coding may or may not exist; the Restructure and Teo trials definitely have behavioural video coding). Jun 2025’s photography dataset. Fazzino’s Prolific behavioural data.

Confidence. About 0.3 overall. The “not liking” half is contradicted by Bellitti 2026; bite-size mediation is untested and is the more promising remaining part of this hypothesis.

H5. Atwater bias is matrix-dependent and diet-level, costing roughly 4 to 8% of labelled calories on an intact, fermentable-matrix diet (confirmed by Baer 1997 and Zou 2007)

[Review verdict: Confirmed by prior work, not a novel candidate. Baer DJ et al. J Nutr 1997 (n=17, nine diets, directly measured metabolizable energy; PMID 9109608) and Zou ML et al. AJCN 2007 (n=27; Atwater error up to 4% on a refined diet vs 11% on a high-fiber diet; PMID 18065582) already establish matrix- and diet-dependent Atwater bias at this magnitude, as does the Livesey/FAO 2003 net metabolizable energy framework. This item should be read as corroborating those results with the almond, cashew, chickpea, lentil and enhancer-diet data above, not as proposing something new. Confidence: about 0.85.]

Claim. Atwater’s general factors embed digestibility coefficients calibrated on mixed, largely refined diets. For a diet built from intact nuts, pulses and whole grains, true metabolizable energy is systematically below label by an amount that is predictable from the diet’s composition, not just from single-food ME studies.

Why. Component-level errors are 10 to 30% for almonds, cashews, chickpeas and lentils. Diet-level: whole vs refined grain gave a 92 kcal/d net swing; the Microbiome Enhancer Diet gave 116 ± 56 kcal/d of extra fecal loss with host metabolizable energy at 89.5% vs 95.4% on the Western control, a 5.9 percentage point spread measured by bomb calorimetry in a metabolic ward. Dairy calcium alone contributes roughly 55 kcal/d.

Confirms. A whole-diet ME study of an intact-matrix diet reproducing an 85 to 91% metabolizable fraction, and a regression of measured ME loss on grams of intact plant cell wall, resistant starch and calcium that predicts individual diets within ±30 kcal/d.

Refutes. Whole-diet ME on an intact-matrix diet comes in above 95%, i.e. indistinguishable from the Western reference. Note Corbin 2023 already makes this unlikely.

Existing data. Corbin 2023 has per-participant fecal bomb calorimetry and diet composition (NCT02939703). Karl 2017 has stool energy content per participant. The four USDA ME studies have per-food digestibility coefficients that could be composed into a diet-level predictor and tested against Corbin’s measured values out of sample.

Confidence. About 0.85, confirmed rather than novel. High that the effect exists at 4 to 6%. Medium on whether it is predictable from composition rather than dominated by between-person microbiome variation, which Corbin explicitly found to be large (individual metabolizable energy on the enhancer diet ranged 84.2 to 96.1%).

H6. Oro-sensory cephalic-phase hypothesis (premise not supported; no fat-mass paradox)

[Review verdict: The premise is not supported and there is no paradox to explain. Liu 2026 is observational (n=33, natural slow versus fast eaters at a single fixed meal), not an experimental manipulation, and its own numbers contradict the “identical glucose” premise this hypothesis rests on: glucose iAUC was +55% (0-180 min, not statistically significant) in slower eaters, and eating rate as a continuous predictor had a significant main effect on glucose, F(1,47)=4.51, p=0.038. Glucose was not flat. Zhu Y & Hollis JH 2013 (n=21, RCT, 15 vs 40 chews; glucose, insulin and GIP all higher with more chewing; PMID 23181989) already gives the ordinary, non-cephalic-phase mechanism for this direction. The fat-mass “paradox” is arithmetic, not a signal that oro-sensory exposure needs a separate pathway: 369 kcal/d over 14 days is about 5,200 kcal, roughly 0.6 to 0.7 kg of fat-equivalent, against an observed DEXA fat-mass loss of 0.43 kg (CI 0.23 to 0.63). Body weight is simply the noisier measure over 14 days, and its null result is not evidence of a dissociation between fat mass and weight. Confidence: about 0.1.]

Claim. Time in the mouth is a separate signal from nutrient arrival in the gut. Longer oro-sensory exposure raises early insulin independent of glucose load, shifting post-prandial substrate partitioning toward storage of carbohydrate and away from de novo lipogenesis, so slow eating changes body composition faster than it changes body weight.

Why. Liu 2026 (n=33): slower eaters had 91.2% longer oro-sensory exposure, identical total post-prandial glucose, and significantly higher insulin and C-peptide, with insulin iAUC correlating with oro-sensory time (R=0.41) and chews per bite (R=0.44). Restructure (n=41): 14 days of slow-texture UPF produced −0.43 kg fat mass (P=0.0002) with no body weight difference between arms. Cassady 2009: chew count moved GLP-1 in the same direction.

Confirms. A trial holding total intake and macronutrients fixed and manipulating only chew count or oral exposure duration, measuring respiratory quotient by indirect calorimetry and tracking fat mass by DXA over 2 to 4 weeks. Confirmation looks like a lower post-prandial RQ excursion and a fat-mass divergence without a weight divergence.

Refutes. Under fixed intake, oral exposure duration has no effect on RQ or fat mass. Then the Restructure fat-mass result is simply the 369 kcal/d deficit landing in fat while fluid and glycogen mask it on the scale, which is the boring and probably correct explanation.

Existing data. The Restructure trial measured metabolic markers and body composition as secondary endpoints, and Liu 2026 is from the same group, so the insulin and body composition data may already be linkable at participant level. Cassady 2009 has per-arm GLP-1 and insulin curves.

Confidence. About 0.1. The premise (identical glucose) is contradicted by Liu 2026’s own data, and the fat-mass result needs no separate mechanism beyond the 369 kcal/d deficit and DEXA’s greater precision relative to body weight over 14 days.

4. Open questions

  1. Nobody has manipulated variety inside a processed diet. The Restructure trial deliberately matched meal variety across arms, so its 369 kcal/d is variety-independent, and the sensory-specific-satiety literature (Raynor & Epstein, Psychol Bull 2001;127:325, https://doi.org/10.1037/0033-2909.127.3.325 [S]) is built on single-meal course-switching paradigms, not on multi-week diets. Whether the effective variety of a supermarket ultra-processed diet is larger or, because of shared flavour platforms, functionally smaller than a home-cooked diet is unmeasured. Raynor & Vadiveloo (Curr Obes Rep 2018;7:68, https://doi.org/10.1007/s13679-018-0298-7 [V]) note that the variety effect is moderated by energy density and that the epidemiological evidence is inconsistent, which is where the field still sits.
  2. What is the residual 128 kcal/d? The Hall factorial’s processing-per-se term was +128.4 kcal/d, p=0.065, in n=36. That is exactly the magnitude that a properly powered trial (roughly n=90 at the observed CI width) would resolve. Right now the field is treating a p=0.065 as a null.
  3. Is the ED x ER interaction real? Heuven 2025 reports it at P=0.024 in n=69, and there is a published critical letter I could not read. If it is real, “match energy density” is not a sufficient control and several nulls in section 1g are underpowered rather than informative.
  4. Do eating-rate effects compensate beyond 14 days? Restructure saw no diet-by-time interaction across 14 days (P=0.486), which is reassuring, but 14 days is short relative to the timescale on which energy intake compensates for a deficit. A 12-week texture trial does not exist.
  5. Why are adolescents different? The Obesity 2026 trial found an effect at ages 18 to 21 and none at 22 to 25, in an exploratory subgroup of n=27. Almost certainly noise, but it is the only reported effect modifier in a density-matched UPF trial.
  6. Can protein leverage and micronutrient deleveraging be told apart at all in observational data? Both predict the same NHANES gradient. Only a fortification factorial separates them.
  7. Is the Preston 2025 claim of harm at matched calories survivable? It is the only trial asserting metabolic damage independent of energy intake, and it is under formal challenge on its energy density accounting. This should be settled before anyone builds policy on it.
  8. Between-person variance in metabolizable energy is large and unexplained. Corbin 2023 found individual metabolizable energy ranging 84.2 to 96.1% on the enhancer diet versus 94.1 to 97.0% on the Western diet, explained only in part by fecal short-chain fatty acids and microbial biomass. A 12 percentage point spread is roughly 300 kcal/d at a 2500 kcal intake, which dwarfs most of the intake effects catalogued above.

5. Hardest open questions (from review)

  1. What licenses attributing the Restructure 369 kcal/d to eating rate rather than composition? The slow-eating-rate diet served 11 EN% less fat and 5 EN% more protein than the fast diet, as served. Texture and fat content are not separable in a real food supply, so they co-vary by design, not by accident, and section 1d’s own protein-leverage argument predicts lower intake from the slow arm’s higher protein share on its own. Until an independent trial varies texture while holding fat and protein fixed, the 369 kcal/d figure is a texture-plus-composition effect, not a clean texture effect.
  2. Why does this document not reconcile H1, H2 and H4 with the analysis that already did the reconciliation? Fazzino, Courville, Guo & Hall 2023 (PMID 37117850) regressed 2,733 of Hall’s own meals on energy density, eating rate, hyperpalatability and protein and found all of them independently predictive. A report proposing to reanalyze the same underlying datasets to test mass-rate invariance, protein leverage and hyperpalatability-via-bite-size should start from that regression and explain where it disagrees, rather than proposing fresh individual-participant meta-regressions that ignore it.
  3. What distinguishes “invariant,” “matched” and “not where the effect lives” from “not detected”? Two of the report’s more surprising findings, g/min invariance across the Hall ladder and a palatability-matched hyperpalatable-food effect, are null results from n=36 with confidence intervals of roughly ±7% on g/min and −4 to +9 points on palatability, and the gut-hormone conclusion rested on fasting-only measurements with no postprandial data at all. Each of these was written up as a positive claim (invariance, matching, absence of a mechanism) when the data available can only support “not detected at this sample size.” The report needs a consistent rule for when a wide-CI null is reported as a null versus as a discovery.