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.

Counterintuitive levers on meal healthiness

Curiosity-driven deep research, 2026-09-06. Companion to /research/health/rubric/, which settles what is in the food. This file asks a different question: given a fixed plate of food, what else about how it is assembled, ordered, cooked, timed and eaten actually moves an outcome?

Deliberately excluded because the rubric already settles them: additives, seed oils, MSG, the NOVA label itself, sodium magnitude, fibre as a nutrient target, energy density as the master variable.

Citation discipline (same as the rest of evidence/): [V] = source fetched or searched in this session. [R] = recalled, not re-verified, treat as a lead. [C] = computed in this session from [V] numbers. Design, n and effect size are given for every verdict. Where the honest answer is “nobody measured the endpoint you care about”, it says so.

Verdict scale

Verdict	Meaning
REAL	Randomised human evidence, effect size large enough to matter, replicated
REAL (glycemia only)	Robust postprandial effect, no intake/weight/outcome evidence
SMALL	Real but the magnitude is below the noise floor of daily eating
FOLKLORE	Popular claim that the trial literature does not support
UNKNOWN	Nobody has run the trial

Summary verdict table

Everything in Part 1, ranked within groups. “Endpoint” says what was actually measured, which is usually not the endpoint the lever is sold on.

REAL and large enough to act on

Lever	Best measured effect	Endpoint	Section
Sleep extension in short sleepers	-270 kcal/day (95% CI -393 to -147), n = 80 RCT, doubly-labelled water; 2 weeks, single centre, unreplicated, intake inferred rather than measured	energy intake	§13.1
Portion size served	+423 kcal/day for +50% portions, no habituation over 11 days; meta -235 kcal/day	energy intake	§4.5
Nut metabolizable energy vs label	almonds 4.6 vs 6.05 kcal/g, i.e. -24% against the label or equivalently Atwater over-predicts by +32% (same gap, two denominators); walnuts -21%, cashews -16%; abolished by grinding to butter	measured energy	§4.1
Fat co-ingested with vegetables	5.1x to 15.3x carotenoid AUC vs fat-free; linear to 32 g oil, no saturation	absorption	§8.3
Matrix disruption (paste, juice, cooking)	lycopene AUC 3.8x, beta-carotene AUC 2.09x	absorption	§7.4
Dressing on a first-course salad	-12% to +17% total meal energy, a 29-point swing from dressing alone	energy intake	§9.2
Uncounted free side dishes	provided energy = 245% of the entrée’s stated value	measured energy	§9.4
Distracted eating	later intake SMD 0.76 (0.45, 1.07), about 2x the immediate effect	energy intake	§11.1
Air frying vs deep frying	oil 5.63 vs 1.12 g/100 g DDM; but measurement bases disagree by 10x	fat content	§7.1
Acute exercise	relative energy deficit ES -1.35; absolute intake ES 0.14, CI crosses zero	energy intake	§13.2

REAL on glycemia only, with no demonstrated intake or weight effect

Lever	Best measured effect	What was null	Section
Whey preload	peak glucose -1.4 mmol/L (-1.9, -0.9), 16 RCTs, high GRADE certainty, dose-dependent	12-wk RCT n = 79: no change in energy intake, weight, lean or fat mass; HbA1c -0.1%	§12.3
Food order (carb last), T2D	60-min glucose -42.7 mg/dL, pooled 17 RCTs; time in range +6.2 points	between-group weight p = 0.625 and intake change p = 0.205 (Shukla 2023, n = 39 pilot, powered for neither); pooled HbA1c -0.16%	§1
Vinegar / any food acid	~20-35% postprandial glucose, only with starch	insulin and HOMA-IR pooled null; appetite effect is nausea	§3
Cooling cooked starch	-18% to -25% iAUC in healthy adults; 0% in the two best negatives	calories: ~1% of the meal; no second-meal effect from cold-stored grains	§2
Second-meal effect	-38% at 4 h (n = 7); -29% overnight (n = 10); tissue glucose uptake +30% at fixed insulin	no trial has ever followed it to a clinical endpoint	§2.5
Low-energy-dense soup/salad preload	-81 to -132 kcal (9-14%) at the test meal	1-yr RCT n = 200: soup did not beat the same energy-restricted counselled diet without it (all four arms were dieting; the designed contrast was soup vs snack)	§12.1
Early time-restricted eating (eucaloric)	insulin resistance index -36 ± 10, n = 8	mean glucose unchanged; triglycerides +57 mg/dL	§5.1

SMALL

Lever	Best measured effect	Section
Eating rate (slow down)	SMD 0.45 (0.25, 0.65), 22 studies, single meal; kcal not reported	§4.2
Chewing more per bite	-9.5% to -14.8% single-meal intake; meta has publication bias, I² = 93.4%, industry first author	§4.2
Time-restricted eating vs calorie restriction	-1.40 kg (-1.81, -1.00) with zero metabolic co-movement; NS vs CR in the calorie-matched subgroup (p = 0.295)	§5.1
Late eating, isocaloric	waketime expenditure -59.4 kcal/day; ghrelin:leptin +34%; no weight endpoint	§5.2
Evening glucose intolerance	postprandial glucose +17% at 20:00 vs 08:00, well replicated	§5.2
Viscous fibre supplement	-0.33 kg free-living, -0.81 kg with a deficit	§8.5
Capsaicin	-74 kcal/meal acute, +59 kcal/day expenditure (null in lean), -0.51 kg across 15 RCTs	§4.7
Bite size / smaller spoon	-8% intake, lean men only	§4.3
Variety within a meal	+15 g vs portion size’s +57 g in the same trial	§11.2
Vinegar and weight	-1.06 kg pooled across 10 RCTs, 8 nominally positive; mostly small and low-quality, one retracted in 2025, no sensitivity analysis without it; the manufacturer-run trial rebounded in 4 weeks	§3.3
Water preload, over 55	-74 kcal/meal; Parretti -1.2 kg, CI -2.4 to +0.07 after adjustment	§12.2
Water preload, young adults, immediately pre-meal	-23% single-meal intake (Corney 2016, n = 14, p < 0.001); null at a 30-min gap (Van Walleghen); no weight endpoint has ever been measured in young adults	§12.2
Fried food and cardiac events	MACE 1.28, but +46% to +17% on BMI adjustment; all-cause and CVD mortality null	§7.1
Well-done meat and cancer	CRC 1.19 vs red meat 1.24 in the same cohort; NHS/HPFS null; IARC “not enough data”	§7.2
Low-AGE diet	insulin sensitivity +1.3 mg/kg/min (n = 20); all inflammatory and CV markers null in the same n = 20	§7.3
Boiling and vitamin C	33% loss boiling, 0% steaming	§7.4

FOLKLORE

Claim	Why it fails	Section
Plate size changes intake	pre-registered RCT d = 0.07 (CI -76.5 to +115 kcal); d = 0.03 with portion held fixed; 12 retractions in the source literature	§4.4
More frequent meals raise metabolic rate	three independent chamber-calorimetry studies, all null; six meals made people hungrier	§6.1
Meal frequency improves body composition	meta-analytic signal collapses to a single study; USDA/DGAC 2025 grade “Not Assignable”	§6.2
Breakfast aids weight loss	pooled RCTs favour skippers by 0.44 kg; eaters consume +260 kcal/day	§5.3
Morning calories burn hotter	TEF is flat (p = 0.680) after adjusting for the circadian RMR rhythm; identical weight loss and identical doubly-labelled-water expenditure	§5.3
Cold water burns meaningful calories	failed direct replication with a working positive control; 2-3 kcal, not 24	§4.6
Vinegar suppresses appetite	the study designed to separate palatability from metabolism concluded nausea, and said so	§3.5
Cooling rice halves the calories	~1%; source is an unpublished conference abstract whose calorie figure was a projection	§2.2, §2.3
The BBC pasta result	unrandomised TV demo, n ~10; best replication -4.4% total AUC and nothing for cold pasta	§2.4
Dietary acrylamide causes cancer	1,151,189 participants, 48,175 cancers, every site null	§7.1
Air frying makes more acrylamide	the only supporting study is p = 0.789 and entirely below its own LoQ, and has a published rebuttal	§7.1
Dietary PAHs from grilling are a real risk	within EU limits, lifetime risk <10⁻⁶; the 1,400x exposure is inhaled smoke	§7.2
Short eating windows raise CV mortality	31 CV deaths, two 24-hour recalls, severe baseline imbalance; 34 researchers objected to the press release	§5.1
Table GI predicts a mixed meal’s GI	no association; measured GI tracked meal energy (R² = 0.93), of which fat was the main contributor (R² = 0.88)	§8.1

UNKNOWN, nobody ran the trial

Question	Section
Utensils vs hands and energy intake	§4.3
Food temperature and intake or satiety	§4.6
Sauce as a share of restaurant meal calories	§9
Metabolizable energy of cooled-then-reheated starch in humans	§2.3
Second-meal effect from domestic cooled starch specifically	§2.5
Whole-meal analogue of the nut chewing/energy effect	§4.1, T4
Snacking per se at fixed calories (USDA/DGAC: “Grade Not Assignable”)	§6.4

Part 1. The levers

1. Food order within a meal

Verdict: REAL for glycemia, large in type 2 diabetes, smaller and less reliable in healthy people. NOT DETECTED for energy intake or weight, in a trial set that was powered for neither. [Corrected on review: this previously read “NULL for energy intake and weight, in every design-matched randomised trial that measured them.” The trials are Tricò 2016 (n = 17, 8 wk), Shukla 2023 (n = 39, 16 wk) and Touhamy 2025 (12 d, eucaloric). None could detect 1 to 2 kg, and one could not produce a weight difference at all by design. “Not detected” is the honest word.]

This is the cleanest example in the file of a lever that is genuinely real on one endpoint and has been silently promoted to a different endpoint it does not touch.

1.1 The acute glycemic trials

Study	Design, n	Population	Effect
Shukla 2015, Diabetes Care 38(7):e98-9	within-subject crossover, n = 11	metformin-treated obese T2D	glucose -28.6% at 30 min, -36.7% at 60 min, -16.8% at 120 min. Glucose iAUC 0-120 -73% (2,001 ± 377 vs 7,545 ± 804 mg/dL·min, p = 0.001). Insulin iAUC -48.5%
Shukla 2017, BMJ Open Diab Res Care 5:e000440	randomised crossover, 3 arms, n = 16	T2D	iAUC 0-180 -53.4% vs carbohydrate-first (p < 0.001); peak -53.8%; insulin iAUC -24.8%; GLP-1 iAUC +38.4% (p = 0.019)
Shukla 2019, Diabetes Obes Metab 21(2):377-81	randomised crossover, 3 arms, n = 15	prediabetes	incremental peak -45.8% (protein+veg first) and -43.1% (veg first), both p < 0.001. Glucose iAUC -38.8% for protein+veg (p = 0.008) but veg-only -23.4%, NOT significant (p = 0.205)
Kuwata 2016, Diabetologia 59:453-61	randomised crossover, ¹³C-acetate breath test	T2D n = 12 and healthy n = 10	see below
Nishino 2018, J Nutr Sci Vitaminol 64(5):316-20	crossover, 3 orders, n = 8	healthy, age 20	rice-last lowest 120-min glucose and insulin AUC
Tricò 2019, Eur J Nutr 58:2253-61	crossover, n = 16	T2D	Parmesan + egg preload raised plasma amino acids +24% (p < 0.0001); AA rise correlated with beta-cell function (r = 0.58) and the GLP-1 gradient (r = 0.57); leucine to GLP-1 r = 0.74

All [V].

Kuwata 2016 is the one to read carefully, because it contains the healthy-subject caveat that popularisers drop.

T2D total AUC (-15 to 240 min): fish-first 2,326.6 ± 114.7 and meat-first 2,257.0 ± 82.3 vs rice-first 2,475.6 ± 87.2 (p < 0.05). That is a 6% reduction in TOTAL AUC, not incremental. Quoting Shukla’s 73% iAUC alongside Kuwata mixes two different metrics.
Healthy subjects: 1,419.8, 1,389.7 and 1,483.9. Not significant. The authors state it directly: the reduction “reached statistical significance in type 2 diabetes but not in controls, possibly due to the milder glucose elevation after rice ingestion.”
The mechanism is verified in both groups. Gastric emptying T50 was 83.2 / 82.3 min vs 29.8 min in T2D and 66.3 / 74.4 vs 32.4 min in healthy subjects, so emptying is delayed roughly 2.5x regardless of glucose status (p < 0.05). GLP-1 AUC rose in both. T50 correlated with glucose AUC (T2D r = -0.746; healthy r = -0.433).

So the mechanism is real in everyone; only the glycemic consequence scales with baseline hyperglycemia.

1.2 Free-living CGM

Touhamy S, Shukla AP, et al. Diabetes Care 2025;48(2):e15-e16. Crossover, n = 20 metformin-treated T2D, 12 days (6 carbohydrate-last, 6 carbohydrate-first), eucaloric standardised meals at 2,100-2,600 kcal/d, Dexcom G6 Pro. [V]

Controlled day: incremental peak 48.3 ± 25.8 vs 85.7 ± 25.5 mg/dL (p < 0.001); iAUC 0-180 4,053.8 vs 8,112.0 (p < 0.0001).
Free-living: time in range 84.8% vs 78.6% (p = 0.041); coefficient of variation 19.2% vs 23.0% (p = 0.001); mean glucose 146.0 vs 152.3 mg/dL, p = 0.202, NOT significant.
Weight fell 1.97 kg over 12 days in both arms, with no between-arm difference.

1.3 The critical question: does it change intake or weight?

No trial has shown that it does, and no trial has been built that could. [Corrected on review: this heading previously asserted “Every design-matched comparison is null.”] The design-matched comparisons below are n = 17, n = 39 and n = 20. Detecting a 1 kg between-arm difference at 80% power against the ~5 kg standard deviation of 3-to-6-month weight change needs on the order of 400 per arm (2 x (1.96 + 0.84)² x 5² / 1²) [C]. These trials are absence of evidence, not evidence of absence.

Tricò 2016, Nutr Diabetes 6:e226. 8-week randomised free-living trial, n = 17 T2D, carbohydrate only after protein and fat at lunch and dinner. Weight fell equally: experimental -1.9 kg (95% CI -3.4 to -0.4), control -2.0 kg (-3.6 to -0.5). Waist -2.9 vs -3.3 cm. Only the experimental arm improved HbA1c -0.3% (-0.50 to -0.02) and fasting glucose. Glycemia moved, weight did not. PMID 27548711. [V]
Shukla 2023, Nutrients 15(20):4452. Open-label RCT pilot, n = 45 randomised / 39 completed, 16 weeks, prediabetes. Between-group results, which are the ones that answer the question: HbA1c p = 0.364; body weight p = 0.625; change in energy intake p = 0.205. All NS. [Corrected on review: earlier text called -64.8 ± 594 kcal/d, p = 0.649 “the decisive number”. That is a within-arm p for the food-order group, as is the control’s -292.2 ± 506 kcal/d, p = 0.016. Comparing two within-arm p-values is not a between-group test, and the actual between-group test is p = 0.205.] For symmetry, the within-arm weight change ran toward the intervention: the food-order arm lost -3.6 ± 5.7 lb, p = 0.017, while the control did not (p = 0.102). At n = 39 this is a pilot and neither direction should be read as a result. [V] https://pmc.ncbi.nlm.nih.gov/articles/PMC10610476/
Mishra, McLaughlin & Monro 2023, Nutrients 15(14):3269, randomised crossover, n = 20 healthy adults. Glycemia improved 20-30% iAUC and “Mean incremental satiety did not differ among meal types (p = 0.65).” [Corrected on review: this is not a food-order-versus-appetite test. The manipulation exchanged kiwifruit for cereal, with an order and a 30-minute separation, so it is a fruit preload / substitution design. p = 0.65 therefore means satiety was not lost when fruit replaced cereal, which is a favourable isosatiety finding, not a null on appetite suppression by ordering.] [V] https://doi.org/10.3390/nu15143269
The 2025 CGM trial above: identical weight change in both arms.
Roe/Rolls 2012 (see §12.1 for the full design) held the food constant and varied only the order, with ad libitum intake as the outcome: timing had no significant effect on energy intake; only the amount of salad did.

The one positive weight signal is Yabe 2019, J Diabetes Complications 33(12):107450: exploratory cluster-randomised, open-label, 6 months, prediabetes, n = 42 across three clusters (11 / 18 / 13), with industry co-authors (Kao Corporation). Weight fell more in the meal-sequence cluster while glycemia effects were similar. Actual kg values were not retrievable (publisher 403). Not a basis for anything. [V for existence]

1.4 Longer-term glycemic control and the meta-analysis

Imai 2011, Asia Pac J Clin Nutr 20(2):161-8. RCT, n = 101 T2D, 24 months. HbA1c fell 8.3 to 6.8% on vegetable-before-carbohydrate versus 8.2 to 7.3% on exchange-based counselling, significantly lower at 6, 9, 12 and 24 months. Caveats: unbalanced allocation (69 vs 32), open-label, single centre, and the intervention arm also increased green vegetable intake and decreased fruit, so sequence is confounded with composition. PMID 21669583. [V]

Saldarriaga-Callejas et al. 2026, Acta Diabetol 63(3):399-411 (epub 2025-10-13), PMID 41081830. 17 RCTs, n = 389. Random effects. [V]

Outcome	Pooled effect
Postprandial glucose at 60 min	-42.73 mg/dL (95% CI -55.51 to -29.96), p < 0.01
Postprandial glucose at 120 min	-13.00 mg/dL (-21.07 to -4.94), p < 0.01
GLP-1	+8.21 pmol/L (2.34 to 14.09), p < 0.01
Gastric emptying T50	+28.14 min (16.06 to 40.23), p < 0.01
HbA1c	-0.16% (95% CI -0.31 to -0.01), p = 0.04

The authors’ own conclusion is “minimal effect on HbA1c in individuals with mild T2D.” No weight or energy-intake outcome was pooled at all.

Two caveats that should travel with this table. I² values are not in the abstract and the full text is paywalled, so heterogeneity is UNVERIFIED, and no publication-bias assessment is mentioned. And a Correction was published 2026-08-24 (PMID 42635734) whose content could not be retrieved. Treat the pooled figures as provisional.

For healthy adults specifically: Kim, Jang & Lee 2026, Clin Nutr Res 15(1):55-63, systematic review, 6 studies, n = 107, ages 20-36.7. Direction-consistent reductions, no pooled estimate, and a call for large long-term RCTs. PMID 41837403. [V]

1.5 Where the popular claim outruns the evidence

“Eating vegetables first cuts your glucose spike 73%.” That is Shukla 2015: n = 11, iAUC not peak, 120-minute window, in metformin-treated obese type 2 diabetes. The meta-analytic 60-minute figure is -42.7 mg/dL. In healthy people Kuwata found no significant glucose AUC change at all.
“Food order helps you lose weight.” Not demonstrated by any randomised design-matched trial. Two measured weight (n = 17, n = 39) and found no between-group difference; neither had the power to find one.
“It reduces appetite.” The between-group intake comparison is p = 0.205 (Shukla 2023). [Corrected on review: the “satiety p = 0.65” previously cited here is Mishra 2023, a kiwifruit-for-cereal substitution study, where p = 0.65 means satiety was preserved by the swap. It is not evidence that ordering fails to change appetite.]
“HbA1c improves meaningfully.” Pooled -0.16%, with the upper confidence bound at -0.01%.

2. Resistant starch from cooling cooked starch, and the second-meal effect

2.1 How much resistant starch does cooling actually create?

Verdict: REAL but SMALL. Typical gain is +1 to +1.5 g RS per 100 g cooked food, taking resistant starch from about 2% to about 5% of total starch. Not the 10x of the press releases.

Food	Condition	RS (g/100 g)	RS as % of total starch	Source
White rice (IR-64)	fresh cooked	0.64	2.0%	Sonia 2015
White rice	cooled 10 h at ~27 °C	1.30	4.1%	Sonia 2015
White rice	cooled 24 h at 4 °C then reheated	1.65	5.2%	Sonia 2015
White rice, best of 12 combos	long grain, rice cooker, 3 d at 4 °C	2.55	n/r	Chiu & Stewart 2013
White rice, worst of 12 combos	short grain, pressure cooker, 3 d at 4 °C	0.20	n/r	Chiu & Stewart 2013
Potato	hot, 65 °C	3.1	n/r	Raatz 2016
Potato	chilled, 4 °C	4.3	n/r	Raatz 2016
Potato	chilled 6 d then reheated	3.5	n/r	Raatz 2016

Sonia S, Witjaksono F, Ridwan R. Asia Pac J Clin Nutr 2015;24(4):620-625, PMID

[V] Chiu YT, Stewart ML. Asia Pac J Clin Nutr 2013;22(3), PMID 23945407. [V] Raatz 2016 Food Chemistry values were taken from the press release, not the paper PDF, which was paywalled on every route: treat those three numbers as UNVERIFIED at source.

Three structural findings matter more than the averages.

Reheating does not destroy the resistant starch. In Sonia the reheated sample had the highest RS of all (1.65 > 1.30 room-temperature-cooled > 0.64 fresh). Retrograded amylose is heat-stable to 117-125 °C; only retrograded amylopectin reverts at 40-60 °C. In Raatz reheating cost about 19% of the chilled RS but stayed above the never-chilled value.
Variety and cooking method matter more than cooling does. Chiu and Stewart’s 12-way factorial spanned 0.20 to 2.55 g/100 g, a 12x range, with every sample chilled for 3 days. Which rice you buy and how you cook it swamps whether you refrigerate it.
Pasta has no trustworthy measured RS delta. The two most-cited pasta chilling trials, Hodges 2019 and Robertson TM 2021, did not measure resistant starch at all; both inferred it. The one trial reporting RS in cooled pasta (Rogowicz-Frontczak 2026) gives numbers implying 36% of starch is resistant in freshly cooked pasta, 10-20x every other measurement in the field, without stating the basis. Unusable until clarified.

2.2 The coconut-oil rice study

Verdict: FOLKLORE as popularly reported.

Presented by an undergraduate at the 249th ACS National Meeting, March 2015 (College of Chemical Sciences, Colombo). A conference abstract, never peer-reviewed: Europe PMC and Crossref were both searched this session for any peer-reviewed publication of this work and none exists. [V]

What was measured: in-vitro RS content of rice cooked with coconut oil, refrigerated 12 h, oven-dried. RS across 38 Sri Lankan varieties ranged 0.30-4.65%.
The calorie claim was explicitly a projection. The release says the best variety “might reduce the calories by about 50-60 percent” and states that human studies are future work.
The claim is arithmetically incompatible with its own data. Even 4.65% to 46.5% RS cannot halve the calories of a food that is ~80% starch by dry weight, because RS still yields energy (see 2.3).
Independent peer-reviewed work with the same idea (Luang-In 2026, Foods, doi:10.3390/foods15050834) got RS 0.65 to 1.39 g/100 g DW with coconut oil: a 2.1x increase, not 10x. [V]

2.3 The calorie question: about 1%

Verdict: FOLKLORE for any meaningful calorie saving. The real number is 2.5 to 4 kcal on a 346 kcal serving of rice.

The critical input is the measured energy value of resistant starch, not the assumed one. Whole-room indirect calorimetry, n = 18, randomised double-blind crossover, 24-h chambers, test starch at 26% of daily intake: net energy of resistant starch = 2.74 ± 0.41 kcal/g (33%-fiber form) and 3.16 ± 0.27 kcal/g (56%-fiber form). Nutrients 2019, PMC6835355. [V]

That is 66-74% of the 4 kcal/g of digestible starch, and higher than the 2 kcal/g regulatory convention (FAO/WHO 1998). The measured number makes the saving smaller than the popular 4-to-2 assumption, not larger.

Arithmetic for 200 g of cooked white rice using Sonia’s own composition (173 kcal/100 g, so 346 kcal; total starch 63.2 g): [C]

Assumption	RS fresh	RS chilled+reheated	Δ RS	kcal saved	% of the meal
RS = 2 kcal/g (convention)	1.28 g	3.30 g	2.02 g	4.0	1.2%
RS = 2.74 kcal/g (measured)	1.28 g	3.30 g	2.02 g	2.5	0.7%
Best case, RS forced to Chiu max	1.28 g	5.10 g	3.82 g	4.8-7.6	1.4-2.2%

For 200 g of potato using Raatz (3.1 to 4.3 g/100 g): about 3.0 to 4.8 kcal.

Ileostomy work confirms the starch reaches the colon: Langkilde 1998, Eur J Clin Nutr 52, n = 7, 3-day controlled inpatient diet, 39 g RS/d from autoclaved retrograded high-amylose corn starch vs 5 g/d control, gave +41% ileal energy excretion (p = 0.036) and +70% dry-weight excretion (p = 0.001). PMID 9846590. [V] That is delivery, not net loss: the colon recovers most of it.

No study was found that measured metabolizable energy of cooled-then-reheated rice, potato or pasta in humans. The 2.74 kcal/g figure comes from supplemental RS2/RS4 test starches at 26% of energy intake, so applying it to domestic cooled rice is reasonable but unvalidated.

2.4 The glucose question

Verdict: REAL but heterogeneous. In healthy adults the honest range is 0% to -25%, centred near -15 to -20%. The two -60% results come from insulin-dosed type 1 diabetes and are not transferable.

Study	Food	n	Design	Effect on iAUC	p
Sonia 2015, PMID 26693746	white rice 125 g, chilled 24 h + reheated	14 analysed	randomised single-blind crossover	152 to 125 mmol·min/L, -17.8%	0.047
Millet/rice trial 2024, PMC11643557	rice and millet, 50 g CHO, 24 h at 4 °C	18 women	randomised crossover, 8 meals	cold rice -24.4%, cold millet -29.5%. Glutinous varieties formed no RS and showed no effect	0.013 / 0.012
Hodges 2019, PMC7022949	pasta 100 g dry, hot vs cold vs reheated	45	randomised crossover	reheated -4.4% total AUC; cold vs hot NS	0.003
Robertson TM 2021, Eur J Clin Nutr 75(3)	pasta + oil + sauce	10	randomised crossover	overall difference p = 0.006; fresh vs chilled/reheated p = 0.041; no percentage in the abstract, full text paywalled	0.006
Alhussain 2022, Prog Nutr 24(4)	pasta chilled overnight + reheated	8 men	randomised crossover	NULL for glucose and insulin; satiety iAUC +43% (p = 0.03)	NS
Chiu & Stewart 2013, PMID 23945407	highest-RS (2.55) vs lowest-RS (0.20) rice	21	pilot clinical trial	NULL. A 12.75x difference in measured RS produced no glycemic difference	NS
Strozyk 2022, PMC9013350	white rice, chilled + reheated	32, type 1 diabetes	crossover, Libre CGM	AUC -60%; peak 11 to 9.9 mmol/L	<0.0001
Rogowicz-Frontczak 2026, PMC13074698	pasta, chilled + reheated	32, T1D on pumps	randomised single-blind crossover	iAUC -60%	<0.0001

All [V].

Two readings that change the conclusion.

The -60% trials are confounded by fixed insulin dosing. Strozyk explicitly reports 12 vs 3 hypoglycemic episodes (p = 0.0039) on the cooled meal. Part of that “iAUC reduction” is the patient being over-insulinized relative to absorbed carbohydrate. It is not a pure food effect and does not transfer to a non-diabetic.
Chiu and Stewart is the most important negative result in this section. A 12.75-fold difference in measured resistant starch content produced no glycemic difference at n = 21. That argues the glycemic effect of cooling is not driven by RS mass at all, but by broader starch reorganisation (slowly digestible starch, granule structure). The 2024 millet trial supports exactly that reading: its strongest effect came from fresh-cooked millet with minimal RS and 64.8% slowly digestible starch.

The BBC “Trust Me I’m a Doctor” pasta experiment: FOLKLORE. A television demonstration, ~10 participants, hot on day 1, cold on day 2, reheated on day 3. No randomisation (treatment fully confounded with day order), no blinding, one measurement per condition, never published. The same lab’s properly randomised peer-reviewed follow-up (Robertson TM 2021, n = 10) reports significance without a 50% reduction, and the best-powered independent replication (Hodges 2019, n = 45) found -4.4% total AUC for reheated pasta and nothing at all for cold pasta, an order of magnitude below the broadcast claim. [V]

2.5 The second-meal effect

Verdict: REAL for the phenomenon. UNKNOWN, and leaning negative, for cooled domestic starch specifically. Glycemia only: no trial has ever followed a second-meal design to a clinical endpoint.

Origins, with a citation correction worth noting: the canonical lente-carbohydrate paper is Jenkins DJ, et al. Am J Clin Nutr 1982;35(6):1339-46, not BMJ. n = 7, crossover, lentil vs wholemeal-bread breakfast then a standard bread lunch 4 h later: first meal -71% glucose area (p < 0.001), second meal -38% (p < 0.01). A quarter-carbohydrate bread breakfast worsened the lunch response to 168% of control. Breath hydrogen in a separate n = 4 ruled out malabsorption. PMID 6282105. [V]

Wolever TM, et al. Am J Clin Nutr 1988;48(4):1041-7: low-GI dinner lowered next- morning glycemia, and fibre content at matched GI had no effect. PMID 2844076. [V] An early signal that the overnight effect needs fermentable carbohydrate specifically, not fibre generally.

Study	n	Interval	Dose	Effect
Robertson MD 2003, Diabetologia 46:659-65, PMID 12712245	10	overnight	60 g RS	insulin sensitivity +69% (44 ± 7.5 vs 26 ± 3.5), p = 0.028; NEFA and 3-hydroxybutyrate suppressed
Robertson MD 2005, AJCN 82:559-67, PMID 16155268	10	chronic 4 wk	30 g/d RS	clamp SI +14% (p = 0.03), MTT SI +33% (p = 0.05), forearm muscle glucose clearance +44%; acetate and propionate up
Nilsson AC 2008, AJCN 87:645-54, PMID 18326603	12	evening to breakfast (9.5 h)	barley/rye kernels	lower iAUC at breakfast, lunch and cumulatively (p < 0.05); breath H2 up (p < 0.001) and negatively correlated with lunch iAUC, r = -0.33
Nilsson AC 2010, J Nutr 140:1932-6, PMID 20810606	15	evening to morning	barley kernels	plasma butyrate higher next morning (p < 0.05); breakfast iAUC inversely related to butyrate r = -0.26 and acetate r = -0.20
Priebe/Vonk 2010, AJCN 91, PMID 19889821	10 men	evening to morning, dual isotope	barley kernels vs white wheat	glucose response -29% (p = 0.019) with identical insulin; tissue glucose uptake +30% (p = 0.016); IL-6 and TNF-alpha higher after the wheat evening meal
Sandberg 2016, PLOS One 11:e0151985	19	evening to breakfast	rye kernel bread, 15.6 vs 3 g fibre	breakfast iAUC lower (p = 0.001); PYY up (p < 0.001), GLP-1 up, satiety up, hunger down; no difference between 1-day and 3-day priming
MacNeil 2013, APNM 38, PMID 24195618	12 T2D	breakfast to lunch, 3 h	high-RS bagels	first meal improved; second meal: only the GIP-insulin slope changed, no significant improvement in glucose disposal
Bodinham 2013, Br J Nutr 110, PMID 23507477	30 men	breakfast to lunch	48 g RS	GLP-1 was LOWER, not higher (p = 0.025); insulin lower after lunch (p = 0.034); glucose unchanged throughout
Millet trial 2024, PMC11643557	18 women	breakfast to lunch	50 g CHO	fresh-cooked millet -39.1% second-meal iAUC (p < 0.05). Cold-stored grains showed NO second-meal effect despite higher RS

All [V].

Effect size by interval. Breakfast to lunch (3-4 h): -38% in Jenkins with lentils, but modern resistant-starch supplements at this interval are weak to null (MacNeil, Bodinham), and the 2024 -39.1% came from slowly digestible starch, not RS. Evening to next breakfast (9.5-12 h): -29% glucose response in Priebe, significant in Nilsson and Sandberg. Overnight after a single 60 g RS load: +69% insulin sensitivity, but note the dose. 60 g of RS in one day is roughly 30 to 60 times the RS you gain by cooling one serving of rice.

Mechanism, demonstrated vs hypothesised.

Demonstrated in humans, in these papers: colonic fermentation occurs and correlates (breath H2 up, and negatively correlated with subsequent-meal iAUC, r = -0.33); plasma SCFAs rise and correlate (butyrate r = -0.26, acetate r = -0.20; adipose acetate uptake measured directly by arteriovenous difference, p = 0.03); free fatty acid suppression, which is the best-supported single mechanism (breakfast glucose correlated with FFA r = +0.37, p < 0.001); and peripheral tissue glucose uptake +30% at identical insulin, which rules out “just more insulin.”

Contested: GLP-1 is contradictory. Elevated after rye evening meals and inversely correlated with breakfast glucose in Nilsson, but Bodinham gave 48 g HAM-RS2 to n = 30 and found GLP-1 significantly lower, concluding RS “does not acutely increase endogenous GLP-1 concentrations in human subjects.”

Not demonstrated at all: RS3 from domestic cooling has never been tested in a second-meal design. Every positive second-meal trial above used barley or rye kernels, lentils, or 30-60 g of supplemental RS2. The one trial that did test cold-stored grains in a second-meal design found no second-meal effect. That is a direct negative for the popular claim.

2.6 Does any of it reach a clinical endpoint?

Verdict: SMALL to NULL for weight; the HbA1c literature is internally contradictory; the second-meal effect has zero long-term data.

Meta-analysis	Trials / n	Findings
Snelson 2019, Nutrients 11:1833, PMC6723691	22 RCTs, 670 participants, ≥8 g RS2/d	Triacylglycerol -0.10 mmol/L (p = 0.03) in healthy; body weight -1.29 kg (p = 0.02) in T2DM only, n = 90. No effect on any other outcome including HbA1c, fasting glucose, insulin resistance, satiety. Authors: results “heavily influenced by positive results from a small number of individual studies which contradicted the conclusions of the majority of trials”; conclusion “of limited cardiometabolic benefit”
Halajzadeh 2020, Crit Rev Food Sci Nutr 60, PMID 31661295	19 trials	FPG -4.28 mg/dL, insulin -1.95, HbA1c -0.60% (95% CI -0.95 to -0.24) from 8 studies. Null for HOMA-IR, triglycerides, HDL, CRP, IL-6
Xiong 2021, Br J Nutr 125	19 RCTs	fasting glucose -0.09 mmol/L (p = 0.001); HOMA-IR SMD -0.33 (p = 0.001). Null for fasting insulin, insulin sensitivity index, disposition index, HOMA-beta
Nutr Diabetes 2019, PMC6551340	13 trials, 428 subjects	HbA1c SMD -0.43; fasting insulin SMD -0.72; HOMA-IR NS; significant heterogeneity flagged

All [V].

The HbA1c signal is unstable and should be discounted. Halajzadeh reports -0.60%, a drug-sized effect, from 8 trials. Snelson, working from 22 RCTs and 670 participants, reports no HbA1c effect at all. Both are peer-reviewed and they cannot both be right. HOMA-IR is null in three of the four meta-analyses. The weight result rests on n = 90 with the meta-authors themselves warning it is outlier-driven.

And the scale mismatch is the point: all of this is for 10-45 g/day of supplemental resistant starch. Cooling one serving of rice yields 1-2 g. Nothing in these meta-analyses is evidence for the cooled-rice practice.

2.7 The energy-harvest counterweight

Verdict: REAL, counterintuitive, and roughly 40-60x larger than the effect it would offset.

Jumpertz R, Le DS, Turnbaugh PJ, et al. Am J Clin Nutr 2011;94(1):58-65, NCT00414063, PMC3127503. n = 12 lean + 9 obese, inpatient, 3-day 2,400 kcal/d and 3,400 kcal/d diets with washout, bomb calorimetry on both ingested food and stool, 16S sequencing. [V]

Stool energy loss in lean subjects: 134.3 ± 48.9 kcal/d (4.9% of intake) at 2,400 kcal/d and 145.1 ± 42.7 kcal/d (3.8%) at 3,400 kcal/d. Absolute loss barely moved while intake rose 1,000 kcal, so fractional loss fell by ~1.1 percentage points during overfeeding: the gut got more efficient under caloric load.
A 20% increase in Firmicutes with a corresponding Bacteroidetes decrease was associated with ~150 kcal of increased nutrient absorption. Bacteroidetes r = 0.52 (p = 0.01), Firmicutes r = -0.50 (p = 0.02) in lean subjects; absent in obese subjects.
Caveat: the microbiome was observed, not manipulated. This is an association inside an experimental feeding study.

Scale reference: Basolo 2020, Obesity 28, PMID 33029899, bomb-calorimetry protocol with dye-marked stool: mean stool calorie loss 7.3 ± 1.6% of intake. At 2,500 kcal/d that is ~180 kcal/d. [V]

Net arithmetic [C]. Making ~2 g of RS3 by cooling 200 g of rice removes at most 2.5 to 4.0 kcal. The measured microbiome-associated swing in energy harvest is ~150 kcal/d, roughly 40 to 60 times larger and of uncertain sign. Any calorie-reduction claim for cooled rice sits well below the resolution of human energy-balance measurement.

3. Vinegar and acetic acid with meals

Verdict: REAL but modest for postprandial glucose, and only with starch. FOLKLORE for appetite suppression. OVERSTATED for HbA1c. UNSUPPORTED for insulin sensitivity. And the effect is not specific to acetic acid.

3.1 Acute glycemia

Study	Design, n	Dose	Result
Ostman 2005, Eur J Clin Nutr 59:983-8	randomised crossover, n = 12 healthy	18 / 23 / 28 mmol acetic acid with 50 g CHO white bread	dose-response at 30 min only. GI and II lowered only at the highest dose, and only on the 90-min incremental area; 120-min values did not differ from reference at any dose
Liljeberg & Björck 1998, Eur J Clin Nutr 52:368-71	crossover, n = 10 healthy	vinegar with starchy meal	GI 64, II 65 vs reference 100; paracetamol marker indicated delayed emptying
Johnston 2004, Diabetes Care 27(1):281-2 (research letter)	randomised crossover	20 g vinegar pre-meal, 87 g CHO	n = 11 insulin-resistant: PPG 64% of placebo (p = 0.014), insulin sensitivity +34% (p = 0.01). n = 10 T2D: PPG -17% (p = 0.149, NS), sensitivity +19% (p = 0.07, NS)
Johnston 2010, Ann Nutr Metab 56(1):74-9	4 randomised crossovers, n = 9-10 each	10 vs 20 g, mealtime vs 5 h prior	~20% PPG reduction; 10 g sufficient; no effect with monosaccharides (dextrose); sodium acetate had no effect
Liatis 2010, Eur J Clin Nutr 64:727-32	T2D, two matched groups, crossover within group	vinegar with high-GI vs low-GI meal	high-GI arm (n = 8): iAUC 181 ± 78 vs 311 ± 124 mmol·min/L, p = 0.04 (~-42%). Low-GI arm (n = 8): 229 ± 38 vs 238 ± 25, p = 0.56, NO effect
Mitrou 2015, J Diabetes Res 2015:175204	randomised crossover, n = 11 T2D, arteriovenous forearm balance	vinegar pre-meal	forearm glucose uptake up (p = 0.036), glucose down (p = 0.028), insulin down (p = 0.046), TG down (p = 0.044). Four p-values clustered 0.028-0.046 with no multiplicity correction

All [V]. Note: the commonly cited “Mitrou 2010” appears to be the 2015 paper; a separate 2010 vinegar paper by that group could not be located.

The starch dependency is the under-discussed part. Vinegar does nothing with monosaccharides (Johnston 2010) and nothing with a low-GI meal (Liatis, p = 0.56). It works by interfering with starch digestion, so it only helps meals that were going to spike anyway.

3.2 HbA1c and insulin sensitivity: the numbers do not hold up

The originating lab’s own direct measurement is the smallest one in the field. Johnston 2009, Diabetes Res Clin Pract, PMID 19269707: HbA1c -0.16% with vinegar, versus +0.06% (vinegar pill) and +0.22% (dill pickle), p = 0.018. [V]

The meta-analyses report 4 to 10 times more:

Meta-analysis	Studies / n	HbA1c or FBG	Heterogeneity
Shishehbor 2017, Diabetes Res Clin Pract 127:1-9	not retrieved	glucose AUC SMD -0.60 (-1.08 to -0.11); insulin AUC SMD -1.30 (-1.98 to -0.62)	I² and publication-bias tests NOT RETRIEVED. An insulin SMD of -1.30 for a condiment is the classic signature of pooling underpowered crossovers
Tehrani 2025, Curr Med Chem 32(11):2257-74	25 trials, 33 arms, n = 1,320	FBG -21.20 mg/dL; HbA1c -0.91	I² = 95.8% and 98.9%. Uninterpretable. BMI, HOMA-IR, insulin, TG, LDL, HDL all NS
Arjmandfard 2025, Front Nutr 12:1528383	7 CTs, GRADE + dose-response	FBS -21.93 mg/dL; HbA1c -1.53	An HbA1c drop of -1.53% from vinegar exceeds metformin monotherapy. And insulin went UP (+2.06 µU/mL, p = 0.025) while HOMA-IR was null, which contradicts the insulin-sensitivity story

The most careful synthesis is an umbrella review: Shahmohammadi 2026, Food Sci Nutr, PMID 42110393, 10 meta-analyses, 38 primary studies, n = 1,781, effects re-estimated. [V]

Outcome	MD (95% CI)	I²	Egger	GRADE
Fasting glucose	-9.40 mg/dL (-12.46, -6.34)	96.3%	NS	Moderate
Postprandial glucose	-14.59 mg/dL (-27.11, -2.06)	93.6%	NS	Moderate
HbA1c	-0.70% (-1.07, -0.33)	98.1%	p = 0.037, BIAS DETECTED	Moderate
Fasting insulin	+0.82 (-1.40, 3.05)	44.1%	NS	NULL
HOMA-IR	+0.13 (-0.37, 0.64)	57.5%	NS	NULL
Body weight	-1.06 kg (-1.60, -0.52)	11.1%	NS	High
BMI	-0.15 (-0.35, 0.04)	30.5%	NS	NULL
Waist	-0.91 cm (-1.92, 0.11)	61.8%	NS	NULL
Systolic BP	-2.94 mmHg (-4.81, -1.07)	0%	p = 0.026, BIAS	High

And the quality audit from that same review: 32 of 38 primary trials rated LOW quality on Cochrane risk of bias, with inadequate blinding in over half. Five of the ten meta-analyses are AMSTAR-2 LOW and two are CRITICALLY LOW. Overlap between the meta-analyses is not quantified, so the same handful of small crossovers is being counted repeatedly.

“Vinegar improves insulin sensitivity” is unsupported. Pooled fasting insulin and HOMA-IR are both null across 8-9 trials and ~670-690 people. The claim traces back to n = 11 in a 2004 research letter whose type 2 diabetes arm missed significance on both endpoints.

3.3 The weight result and where it comes from

The one clean-looking pooled estimate is -1.06 kg (I² = 11.1%, GRADE High).

[Corrected on review: this file previously stated that the pooled weight effect “is carried by a single trial.” That was asserted without a source and is very likely wrong. A 2025 meta-analysis pools 10 RCTs of vinegar on body weight, of which 8 report a significant reduction, so the pooled estimate rests on many small trials, not one. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12472926/ The real problem with that pool is quality rather than count: the constituent trials are small and mostly low quality, one of them (Abou-Khalil 2024) was retracted on 2025-09-23 after the journal found data inconsistent with the stated randomisation, and no sensitivity analysis excluding it has been published. https://retractionwatch.com/2025/09/23/apple-cider-vinegar-weight-loss-study-retracted-bmj/ Kondo below is the manufacturer-run trial in this literature, not the only positive one.]

Kondo T, et al. Biosci Biotechnol Biochem 2009;73(8):1837-43. Parallel-group, randomised, double-blind, placebo-controlled, 12 weeks plus 4-week washout. 175 randomised, 155 analysed. Obese Japanese adults, BMI 25-30. 500 mL/day of a beverage containing 0, 15 or 30 mL apple vinegar (0 / 750 / 1,500 mg acetic acid). [V, PDF read at source]

Body weight: placebo 74.2 to 74.6 kg (+0.4); low dose 74.9 to 73.7 (-1.2); high dose 73.1 to 71.2 (-1.9). Visceral fat area change: +4.4%, -2.8%, -4.9%.
All authors are employees of the Central Research Institute of the Mizkan Group, Japan’s largest vinegar manufacturer. Ethics review was conducted in part by Mizkan’s own committee.
No confidence intervals appear anywhere in the paper, only SDs and Dunnett p-values.
Energy intake and steps per day were statistically unchanged in all arms. 1.9 kg over 12 weeks with no measured change in intake or activity implies a ~165 kcal/day deficit from nowhere. That is not energetically coherent.
The weight returned within 4 weeks of stopping: low dose rebounded fully to 74.9, high dose regained to 72.7.
The placebo contained 1,250 mg of lactate to mimic taste, so the trial is simultaneously an acetic-versus-lactic-acid comparison.

3.4 It is not the acetic acid

Verdict: the effect belongs to food acids generally, and the acetate ion is inert.

Liljeberg & Björck 1996, Am J Clin Nutr 64:886-93: bread with added lactic acid or sodium propionate both lowered glucose and insulin. Propionate also delayed gastric emptying; lactic acid lowered glycemia with no emptying change. A hydrochloric acid solution of similar pH was much less effective in rats, so it is not pH alone either. PMID 8942413. [V]
Johnston 2010: sodium acetate did not alter postprandial glucose. The salt is inert; the proton matters.
Brighenti 1995, Eur J Clin Nutr 49:242-7, n = 5: vinegar versus the same vinegar neutralised to pH 6.0 with bicarbonate. Glucose response -31.4% (p = 0.023) for the acid form, with no difference in gastric emptying by ultrasonography. PMID 7796781.
Freitas 2021, Eur J Nutr, PMID 32201919: lemon juice with bread lowered the glucose peak 30% (p < 0.01) and delayed it by more than 35 min. No effect on ad libitum energy intake.
Freitas 2022, Eur J Nutr, PMID 35013789, n = 10, MRI gastric volumetry: lemon juice lowered glucose at 55 min by 35% (p = 0.039) while gastric emptying was 1.5x FASTER. Proposed mechanism: acid inhibition of salivary alpha-amylase. This undercuts the delayed-emptying story that dominates vinegar popularisations.
Pickled foods do not carry over: Johnston 2009 found dill pickles raised HbA1c +0.22% while liquid vinegar lowered it -0.16%.

No single-protocol head-to-head of acetic vs lactic vs citric acid exists in the reachable literature. The comparison above is assembled across labs and decades.

3.5 The satiety effect is nausea

Verdict: FOLKLORE, and the authors of the decisive study said so.

Darzi J, Frost GS, Montaser R, Yap J, Robertson MD. Int J Obes 2014;38(5):675-81. Two acute randomised crossovers in healthy young unrestrained eaters. Study 1 (n = 16): vinegar in a palatable versus an unpalatable drink versus control. Study 2 (n = 14): modified sham feeding, taste only, no swallowing. PMID 23979220. [V]

Vinegar “significantly reduced quantitative and subjective measures of appetite, which were accompanied by significantly higher nausea ratings, with unpalatable treatment having the greatest effect.” Orosensory stimulation alone had no effect. The authors’ conclusion, verbatim: “these effects are largely due to poor tolerability following ingestion invoking feelings of nausea. On this basis the promotion of vinegar as a natural appetite suppressant does not seem appropriate.”

Supporting: Johnston 2008, J Med Food, n = 27, 12 weeks: 50-56% of the vinegar and pickle arms reported at least one treatment-emergent adverse event versus 11% on a near-zero-dose pill (p = 0.110). And Johnston 2005, J Am Diet Assoc, n = 11: vinegar cut the 60-minute glucose response ~55%, but later energy intake fell only ~200-275 kcal at p = 0.111 (NS), and the glucose response explained only 11-16% of the variance in later intake.

Worth stating plainly: the glycemic effect and the appetite effect are not the same mechanism, and only one of them survives scrutiny.

4. Eating mechanics: chewing, rate, tableware, temperature, spice

This cluster contains both the largest unrecognised effect in the file and its most thoroughly debunked claim.

4.1 Nut metabolizable energy: the biggest under-appreciated number here

Verdict: REAL, large, mechanistically explained, replicated across four nut types. And it is destroyed by grinding.

Controlled full-feeding crossover studies with complete urine and faecal collection and bomb calorimetry, all from USDA Beltsville:

Nut	Study	n	Measured metabolizable energy	Atwater predicted	Gap
Almonds	Novotny 2012, AJCN 96(2):296-301	18, 18-day periods	4.6 ± 0.8 kcal/g = 129 kcal/28 g	6.0-6.1 kcal/g = 168-170 kcal	-24% vs label
Walnuts	Baer 2016, J Nutr 146(1):9-13	18, 3-week periods	5.22 kcal/g = 146 kcal/28 g	6.61 kcal/g = 185 kcal	-21%, p < 0.0001
Cashews	Baer 2019, Nutrients 11(1):33	18, 4-week periods	137 ± 3.4 kcal/28 g	163 kcal	-16%, p < 0.0001
Pistachios	Baer 2012, Br J Nutr 107(1):120-5	16, 3-week periods	22.6 kJ/g	23.7 kJ/g	-5%

All [V].

State the denominator, because two different numbers get quoted for the same gap. [Added on review.] Almonds measure 4.6 kcal/g against an Atwater-predicted 6.05 kcal/g. Against the label denominator that is -24%. Against the measured denominator, which is how Novotny herself frames it, Atwater over-predicts by 32% (6.05 / 4.6 = 1.32). “-24%” and “+32%” are the same finding, not two. https://pubmed.ncbi.nlm.nih.gov/22760558/

Processing abolishes it. Gebauer 2016, Food Funct 7(10):4231-8, n = 18, five-period crossover, 42 g/d: whole natural almonds 4.42 kcal/g, whole roasted 4.86 kcal/g (p < 0.05 vs natural), almond butter 6.53 kcal/g, statistically indistinguishable from the Atwater prediction (p = 0.08). Mechanism measured directly: roasted almonds fracture at lower force (298 ± 1.3 N vs 345 ± 1.6 N), releasing more lipid from intact cell walls. [V]

And the chewing link is measured: Cassady 2009, AJCN 89(3):794-800, n = 13, randomised crossover plus 4-day controlled feeding with complete faecal collection, 55 g almonds chewed 10 / 25 / 40 times: faecal fat excretion significantly higher at 10 chews than at 25 or 40 (p < 0.05); every participant lost more faecal energy at 10 and 25 than at 40 chews (p < 0.005). NCT00768417. [V]

So the calorie content of a nut is not a property of the nut. It is a property of the nut and your molars jointly. 28 g of whole almonds is ~129 kcal, not the 170 on the label. The same almond as butter is the full label value. This is a 40 kcal per serving error in a food eaten daily by many people, and it is one of the very few places where a nutrition label is systematically and knowably wrong.

The single weakness: essentially all human nut-ME data comes from one lab. No independent replication was established.

4.2 Chewing count and eating rate

Verdict: REAL but small, acute only, and the meta-analysis has publication bias.

Study	Design, n	Manipulation	Effect
Li 2011, AJCN 94(3):709-16	randomised crossover, n = 30	15 vs 40 chews per 10 g bite	-11.9% intake; lower ghrelin, higher GLP-1 and CCK
Zhu & Hollis 2014, J Acad Nutr Diet 114(6):926-31	randomised crossover, n = 45	100 / 150 / 200% of baseline chew count	-9.5% at 150%, -14.8% at 200%. No difference in subjective appetite at meal end or over 60 min

[V] both. Two independent labs, different foods, consistent direction.

But: Miquel-Kergoat 2015, Physiol Behav 151:88-96, meta-analysis of 15 papers / 17 trials: only 10 of 16 experiments found chewing reduced intake and only 5 of 16 found a satiety effect. Evidence of publication bias, I² = 93.4%. Pooled hunger effect -2.31 VAS points. The first author is employed by Wrigley (Mars Inc.), a chewing-gum manufacturer. Treat the pooled estimate as low-confidence. [V]

Eating rate as a separate lever: Robinson E 2014, AJCN 100(1):123-51, 22 experimental studies, random-effects SMD 0.45 (95% CI 0.25 to 0.65), p < 0.0001, with no significant effect on hunger at meal end or up to 3.5 h later. The pooled effect in kcal is not in the abstract and the full text was inaccessible, so any kcal conversion is UNVERIFIED. [V]

Note the recurring mechanism signature: intake falls without hunger rising. That is the same pattern as the chewing trials and it is what makes these levers different from restriction.

Cohort evidence, and why the headline misleads: Hurst & Fukuda 2018, BMJ Open 8(1):e019589, n = 59,717 Japanese T2D patients from insurance claims and health checkups. Slow eaters had OR 0.58 for obesity versus fast eaters (p < 0.001). But the fixed-effects BMI coefficient is -0.11 BMI units, which is about 300 g of body weight for a 1.7 m adult. Quoting OR 0.58 while omitting the -0.11 BMI effect is the single most misleading move in this literature. Exposure is one self-reported categorical question. [V] Zhu 2015, J Epidemiol 25(4):332-6, n = 8,941, 3 years: fast eating and incident metabolic syndrome HR 1.30 (1.05-1.60). [V]

A relevant bonus: Hamano 2024, Diabetes Obes Metab 26(11):5431-43, randomised open-label crossover, ultra-processed vs non-ultra-processed meals matched for total energy and macronutrients, 1 week each, n = 9 overweight Japanese men. The UPF period produced +1.1 kg body weight (95% CI 0.2 to 2.0, p = 0.021) and +813.5 kcal/day (95% CI 342.4 to 1284.7, p = 0.0041), with significantly fewer chews per calorie (p = 0.016). n = 9, open label, single site: hypothesis generating, but it is the only trial linking the oral-processing mechanism to a weight endpoint. [V]

4.3 Utensils versus hands

Verdict: UNKNOWN for intake. One n = 11 study exists and it measured glycemia.

Sun 2015, Physiol Behav 139:505-10, n = 11, white rice eaten with spoon vs chopsticks vs fingers, with EMG mastication measurement. Chopsticks produced more chews per mouthful, longer chewing time and a lower glycemic index (68 vs 81 for spoon). Fingers did not differ from either. Energy intake was not an outcome. [V] Single site, unreplicated.

No trial measuring ad libitum intake by eating implement was located. This is a genuine gap, and given the eating-rate literature it is a cheap experiment that nobody appears to have run.

The adjacent lever that is supported is bite size: James 2018, Br J Nutr 120(7):830-7, two studies in lean young men, teaspoon vs dessert spoon: -8% ad libitum intake (532 ± 189 vs 575 ± 227 g, p = 0.006), bite size 10.5 vs 13.7 g, eating rate 92 vs 108 g/min. No effect on subjective appetite. Small, male-only, lean-only. [V]

4.4 Plate size

Verdict: FOLKLORE, with high confidence in the null. Portion size is the real lever and it is being smuggled in under the plate.

Wansink retraction accounting, measured this session. A PubMed query for Wansink B[au] AND Retracted Publication[pt] returns 12 records, including “Super Bowls: serving bowl size and food consumption” (JAMA 2005), “Consequences of belonging to the clean plate club” (Arch Pediatr Adolesc Med 2008) and “The joy of cooking too much” (Ann Intern Med 2009). Retraction Watch reported 13 retractions plus at least 15 corrections as of September 2018; press reporting since gives 18 total across all venues. Cornell’s committee found “misreporting of research data, problematic statistical techniques, failure to properly document and preserve research results, and inappropriate authorship.” [V]

The bottomless-bowl study specifically is NOT retracted. Wansink 2005, Obes Res 13(1):93-100, n = 54, self-refilling soup bowls, reported +73% intake with no difference in perceived consumption or satiety. Raw PubMed XML checked directly this session: publication types are Clinical Trial / RCT / Journal Article, there is no RetractionIn link and no “Retracted Publication” type, record last revised 2022-04-08. It is from the lab found guilty of data misconduct, has been publicly questioned as to whether the experiment was run as described, and has not been independently replicated. Verdict: unreliable, not formally retracted. [V]

Independent evidence:

Source	Design	Result
Kosīte 2019, Int J Behav Nutr Phys Act 16:75	pre-registered (OSF) two-group RCT, general population, n = 134, self-served lunch, 29 cm vs 23 cm plate	Null. d = 0.07 (95% CI -0.27 to 0.41); mean difference 19.2 kcal (95% CI -76.5 to 115.0). Authors: “previous meta-analyses of a low-quality body of evidence may have considerably overestimated effects”
Robinson E 2014, Obes Rev 15(10):812-21	meta-analysis, 9 experiments	SMD -0.18 (95% CI -0.35 to 0.00), I² = 77%. “The majority of experiments found no significant difference.” Recommendations to use smaller plates “may be premature”
Holden, Zlatevska & Dubelaar 2016, J Assoc Consum Res 1(1)	meta-analysis	d = 0.70 when food is self-served; d = 0.03 when the portion is held constant. Figures as quoted in the Kosīte full text; the original was inaccessible
Peng 2017, Obes Sci Pract 3(3):282-8	ratings of 20 food photographs	perceptual study, no food consumed. Does not bear on intake

All [V].

The decomposition is the point. Given a fixed portion, plate size does essentially nothing (d = 0.03). The only large estimate is restricted to self-service, where the plate works by changing how much you serve yourself. Plate size is the portion-size effect wearing a costume.

And the Cochrane review does not rescue it. Hollands GJ, et al. Cochrane Database Syst Rev 2015;9:CD011045: 72 studies, all at unclear or high risk of bias. Composition: 49% portion size, 14% package size, 21% tableware. The headline is SMD 0.38 (95% CI 0.29 to 0.46) for consumption, moderate quality, which Cochrane translates to a potential 144-228 kcal/day reduction. But that number pools portion, package, unit and tableware together. The tableware-only strands are 1 to 3 studies rated very low and low quality. Citing SMD 0.38 as evidence for buying smaller plates is a category error. [V]

One standard, applied both ways. [Added on review.] This file rejects Hollands’ SMD 0.38 partly for pooling tableware with portion and package size, while accepting Robinson 2023’s SMD -0.709 (about -235 kcal/day) in §4.5, which pools 14 heterogeneous studies and 85 effect sizes with an acknowledged curvilinear dose-response. If pooling across dissimilar manipulations is disqualifying, it discounts both. The distinction that actually survives is narrower and should be stated as such: Hollands’ tableware-only strands are 1 to 3 studies rated low and very low quality, whereas Robinson’s pool is entirely portion-size manipulations, which is the construct being estimated. Read -235 kcal/day as a heterogeneous pooled central estimate, not a precise per-person number.

4.5 Portion size, which is the real thing

Verdict: REAL, large, and the only lever in this section with sustained-exposure evidence.

Rolls BJ, Roe LS, Meengs JS. Obesity 2007;15(6):1535-43. Crossover, all food provided, two 11-day periods, all portions increased 50%, n = 23. +423 ± 27 kcal/day (p < 0.0001), with no decline over 11 days, cumulative +4,636 ± 532 kcal. Effect present at all meals and all food categories except fruit-as-snack and vegetables. Not moderated by body-weight status. [V]
Robinson E 2023, Br J Nutr 129(5):888-903, meta-analysis, 14 studies / 85 effects with intake measured over ≥1 day: SMD -0.709 (95% CI -0.956 to -0.461), approximately -235 kcal/day for smaller versus larger portions. Curvilinear: shrinking already-very-large portions yields less. Weight subset (4 studies): 0.58 kg less weight gain. [V]

No habituation over 11 days. That is the striking part: people did not adjust.

4.6 Food temperature

Verdict: UNKNOWN for food. FOLKLORE at the claimed magnitude for cold water.

No human trial manipulating the temperature of a meal and measuring intake or satiety was located. Multiple query formulations returned nothing. Genuine gap.

Adjacent findings: Coca 2024, Br J Nutr 132(2):209-26, n = 23, 24-h passive exposure at 16 / 24 / 32 °C with ad libitum meals: total energy intake NOT modified (p = 0.120), hunger unchanged, but implicit wanting shifted toward cold/low-fat foods at 32 °C and warm/high-fat at 16 °C. Ambient temperature moves food choice without moving amount. [V]

Cold-water thermogenesis, which is the closest thing to a testable claim:

Study	Design, n	Result
Boschmann 2003, JCEM 88(12):6015-9	whole-room calorimetry + microdialysis, n = 14, 500 mL at 22 °C	+30% metabolic rate, total thermogenic response ~100 kJ (~24 kcal); ~40% attributed to warming the water; blunted by beta-blockade
Boschmann 2007, JCEM 92(8):3334-7	randomised crossover, n = 16	only 500 mL plain water raised expenditure (+24% over 60 min); saline and 50 mL null
Brown, Dulloo & Montani 2006, JCEM 91(9):3598-602	randomised crossover, with a working positive control (7% sucrose)	No increase after distilled water (p = 0.34) or saline (p = 0.33). Sucrose control worked (p < 0.0001). Cold water at 3 °C: +4.5% over 60 min (p < 0.01), which the authors note is well below the theoretical cost of warming the water

All [V]. First-principles arithmetic [C]: warming 500 mL from 3 °C to 37 °C costs 17 kcal. Brown’s measured +4.5% over 60 min against a ~60 kcal/h resting rate is about 2 to 3 kcal, roughly a tenth of the theoretical warming cost and a tenth of Boschmann’s 24 kcal. The failed replication is the methodologically stronger study because it included a positive control that worked. Not a weight-management lever at any plausible magnitude.

4.7 Spiciness and capsaicin

Verdict: REAL acutely, SMALL at the endpoint that matters, and absent in lean people.

Outcome	Source	Effect
Acute energy intake	Whiting 2014, Appetite 73:183-8, meta-analysis, 8 trials pooled, 191 participants	-309.9 kJ = -74.0 kcal per meal, p < 0.001. I² = 75.7%, authors advise caution. Minimum effective dose ~2 mg capsaicinoids
Energy expenditure	Whiting 2012, Appetite 59(2):341-8, 20 trials, 563 participants	~+50 kcal/day
Energy expenditure	Zsiborás 2018, Crit Rev Food Sci Nutr 58(9):1419-27, 9 trials	+58.6 kcal/day (p = 0.030). BMI < 25: null for both expenditure (p = 0.718) and RQ (p = 0.444). BMI > 25: +69.8 kcal/day (p = 0.023)
Body weight	Zhang 2023, Br J Nutr 130(9):1645-56, 15 RCTs, 762 participants	BMI -0.25 kg/m² (-0.35 to -0.15); body weight -0.51 kg (-0.86 to -0.15). Authors’ own word: “rather modest”

All [V]. The acute numbers look attractive; 15 RCTs deliver half a kilogram, and the expenditure effect is absent in lean people.

Observational mortality, for completeness and not as causal evidence: Lv J, et al. BMJ 2015;351:h3942, China Kadoorie Biobank, n = 487,375, 3.5 million person-years, 20,224 deaths. Versus spicy food less than once a week: HR 0.86 (0.82-0.90) at 6-7 days/week. Single self-reported baseline exposure, observational, residual confounding by regional diet unaddressed by design. [V]

5. Chrononutrition: does timing matter once calories are matched?

Verdicts here are unusually clean, because this is one of the few areas of nutrition with gold-standard isocaloric trials and doubly-labelled-water energy expenditure.

5.1 Time-restricted eating

Verdict: FOLKLORE for weight loss beyond calorie restriction. REAL but small and fragile for weight-independent metabolic effects.

Liu D, Huang Y, Huang C, et al. “Calorie Restriction with or without Time-Restricted Eating in Weight Loss.” N Engl J Med 2022;386:1495-1504. Parallel RCT, n = 139 randomised (118 completed, 84.9%), 12 months, Guangzhou. TRE 08:00-16:00 plus calorie restriction vs calorie restriction alone, with identical calorie prescriptions in both arms (1,500-1,800 kcal/d men, 1,200-1,500 kcal/d women). PMID 35443107. [V]

Weight -8.0 kg (95% CI -9.6 to -6.4) TRE+CR vs -6.3 kg (-7.8 to -4.7) CR.
Net difference -1.8 kg (95% CI -4.0 to +0.4), p = 0.11.
Waist, BMI, body fat, lean mass, blood pressure and metabolic risk factors: all consistent with the primary null.

The single adequately powered 12-month trial with matched calorie prescription found no incremental benefit from the window.

Lowe DA, Wu N, Rohdin-Bibby L, et al. TREAT trial. JAMA Intern Med 2020;180:1491-1499. Parallel RCT, n = 116, 12 weeks, app-based, ad libitum 12:00-20:00 vs three structured meals. Calories were neither prescribed nor measured, which the authors state. PMID 32986097. [V]

Between-group weight -0.26 kg (95% CI -1.30 to +0.78), p = 0.63. Null.
The lean-mass scare, in the n = 25/arm DXA subset: appendicular lean mass index -0.16 kg/m² (95% CI -0.27 to -0.05, p = 0.005) and appendicular lean mass -0.47 kg (-0.82 to -0.12, p = 0.009) favouring the control arm, but total lean mass was not significant: -0.75 kg (99.7% CI -1.96 to +0.45, p = 0.09). Authors report that of 1.70 kg lost in the TRE in-person cohort, ~65% was lean mass versus a normal 20-30%.
Three commentaries followed in JAMA Intern Med 2021, all titled “Caution Against Overinterpreting Time-Restricted Eating Results” (PMIDs 33616609, 33616602, 33616599), plus an author reply. Existence and DOIs verified; the specific arguments are UNVERIFIED (letters carry no abstract).
Reasonable discount on the lean-mass finding: n = 50 subset, primary lean-mass measure null, no protein-intake or resistance-training data, DXA fat-free mass is water-sensitive, and many secondary outcomes were tested.

Sutton EF, Beyl R, Early KS, et al. “Early time-restricted feeding improves insulin sensitivity, blood pressure, and oxidative stress even without weight loss in men with prediabetes.” Cell Metab 2018;27(6):1212-1221. This is the key trial, because it is eucaloric by design: all food provided, participants fed to maintain weight. Supervised controlled-feeding randomised crossover, n = 8 completers (12 enrolled), men with prediabetes, 5 weeks per arm, ~7-week washout. 6-h feeding window with dinner before 15:00 vs 12-h control. PMID 29754952. [V]

Outcome	Effect
Insulin resistance index	-36 ± 10 (p = 0.005)
Fasting insulin	-3.4 ± 1.6 mU/L (p = 0.05)
Mean postprandial insulin	-26 ± 9 mU/L (p = 0.01)
Mean glucose	unchanged, +5 ± 5 mg/dL (p = 0.40)
Systolic BP	-11 ± 4 mmHg (p = 0.03)
Diastolic BP	-10 ± 4 mmHg (p = 0.03)
8-isoprostane (oxidative stress)	-11 ± 5 pg/mL, ~-14% (p = 0.05)
Triglycerides	+57 ± 13 mg/dL (p = 0.0007)
Total cholesterol	+13 ± 5 mg/dL (p = 0.02)
Weight	-1.4 vs -1.0 kg, p = 0.12 (neutral by design)

Two things proponents rarely quote. A -11/-10 mmHg blood pressure effect at n = 8 is implausibly large and has not replicated at that magnitude (Jamshed 2022 found -4 mmHg diastolic only). And triglycerides and total cholesterol went up, materially.

Jamshed H, Steger FL, Bryan DR, et al. JAMA Intern Med 2022;182:953-962. Parallel RCT, n = 90, 14 weeks, eTRE 07:00-15:00 plus energy restriction vs ≥12 h window with identical energy restriction and identical counselling. PMID 35939311. [V]

Weight -2.3 kg (95% CI -3.7 to -0.9, p = 0.002) favouring eTRE. The authors themselves describe this as “equivalent to reducing calorie intake by an additional 214 kcal/d.”
Body fat -1.4 kg (95% CI -2.9 to +0.2, p = 0.09), not significant. Fat-loss to weight-loss ratio, p = 0.43.
Diastolic BP -4 mmHg (p = 0.04); systolic and all other cardiometabolic outcomes not significant. Mood improved.

Weight fell without a matching fat-mass signal, and the effect is expressed by its own authors in kcal/d. That is an intake effect wearing a timing costume.

Jamshed H, et al. Nutrients 2019;11:1234, n = 11, 4 days per arm, genuinely matched feeding: mean 24-h glucose -4 ± 1 mg/dL (p = 0.0003), glycemic excursions -12 ± 3 mg/dL (p = 0.001), plus ketone, SIRT1, LC3A, MTOR and clock- gene changes. PMID 31151228. Real, small, glycemia-only, 4 days, n = 11. [V]

Manoogian ENC, et al. “Healthy Heroes” firefighter RCT. Cell Metab 2022;34:1442-1456. n = 137, 12 weeks, 10-h self-selected TRE plus Mediterranean counselling vs counselling alone, not calorie prescribed. PMID 36198291. [V]

Eating window shortened by -2.99 h (95% CI -3.51 to -2.47) vs -0.60 h. Feasibility endpoint met.
Weight: -0.94 kg vs -0.43 kg, between groups p = 0.163. Not significant.
The only whole-cohort between-group hit was VLDL particle size -1.34 nm vs -0.25 nm, p = 0.044.
Subgroup hits in elevated-risk participants only: HbA1c p = 0.012 (n = 10 vs 11), diastolic BP p = 0.033 (n = 9 vs 6).
Null: fasting glucose, insulin, HOMA-IR, LDL, triglycerides, BMI, body fat, sleep, CRP.

The trial demonstrated feasibility in shift workers. The efficacy headline rests on one lipoprotein subfraction plus subgroups of n = 6-11.

Meta-analyses, stratified by whether calories were matched:

Analysis	Comparator	Pooled effect
Li 2023, Eur J Clin Nutr, 8 RCTs, n = 579	TRE+CR vs CR alone	Weight -1.40 kg (95% CI -1.81 to -1.00), I² = 0%; fat mass -0.73 kg; waist -1.87 cm. No benefit on BP, glucose or lipids. PMID 37488260
Front Nutr 2025, 13 RCTs, n = 612, women	Both	Weight -1.93 kg (-3.69 to -0.17). Greater than conventional diets (p = 0.046) but NOT greater than calorie restriction alone (p = 0.295). No lean-mass loss. All metabolic outcomes null. PMID 41036194
Sun 2024, Nutrients, 26 studies	TRE vs regular diet without CR	Weight -1.62 kg; measures spontaneous intake reduction, not timing. PMID 39519533
Chen 2024, Adv Nutr, 19 RCTs, n = 568	TRE+exercise vs diet+exercise	Body mass -1.86 kg; fat -1.52 kg; no glucose/insulin benefit. PMID 38897385

Pooled reading: roughly 1-2 kg, and the two analyses that isolate the calorie-matched contrast give either a ~1.4 kg residual with zero metabolic co-movement, or no significant difference at all.

The 2024 “8-hour window, 91% higher CV mortality” claim. Originally an AHA EPI|Lifestyle 2024 poster abstract, not peer reviewed, and 34 TRE researchers formally objected to the press release. Documented criticisms: exposure derived from two 24-h dietary recalls assumed to represent years of behaviour; only ~414 of 20,078 participants in the <8 h stratum with 31 CV deaths; three to four times more Black participants in the short-window group. It has since been published: Zhong et al., Diabetes Metab Syndr 2025, n = 19,831, median 8.1 y follow-up: CV mortality HR 2.35 (95% CI 1.39-3.98), no association with all-cause or cancer mortality, and the all-cause association “did not survive many sensitivity analyses.” PMID 40849219. [V]

Note the published hazard ratio (2.35) is larger than the 1.91 in the press release, so the model changed between abstract and publication. Verdict: the association is real and the causal claim is not. Short eating windows in NHANES are enriched for illness, poverty, shift work and food insecurity.

5.2 Late eating with calories matched

Verdict: REAL for mechanism, UNKNOWN for outcome.

Vujović N, Piron MJ, Qian J, et al. “Late isocaloric eating increases hunger, decreases energy expenditure, and modifies metabolic pathways in adults with overweight and obesity.” Cell Metab 2022;34:1486-1498. Randomised controlled crossover, n = 16, all meals delayed 250 min, identical meals in both conditions, with control of nutrient intake, activity, sleep and light. NCT02298790. PMID 36198293. [V]

Hunger increased, p < 0.0001; odds of a hunger rating >50 doubled (OR 2.02).
Waketime ghrelin:leptin ratio +34% (p < 0.0001); 24-h ratio +12% (p = 0.0063); waketime leptin -16% (p < 0.0001).
Waketime energy expenditure -59.4 ± 13.9 kcal/day, -5.03% (p = 0.002).
24-h core body temperature -0.19 °C (p = 0.019).
Adipose transcriptome shifted toward lipogenesis: lipid-breakdown genes down (PLD6, DECR1), lipid-synthesis genes up (GPAM, ACLY).

Six days, 16 people, no weight endpoint. About 60 kcal/d is ~2% of daily expenditure. Whether it compounds is untested, and see §5.3 for a 4-week doubly-labelled-water study that found no timing effect on total expenditure.

Glucose tolerance really is worse in the biological evening. Well replicated, mechanistically clean, and independent of calories:

Morris CJ, et al. PNAS 2015;112:E2225-34. Postprandial glucose 17% higher at 20:00 than 08:00 as a circadian-phase effect; circadian misalignment alone added 6%. Evening deficit was 27% lower early-phase insulin. PMID 25870289. [V]
Morris CJ, et al. J Clin Endocrinol Metab 2016;101:1066-74, chronic shift workers, randomised crossover: +6.5% at 20:00 vs 08:00 (p = 0.0041), misalignment +5.6% (p = 0.0042). PMID 26771705. [V]
Bandín C, et al. Int J Obes 2015;39:828-833, randomised crossover, n = 32 women, provided meals, lunch at 13:00 vs 16:30: glucose AUC above baseline +46% (p = 0.002), lower pre-meal resting expenditure (p = 0.048) and carbohydrate oxidation (p = 0.006). PMID 25311083. [V]

This is a metabolic fact. It is not the claim “late eating makes you fat.”

The observational late-eating cohorts are SMALL and CONFOUNDED: Garaulet 2013 Int J Obes (n = 420, 20-week program, split at 15:00 lunch) found early eaters lost -9.9 ± 5.8 kg vs -7.7 ± 6.1 kg, p = 0.008 with self-reported intake similar (1,426 vs 1,388 kcal/d, p = 0.191). PMID 23357955. Ruiz-Lozano 2016 Clin Nutr (n = 270 post-bariatric) similar direction, p = 0.011. PMID 26948400. Self-reported intake “similar” is weak evidence of calorie matching, since under-reporting is differential by chronotype and by weight-loss success. [V]

5.3 Breakfast, and the big-breakfast question

Verdict: FOLKLORE, and inverted.

Sievert K, Hussain SM, Page MJ, et al. “Effect of breakfast on weight and energy intake: systematic review and meta-analysis of randomised controlled trials.” BMJ 2019;364:l42. 13 RCTs (7 for weight, 10 for intake), high-income countries. PMID 30700403. [V]

Weight: mean difference 0.44 kg (95% CI 0.07 to 0.82) FAVOURING BREAKFAST SKIPPERS, I² = 43%.
Energy intake: breakfast eaters consumed 259.79 kcal/d MORE (95% CI 78.87 to 440.71), I² = 80%.
All trials at high or unclear risk of bias; mean follow-up 7 weeks for weight.

Betts JA, Richardson JD, Chowdhury EA, et al. Bath Breakfast Project. Am J Clin Nutr 2014;100:539-547. RCT, n = 33 lean adults, 6 weeks, ≥700 kcal before 11:00 vs nothing until 12:00, free-living, with all components of energy balance measured. PMID 24898233. [V]

No metabolic adaptation: RMR stable within 11 kcal/d.
Breakfast group’s total intake stayed +539 kcal/d (95% CI 157 to 920).
Physical activity thermogenesis +442 kcal/d (95% CI 34 to 851) with breakfast.
No difference in body mass or adiposity. Afternoon/evening glycemia more stable with breakfast.
Obese replication (Chowdhury 2016, n = 23, PMID 26864365): RMR stable within 8 kcal/d, no body-composition effect, reduced insulinemic OGTT response with breakfast (p = 0.05). [V]

So breakfast does not “start your metabolism”. It adds calories, and the body partly compensates by moving more, and the two roughly cancel.

The decisive isocaloric big-breakfast trial. Ruddick-Collins LC, Morgan PJ, Fyfe CL, et al. “Timing of daily calorie loading affects appetite and hunger responses without changes in energy metabolism in adults with obesity.” Cell Metab 2022;34:1472-1485. Randomised crossover, n = 30, two 4-week isoenergetic calorie-restricted diets, morning-loaded 45:35:20 vs evening-loaded 20:35:45, all food provided, total energy expenditure by doubly labelled water across the full 4 weeks, plus indirect calorimetry and 4-compartment body composition. PMID 36087576. [V] for the design; the point estimates and p-values below are [R], downgraded on review. The abstract carries no p-values, so p = 0.848 and p = 0.184 could not be confirmed from anything fetched in the originating session.

Weight loss: -3.33 kg (morning-loaded) vs -3.38 kg (evening-loaded), SED 0.24, p = 0.848. Effectively identical. [R]
TDEE by DLW: 2,871 vs 2,846 kcal/d, SED 100.7, p = 0.184. No difference. [R]
RMR 1,675 vs 1,690 kcal, no timing effect. Breakfast thermic effect did differ (147 vs 98 kcal) with no consequence for total expenditure.
Morning loading produced significantly lower hunger, desire to eat, prospective consumption, thirst and composite appetite score (all p < 0.05).

And the companion paper kills the mechanism directly: Ruddick-Collins et al. J Clin Endocrinol Metab 2022;107:e708-e715, n = 14. The standard thermic-effect calculation makes morning TEF (60.8 ± 5.6 kcal) look 1.6× lunch and 2.4× dinner (p = 0.022). After adjusting for the underlying circadian rhythm in resting metabolic rate, TEF is flat: 54.1 vs 49.5 vs 49.1 kcal, p = 0.680. PMID 34473293. [V]

“A calorie in the morning burns hotter” is a measurement artifact of not adjusting for the circadian RMR rhythm. Front-loading helps, and it helps entirely through appetite, i.e. through eating less.

For completeness, Jakubowicz D, et al. Obesity 2013;21:2504-2512, n = 93 women with metabolic syndrome, 12 weeks, two isocaloric ~1,400 kcal/d arms (700/500/200 vs 200/500/700): greater weight and waist reduction in the big-breakfast arm, and triglycerides -33.6% vs +14.6%. PMID 23512957. Abstract-level findings verified; the frequently quoted -8.7 kg vs -3.6 kg figures are UNVERIFIED (Wiley returned 403 on every route). Read alongside Ruddick-Collins, the Jakubowicz advantage is best interpreted as an adherence and intake effect under free-living conditions.

6. Meal frequency and snacking

6.1 “More meals stoke the metabolic fire”

Verdict: FOLKLORE, definitively, on three independent chamber-calorimetry studies.

Study	Design	24-h energy expenditure
Verboeket-van de Venne & Westerterp 1991, Eur J Clin Nutr 45:161-9, n = 13, respiration chamber, fed to energy balance	2 vs 7 meals/d	5.57 ± 0.16 vs 5.44 ± 0.18 kJ/min. Null. Only oxidation periodicity changed
Taylor & Garrow 2001, Int J Obes 25:519-528, women BMI>25, fixed 4.2 MJ/d, chamber calorimetry	6 vs 2 meals	10.00 vs 9.96 MJ, p = 0.88. Night EE was higher with 2 meals (9.12 vs 8.34 MJ/24 h, p = 0.02): TEF is delayed, not lost. Later intake unaffected (p = 0.58)
Ohkawara K, et al. 2013, Obesity 21:336-343, n = 15 lean, whole-room calorimeter, randomised crossover, isoenergetic	3 vs 6 meals	8.7 ± 0.3 vs 8.6 ± 0.3 MJ/d; RQ 0.85 vs 0.85; fat oxidation 82 ± 6 vs 80 ± 7 g/d. All null. Hunger AUC greater with 6 meals (p = 0.03)

[V] for all three. Bellisle F, McDevitt R, Prentice AM. Br J Nutr 1997;77 Suppl 1:S57-70 reached the same conclusion and attributed the epidemiological frequency-adiposity association to dietary under-reporting: “at best very weak, and almost certainly represents an artefact.” PMID 9155494. [V]

The thermic effect of food scales with total calories, not with how many times you eat them. Splitting a fixed load into more meals splits the same thermic response. The 2013 chamber study additionally found six meals made people hungrier.

6.2 Meal frequency and body composition

Verdict: FOLKLORE / UNKNOWN.

Schoenfeld BJ, Aragon AA, Krieger JW. Nutr Rev 2015;73:69-82, meta-regression of 15 studies: frequency was positively associated with fat loss, but sensitivity analysis showed the positive findings were the product of a single study, “casting doubt as to whether more frequent meals confer beneficial effects.” PMID 26024494. [V]
Cameron JD, Cyr MJ, Doucet É. Br J Nutr 2010;103:1098-1101. RCT, n = 16, 8 weeks, identical -2,931 kJ/d restriction, 3 meals vs 3 meals + 3 snacks: no significant between-group difference in any adiposity index, appetite measure, PYY or ghrelin. PMID 19943985. [V]
Canuto R, et al. Public Health Nutr 2017;20:2079-2095, 31 observational studies, n = 136,052; only 6 accounted for misreporting. Conclusion: “not sufficient evidence confirming the association … when misreporting bias is taken into account.” PMID 28578730. [V]
2025 US Dietary Guidelines Advisory Committee / USDA NESR systematic review, “Frequency of Meals and/or Snacking and Energy Intake”: grade “Not Assignable” for number of eating occasions in adults, for breakfast in adults (19 articles, 18 RCTs, mixed results), and for snacking in adults (5 RCTs). PMID 39812565. [V]

The US federal evidence review declines to assign a grade at all.

6.3 The one trial that looks like an exception, and why it is not

Kahleova H, Belinova L, Malinska H, et al. Diabetologia 2014;57:1552-1560. Randomised open crossover, n = 54 T2D patients, 12 weeks per regimen, with the same macronutrient and energy content in both regimens. Six meals/day vs two meals/day (breakfast and lunch only). Hepatic fat by ¹H-MRS, insulin sensitivity by clamp. PMID 24838678. [V]

Weight: -2.3 kg (6 meals) vs -3.7 kg (2 meals), p < 0.001. ~1.4 kg favouring two meals over 12 weeks.
Hepatic fat content fell more with 2 meals (p = 0.009); OGIS rose more (p = 0.01); fasting glucose (p = 0.004) and C-peptide (p = 0.04) fell more. Fasting glucagon fell with 2 meals but rose with 6.

This is a timing result wearing a frequency label. The two-meal arm did not just eat fewer times, it ate breakfast and lunch and then nothing after midday, so frequency, eating window and circadian phase are fully confounded. The direction matches Ruddick-Collins and the evening-glucose-tolerance literature, and it is contradicted as a pure frequency effect by Ohkawara, Taylor and Cameron. Open-label with self-reported adherence.

6.4 Snacking per se

Verdict: UNKNOWN, trending FOLKLORE. The USDA/DGAC 2025 review graded snacking in adults “Not Assignable” on five RCTs. There is no credible evidence that the act of snacking, at fixed calories, changes weight or metabolic outcomes. What snacking reliably does in free-living settings is add calories, which is a different claim.

7. Cooking method: real risk sizes, not mechanisms

The organising question here is not “does this chemistry happen” but “has anyone ever measured a human outcome.” For three of the four sub-topics the answer is no, and for the fourth the answer runs opposite to the folk model.

7.1 Air frying versus deep frying

Fat and calories: REAL, and the largest measured effect in this section.

Study	Measured
Teruel 2015, J Food Sci 80:E349-58	deep fried 5.63 g oil/100 g defatted dry matter vs air fried 1.12; full ranges air 0.37-1.12, deep 5.63-13.77; 9 min at 180 °C vs 21 min air
LWT 2026, air vs deep fried French fries	deep fry oil content 44.8%, air fry ~1.2%, at comparable moisture, colour and texture
Air jet impingement, Food Sci Nutr 2026, PMC13260683	oil absorption -36.2% (fries), -38.8% (wedges); energy use -80%
Vegetable patties, Foods 2026, PMC12841028	deep frying raised fat 45-75%; baking and air frying 0-10%

All [V]. But note the measurement-basis trap: these papers report oil on defatted-dry-matter or unstated bases and disagree by an order of magnitude on the deep-fried number. Anyone quoting “air frying saves X kcal per 100 g” is almost certainly converting across bases silently. [C] Teruel’s 4.51 g/100 g DDM gap translates to roughly 16-18 kcal/100 g of finished product at ~55% moisture; the LWT numbers imply 150-300 kcal/100 g. These could not be reconciled. For scale, USDA oven-heated frozen fries themselves span 3.76 to 18.7 g fat/100 g by cut, which is the same size as the entire appliance effect. Product choice matters as much as appliance.

Acrylamide: the “air frying makes more” claim is FOLKLORE.

Sansano 2015, J Food Sci 80:T1120-8, both at 180 °C: air frying reduced acrylamide ~90% versus deep-oil frying, with no pretreatment needed. [V]
The one study pointing the other way, Al-Dalali 2023, Front Nutr 10:1297069, found air 12.19 ± 7.03 vs deep 8.94 ± 9.21 µg/kg, p = 0.789, not significant, and every reported value sits below the paper’s own limit of quantitation of 18.20 ng/g. A published commentary (Front Nutr 2024, PMID 38933884) notes the LoD/LoQ were established against “blank” chips already containing 575 ± 37 µg/kg, that the values are ~10x below the literature (FDA range for chips and fries is 59 to 5,200 µg/kg), and that the sunflower oil used contains 890-1,200 µg/kg, enough that oil carryover alone could explain the readings. [V]

Air frying at higher temperature and with searing does raise acrylamide within air frying, which is probably the origin of the claim.

Does dietary acrylamide harm humans? FOLKLORE at dietary doses.

Regulatory positions are rodent-derived hazard flags, not measured human risk. EFSA 2015 CONTAM: BMDL10 0.17 mg/kg bw/day for neoplastic effects, margins of exposure 425 (average adults) down to 50 (high-consuming toddlers) against EFSA’s own 10,000 threshold. [C] MOE 425 implies mean adult exposure of about 0.4 µg/kg bw/day. EFSA’s 2022 genotoxicity re-assessment states plainly: “There was no consistent indication for an association between AA exposure and increased risk of cancer in various organs.” FDA’s 2016 guidance sets no action level and no maximum level. [V]

The human epidemiology:

Study	Design, n	Result
Pelucchi 2015, Int J Cancer 136:2912-22	meta-analysis, 32 publications, 14 cancer sites	high vs low summary RRs from 0.87 to 1.14. Only kidney borderline: RR 1.20 (1.00-1.45). All continuous per-10 µg/day estimates between 0.95 and 1.03, none significant
Filippini/Kito dose-response meta, Adv Nutr 2022, PMC9082595	16 studies, 1,151,189 participants, 48,175 cancer cases, median 14.9 y	every site null: oral 0.99, esophageal 1.05, stomach 0.92, colorectal 0.94, pancreatic 0.88, lung 0.91, renal 1.08, prostate 1.00, bladder 0.89. Only lung in smokers 1.16 (1.03-1.31). “No association … with no evidence of thresholds”

[V] both. Over 1.15 million prospectively followed people and 48,175 cancers, the point estimates cluster at 1.00 and several sit below it. Nothing in the human literature supports choosing an appliance to avoid acrylamide.

Fried food and cardiovascular outcomes: SMALL and confounding-dominated.

Qin 2021, Heart 107:1567, meta-analysis: 17 studies, 562,445 participants / 36,727 major adverse cardiac events; mortality arm 6 studies, 754,873 participants / 85,906 deaths. Highest vs lowest: MACE 1.28 (1.15-1.43), I² = 82%; CHD 1.22 (1.07-1.40); heart failure 1.37 (1.07-1.75); stroke 1.37 (0.97-1.94) NS; CVD mortality 1.02 (0.93-1.14) NS; all-cause mortality 1.03 (0.96-1.12) NS. Per additional 114 g/week: +3% MACE. [V]
Sun 2019, BMJ 364:k5420, Women’s Health Initiative, n = 106,966, mean 17.9 y, 31,558 deaths: total fried food ≥1 serving/day vs none, all-cause HR 1.08 (1.01-1.16); CVD mortality 1.08 (0.96-1.22) NS; cancer mortality 0.93 (0.82-1.06) NS. Fried chicken ≥1/week: all-cause 1.13 (1.07-1.19). [V]

The confounding is stated by the authors themselves. Qin’s effect collapses on adjustment: highest-vs-lowest MACE was +46% in studies not adjusting for BMI versus +17% in those that did, and +80% unadjusted for physical activity versus +17% adjusted. I² of 75-92% means the pooled estimate is model-dependent. And the two endpoints least prone to misclassification, all-cause and CVD mortality in Qin, are null. [C] From WHI’s overall death rate of 16.5 per 1,000 person-years, an HR of 1.08 is about +1.3 deaths per 1,000 person-years.

7.2 Grilling: heterocyclic amines and PAHs

Chemistry: REAL and large. Human outcome: SMALL, and not separable from meat intake itself. IARC declines to conclude.

Measured concentrations. Sinha 1998, Food Chem Toxicol, PMID 9651044: MeIQx rises with doneness up to 8.2 ng/g; PhIP in steak 1.9 to 30 ng/g. Roast beef contained none of the five HCAs. Sinha 1995, Cancer Res, PMID 7553619: PhIP 12 to 480 ng/g in pan-fried, oven-broiled and grilled chicken breast, but roasted whole chicken and stewed chicken contained none. So within one food, method spans roughly two orders of magnitude, and it is method rather than temperature alone that drives it. [V]

The home-cooking reality check matters more than either. Keating & Knize 2000, Cancer Causes Control, PMID 11065010, classified samples by photograph. Very well done hamburger: MeIQx 1.88, PhIP 2.04 ng/g. Very well done steak: MeIQx 1.87, PhIP 0.62 ng/g. Roughly an order of magnitude below the restaurant and laboratory figures. And the finding that undermines the entire epidemiology: when samples were categorised by participants’ self-reported doneness preference, HCA concentrations did not differ significantly. They separated only when doneness was assessed at cooking time or by photograph. The doneness variable that cohort food questionnaires actually collect is close to non-informative. [V]

Human outcomes:

Study	Design, n	Effect
Cross 2010, Cancer Res 70:2406	NIH-AARP, n = 300,948, 2,719 colorectal cases, 7 y	Q5 vs Q1: red meat 1.24 (1.09-1.42), MeIQx 1.19 (1.05-1.34), DiMeIQx 1.17 (1.05-1.29)
Taunk 2016, Int J Cancer	NIH-AARP, n = 322,846, 1,417 pancreatic cases	grilled/barbecued 1.24 (1.03-1.50); well/very well done 1.32 (1.10-1.58); significant in men, not women
Sinha/Wu 2015, CEBP	HPFS, n = 26,030, 2,770 prostate cases	PhIP 1.18 (1.03-1.35); MeIQx, DiMeIQx and total mutagenicity all NS. Authors’ own impact statement: “Results do not provide strong evidence that HCAs increase risk of prostate cancers”
Le 2016, Environ Health Perspect	NHS + HPFS pooled, n = 95,490, 1,208 CRC cases, 14 y	all null: MeIQx 1.12, PhIP 1.10, total mutagenicity 1.03
Wu 2010, CEBP	NHS, 2,317 breast cancers	all null or inverse: MeIQx 0.90, PhIP 0.92
Daniel 2012, J Nutr	NIH-AARP, 3,611 non-Hodgkin lymphoma	no association; MeIQx and DiMeIQx inversely associated with CLL/SLL

All [V].

The collinearity problem is fatal to causal inference here. Cross 2010’s HCA hazard ratios (1.17-1.19) are statistically indistinguishable from the red-meat hazard ratio in the same cohort (1.24), because HCA intake is estimated by multiplying meat intake by a doneness-derived database coefficient. The two exposures are near-collinear by construction. And Keating shows the doneness component carries little real signal as collected.

No Mendelian randomisation of meat doneness or HCA exposure exists. Europe PMC was searched; existing meat MR studies use single-item frequency questions with no doneness instrument and produce estimates no cohort supports.

IARC’s own verdict, verbatim: high-temperature and direct-flame cooking “produces more of certain types of carcinogenic chemicals … However, there were not enough data for the IARC Working Group to reach a conclusion about whether the way meat is cooked affects the risk of cancer.” [V]

PAHs are, at measured dietary levels, FOLKLORE. Market surveys by GC-MS (Sci Rep 2025, PMC12394643): benzo[a]pyrene 0.62-1.25 µg/kg in grilled chicken, beef and kebab, all within EU limits, with Monte Carlo incremental lifetime cancer risk in the acceptable range. A second survey (PMC12234855) found total carcinogenic risk below 10⁻⁶, i.e. negligible. [V]

The counterintuitive part: the largest PAH exposure from grilling is inhaled, not eaten. Pork belly over charcoal produced benzo[a]pyrene at 1,431 ng/Sm³ in the smoke against the WHO/EU air guideline of 1 ng/Sm³, roughly 1,400x (Food Chem X 2026, PMC12857362). [V] Standing over the grill is the exposure; eating what comes off it is not.

Marinades: REAL as chemistry, FOLKLORE as health advice, and not even internally consistent.

Salmon 1997, Food Chem Toxicol, PMID 9216741: a sugar-containing marinade on flame-broiled chicken breast cut PhIP by 92-99% but raised MeIQx more than tenfold at 30 and 40 min, with sugar identified as the driver. Ames mutagenicity was lower at 10, 20 and 30 min but higher in the marinated samples at 40 min. [V] So the intervention moves different HCAs in opposite directions and can increase total mutagenicity at long cook times.

Other measured reductions, all chemistry endpoints with zero outcome data: milk or beer marinade cut HCAs -60.6% in air-fried chicken, turmeric -69.4% in beef (Food Sci Biotechnol 2025, PMC12589718); horseradish nanovesicles cut total HCAs -53.4% and AGEs -50.6%; foil wrapping alone dropped benzo[a]pyrene below permissible values. [V]

The genuinely free intervention: Salmon 2000, JNCI 92:1773, turning patties every minute significantly lowered HCAs versus a single turn and reached 70 °C internal temperature sooner, with full E. coli inactivation regardless of method. Less chemistry and equal pathogen kill, at no cost. [V]

Microwave pre-treatment: Toxins 2025, PMC12197532, 2 min pre-treatment of beef sirloin cut norharman -78.4% and harmaline -96.5%. Vanderlaan 1989 reported reduced PhIP immunoreactivity and mutagenicity qualitatively. The often-cited Felton 1994 fold-reduction figures could not be located: UNVERIFIED.

Scale against processed meat. [C] IARC: 50 g/day of processed meat raises colorectal cancer risk ~18%; SEER lifetime US colorectal risk is 3.9%, so a 50 g/day habit moves it to about 4.6%, i.e. +0.7 percentage points. The cooking-method channel, using Cross 2010’s MeIQx HR of 1.19 against a 0.90% seven-year baseline, is about +1.7 cases per 1,000 people over 7 years for the most-exposed fifth versus the least, and that estimate is not separable from red meat intake itself. “Eat less processed meat” has a Group 1 classification behind it; “cook it less aggressively” has an IARC “not enough data.”

7.3 Dietary advanced glycation end products

Verdict: SMALL to UNKNOWN. A reproducible surrogate signal from small short-term trials in an arbitrary unit, and no human hard-outcome evidence of any kind.

The database, from Uribarri 2010, J Am Diet Assoc 110:911-16, kU AGE per 90 g serving of chicken, verified from the published table this session: [V]

Preparation	kU per 90 g
Boiled in water 1 h	1,011
Poached 15 min	968
Raw, skinless	692
Pan fried 13 min	4,444
Grilled	4,364
Broiled 450 °F 15 min	5,245
Roasted 45 min with skin	5,975
Breaded, deep fried 20 min	8,750

Grilled versus boiled is roughly 5x; breaded deep fried versus boiled is ~9x.

But note the red flag visible inside the table itself: raw skinless breast reads 692 kU per serving, only about 30% below boiled. That is hard to reconcile with a Maillard-formation model, and it points at the measurement. The Uribarri values are ELISA immunoreactivity in arbitrary kU, not mass-spectrometric quantitation of specific compounds. Modern work quantifies CML, CEL and MG-H1 individually by LC-MS/MS, and results in kU are not convertible to those. Cross-study comparison of “AGE intake” is therefore not on a common scale.

Randomised human trials, all on surrogate endpoints:

Trial	Design	Result
de Courten 2016, AJCN, PMID 27030534	n = 20, double-blind randomised crossover, isoenergetic macronutrient-matched high- vs low-AGE diets, 2 wk each, hyperinsulinemic-euglycemic clamp, AGEs quantified by MS	insulin sensitivity +1.3 mg/kg/min in favour of the low-AGE diet, p = 0.004 (abstract). No difference in body weight or insulin secretion. [Corrected on review: this row previously read -2.1 mg/kg/min, p = 0.001, which contradicted the +1.3 figure used in the summary table and does not appear in the abstract. Treat -2.1 as UNVERIFIED and use +1.3.] https://pubmed.ncbi.nlm.nih.gov/27030534/
Baye 2017a, Sci Rep, PMC5482825, same n = 20 crossover	inflammation and CVD markers	all null: SBP p = 0.2, DBP p = 0.3, TC p = 0.3, LDL p = 0.7, HDL p = 0.2, TG p = 0.4, IL-6 p = 0.6, MCP-1 p = 0.9, TNF-alpha p = 0.2, CRP p = 0.6, NF-kB p = 0.2
Vlassara 2016, Diabetologia, PMC5129175	randomised parallel, 1 year, 61 vs 77 randomised, 49 and 51 analysed (20-34% attrition), only lab staff blinded	HOMA-IR 3.1 to 1.9 low-AGE (p < 0.001) vs 2.9 to 3.6 regular-AGE. No effect on MRI-measured visceral or subcutaneous fat, or on carotid artery
Baye 2017b, Sci Rep, PMC5442099	meta-analysis, 17 RCTs, 560 participants	insulin resistance MD -1.3 (-2.3 to -0.2); total cholesterol -8.5 mg/dL; LDL -2.4 mg/dL. No change in weight, fasting glucose, 2-h glucose, 2-h insulin, HbA1c, HDL or blood pressure

All [V].

The evidence base is narrow. The food AGE database, the receptor mechanism and the largest positive trial all originate from one group (Vlassara/Uribarri, Mount Sinai). The main independent replication comes from one Australian group whose own crossover was null on every inflammatory and cardiovascular marker. The positive signal is confined to insulin sensitivity and lipids while every downstream measure is null.

Absorption, corrected. Koschinsky 1997, PNAS 94:6474: serum AUC after an AGE-containing meal was about 10% of ingested AGE, and renal excretion of dietary AGE was “normally incomplete (only ~30% of amount absorbed)”. So it is ~10% absorbed and roughly two-thirds of the absorbed fraction retained, not “one-third retained” as the claim is usually stated. Net, on the order of 6-7% of ingested AGE retained by a 1997 ELISA measurement. [V]

Hard outcomes: null.

Meta-analysis of 5 prospective cohorts, 1,220,096 participants, 23,229 incident cancers (Food Sci Nutr 2024, PMC11521677): highest vs lowest dietary AGE, overall cancer HR 1.04 (0.94-1.15); breast 1.12; pancreatic 1.24; colon 0.99; rectal 0.94. Null. [V]
Jiao 2015, AJCN, NIH-AARP, n = 528,251, 2,193 pancreatic cancers: Q5 vs Q1 men 1.43 (1.06-1.93), women 1.14 (0.76-1.72) NS. Red meat HR attenuated from 1.35 to 1.20 after adjusting for the AGE marker, i.e. the two are collinear. Sex discordance plus collinearity makes it weak. [V]
DFG SKLM assessment, Crit Rev Toxicol 2024, PMID 39150724: systematic review, 253 publications screened, 192 meeting quality criteria. Animal evidence exists only “at dose levels by far exceeding estimated human exposures.” Conclusion, verbatim: “There is at present no convincing evidence for a causal association between dietary intake of glycation compounds and adverse health effects.” [V]

Boiling chicken instead of grilling it cuts measured AGE intake about fivefold. Nobody has shown that this changes anything a person would notice.

7.4 The cooking effect that is real and runs the other way

Verdict: REAL, human-measured, 2-4x, and it contradicts the “cooking destroys nutrients” folk model.

Gärtner 1997, AJCN 66:116, human crossover, single 23 mg lycopene dose with 15 g corn oil, chylomicron analysis: tomato paste gave 2.5x higher peak lycopene and 3.8x higher AUC than fresh tomatoes (p < 0.001). Isomer pattern identical and triacylglycerol response not different, so this is matrix disruption, not a fat-delivery artifact. [V]
Nutr Res Pract 2025, PMC11982686, randomised crossover, n = 16, 25 mg beta-carotene: fresh carrot juice gave 2.33x peak plasma and 2.09x AUC versus raw whole carrots. Mechanical disruption alone, no heat. [V]

So the operative variable is physical matrix breakdown, of which cooking is one instance. Compare §8.3: fat co-ingestion multiplies the same carotenoids by 5-15x. Matrix disruption and fat are multiplicative levers on the same nutrient, and both are larger than anything cooking method does on the harm side.

Nutrient loss, for balance. Yuan 2009, J Zhejiang Univ Sci B 10:580, broccoli, five domestic methods, AOAC titrimetric vitamin C in triplicate: steaming no significant loss; microwaving 16%; stir-frying 24%; boiling 33%; stir-fry then boil 38%. Carotenoids retained better than vitamin C across all methods. The mechanism is leaching into the cooking water, not thermal destruction, which is why steaming is nearly free. [V] Folate per-method percentages could not be verified: UNVERIFIED.

And cooking increases usable calories. Carmody & Wrangham 2011, PNAS 108:19199: in mice, cooking substantially increased energy gained from meat, producing body-mass elevations not attributable to intake or activity differences, and cooking beat pounding for both meat and tubers. The authors’ conclusion is the interesting one: this “illuminates a weakness in current food labeling practices, which systematically overestimate the caloric potential of poorly processed foods.” [V] Human data are indirect only: Koebnick 1999, Ann Nutr Metab, cross-sectional questionnaire, n = 513 long-term raw-food dieters, mean 3.7 years: weight loss 9.9 kg (men) / 12 kg (women) from diet onset, BMI < 18.5 in 14.7% of men and 25.0% of women, and ~30% of women under 45 with partial or complete amenorrhoea, worse above 90% raw. Self-selected and cross-sectional, so it cannot isolate cooking from food choice, but the direction matches. [V] No controlled human calorimetry trial of cooked versus raw isocaloric diets was located.

8. Mixed dishes versus components: does combining foods change anything?

This is the most under-discussed cluster in the whole file, and it contains the single most inconvenient result.

8.1 The glycemic index of a mixed meal is not predictable from its parts, and tracks the meal’s total energy

Verdict: REAL, and it invalidates a common practice.

Flint A, Møller BK, Raben A, et al. “The use of glycaemic index tables to predict glycaemic index of composite breakfast meals.” Br J Nutr 2004;91(6):979-89. Randomised crossover, n = 28 healthy young men, 13 breakfast meals plus a reference meal, every meal standardised to 50 g available carbohydrate, venous sampling for 2 h. [V] https://pubmed.ncbi.nlm.nih.gov/15182401/

No association at all between GI predicted from published table values and GI measured in the same meal.
Meal energy content was the single best univariate predictor, R² = 0.93 (inverse, p < 0.001).
Fat content R² = 0.88 (inverse), the largest single contributor to that energy term but a weaker predictor than energy itself.
Carbohydrate as a percentage of energy, R² = 0.80 (positive).
The best multivariate model, fat plus protein, reached R² = 0.93, which does not improve on energy alone.
GI and insulinaemic index were uncorrelated.

Read that carefully. At fixed available carbohydrate, the measured glycemic index of a real breakfast is largely a readout of how much energy is on the plate, and fat is the main contributor to that energy. [Corrected on review: earlier text led with “driven by fat and protein” and put fat’s R² = 0.88 ahead of energy’s R² = 0.93. Energy is the better predictor; fat is best read as its dominant component, not as an independent masking agent acting on glucose.] Looking up the GI of the starch component and adding it up still produces a number with zero predictive validity for the meal you actually eat.

The counterweight, for honesty: earlier Wolever work reported mixed-meal GI within a few percent of predicted when carbohydrate foods of differing GI were combined without added fat and protein [R, Wolever & Jenkins 1985-1986, not re-fetched]. The two are reconcilable: GI addition survives carbohydrate-plus-carbohydrate and fails as soon as the meal contains realistic amounts of fat and protein. Real meals contain fat and protein.

8.2 Adding fat to a fixed-carbohydrate meal lowers the early peak; in type 1 diabetes on fixed insulin dosing it roughly doubles late hyperglycemia

Verdict: REAL in type 1 diabetes on exogenous insulin. UNKNOWN in people with endogenous insulin.

Wolpert HA, Atakov-Castillo A, Smith SA, Steil GM. “Dietary fat acutely increases glucose concentrations and insulin requirements in patients with type 1 diabetes.” Diabetes Care 2013;36(4):810-6. Randomised crossover, closed-loop insulin delivery over 18 h after dinner, n = 7 adults with type 1 diabetes. High-fat and low-fat dinners were identical in carbohydrate and protein, differing only in fat: 60 g vs 10 g. [V] https://pubmed.ncbi.nlm.nih.gov/23193216/

High-fat dinner required 12.6 ± 1.9 U vs 9.0 ± 1.3 U insulin, +40%, p = 0.01.
Despite the extra insulin, hyperglycemia AUC above 120 mg/dL was 16,967 ± 2,778 vs 8,350 ± 1,907 mg/dL·min, roughly double.

Scope this correctly. n = 7, type 1 diabetes, closed-loop insulin. What the trial demonstrates is that exogenous insulin delivered on a carbohydrate-indexed algorithm fails to track the delayed absorption of a high-fat meal. That is a result about insulin dosing, not about digestion in a person with a working beta cell. [Corrected on review: this section previously concluded that “anyone judging a meal by a 2-hour CGM trace is systematically rewarded for adding fat.” No evidence in this file shows that the 0-18 h glucose load rises with fat in people with endogenous insulin. The sentence has been withdrawn.]

The defensible pair of statements:

In T1D on fixed dosing, fat roughly doubles late hyperglycemia despite 40% more insulin. [V]
In non-diabetics, the 0-2 h flattening is real and the late cost is unmeasured. The narrow, supportable version of the warning is that the 2-hour metric alone cannot see a late cost if one exists, which is a limitation of the metric rather than a demonstration that the cost is there.

Partial prior art on the downstream question, which T1 in Part 3 proposes to test: a 2025 meta-analysis of CGM use in non-diabetic and prediabetic adults found that users increased the fat and protein share of their intake within arms, but the change was not significant versus controls. https://pubmed.ncbi.nlm.nih.gov/41015297/ So the drift has been observed once uncontrolled, and the controlled comparison is null.

8.3 Fat co-ingested with vegetables is not optional: without it, carotenoid absorption is roughly zero

Verdict: REAL, very large effect, and the dose-response does not saturate at realistic amounts.

Three independent human trials, all measuring carotenoids in the plasma chylomicron / triacylglycerol-rich lipoprotein fraction, which is the correct compartment for newly absorbed carotenoid rather than a steady-state plasma pool:

Brown MJ, Ferruzzi MG, Nguyen ML, et al. Am J Clin Nutr 2004;80(2):396-403. Randomised, n = 7, three identical spinach/romaine/tomato/carrot salads with dressings containing 0, 6 or 28 g canola oil, 12 h sampling, ≥2 wk washout. With fat-free dressing the appearance of alpha-carotene, beta-carotene and lycopene in chylomicrons was “negligible.” Full-fat was substantially greater than reduced-fat. [V] https://pubmed.ncbi.nlm.nih.gov/15277161/
Unlu NZ, Bohn T, Clinton SK, Schwartz SJ. J Nutr 2005;135(3):431-6. Two crossover postprandial studies, n = 11 each. Adding 150 g avocado to salad multiplied baseline-corrected AUC by 7.2× (alpha-carotene), 15.3× (beta-carotene) and 5.1× (lutein) versus avocado-free salad (p < 0.01). Adding avocado to salsa gave 4.4× lycopene and 2.6× beta-carotene (p < 0.003). Avocado fruit and avocado oil performed the same, so it is the lipid, not the fruit. [V] https://pubmed.ncbi.nlm.nih.gov/15735074/
Kopec RE, Cooperstone JL, Schweiggert RM, et al. “Modeling the dose effects of soybean oil in salad dressing on carotenoid and fat-soluble vitamin bioavailability in salad vegetables.” Am J Clin Nutr 2017. Randomised, n = 12 women, five salads with 0, 2, 4, 8 or 32 g soybean oil, NCT02867488. Across the whole 0-32 g range the relation to chylomicron AUC was linear for alpha-carotene, lycopene, phylloquinone and retinyl palmitate. Beta-carotene was linear across 0-8 g. Absorption of everything measured was highest at 32 g oil (p < 0.002). Large interindividual variation, with some individuals showing a negligible response at any dose. [V] https://pubmed.ncbi.nlm.nih.gov/28814399/

The practical reading is uncomfortable for the low-fat salad: 32 g of oil is about 280 kcal, and the absorption curve for several fat-soluble compounds was still rising there. There is no free lunch here. But the asymmetry at the bottom of the curve is the actionable part: going from 0 to 6-12 g of fat costs 55-110 kcal and moves absorption from approximately nothing to most of the available effect. Fat-free dressing on a vegetable salad is, on this evidence, close to throwing the fat-soluble micronutrients away.

Note the tension with §9. The same fat that rescues carotenoid absorption is the fat that flips a first-course salad from a 12% intake reduction to a 17% intake increase. The two findings are both real and they point in opposite directions. The resolution is dose: 6-12 g of oil is on the steep part of the absorption curve and is a rounding error on meal energy; 30-40 g is neither.

8.4 Fibre inside an intact food matrix does something that fibre added to a drink does not

Verdict: REAL for satiety and insulin, one of the oldest results in the field, and still under-cited.

Haber GB, Heaton KW, Murphy D, Burroughs LF. “Depletion and disruption of dietary fibre. Effects on satiety, plasma-glucose, and serum-insulin.” Lancet 1977;2(8040):679-82. n = 10 normal subjects, three test meals matched at 60 g available carbohydrate: intact apples, fibre-disrupted purée, fibre-free juice. [V] https://pubmed.ncbi.nlm.nih.gov/71495/

Juice could be consumed 11× faster than intact apples and 4× faster than purée.
With ingestion rate equalised, juice was significantly less satisfying than purée, and purée less than apples. So the satiety difference is not only an eating-rate artifact.
Plasma glucose rose to similar levels after all three. The difference was a striking rebound fall after juice, less after purée, absent after apples.
Serum insulin rose higher after juice and purée than after apples.

Three separate levers in one 1977 experiment: eating rate, satiety at matched rate, and reactive hypoglycemia. Note that the peak glucose was the same across all three arms. A peak-glucose readout would have scored these three foods as identical. That is a second independent reason to distrust CGM peak height as a meal-quality metric (see §8.2).

8.5 Supplemental viscous fibre added to a diet: real, and small

Verdict: SMALL.

Two Sievenpiper-group meta-analyses, both registered as NCT03257449:

Jovanovski E, et al. “Can dietary viscous fiber affect body weight independently of an energy-restrictive diet?” Am J Clin Nutr 2020;111(2):471-485. 62 trials, n = 3,877, ≥4 wk, ad libitum diets. Body weight −0.33 kg (95% CI −0.51 to −0.14), BMI −0.28, waist −0.63 cm, body fat null. GRADE moderate for weight, high for waist and body fat. [V] https://pubmed.ncbi.nlm.nih.gov/31897475/
Same group, Eur J Nutr 2021, 15 trials, n = 1,347, viscous fibre added to a calorie-restricted diet: weight −0.81 kg (−1.20 to −0.41), body fat −1.39%. [V] https://pubmed.ncbi.nlm.nih.gov/32198674/

So a fibre supplement buys roughly 0.3 kg free-living, 0.8 kg alongside a deficit. Contrast with the whole-food fibre outcome evidence in the rubric (Reynolds 2019: 15-30% lower all-cause mortality). The supplement does not reproduce the food. This is a clean example of a nutrient whose observational strength does not transfer to its isolated form, and it is a useful prior for every other “add the active fraction” product.

8.6 Protein dilution drives total intake: the protein leverage result

Verdict: REAL, asymmetric, and it is the best mechanistic bridge to the ultra-processing literature in the rubric.

Gosby AK, Conigrave AD, Lau NS, et al. PLoS One 2016;11(8):e0161003. Randomised, n = 22 lean healthy adults, three 4-day in-house ad libitum periods at 10%, 15% and 25% protein. Dropping protein from 25% to 10% produced +14% energy intake (p = 0.02) and a 6-fold rise in fasting plasma FGF-21. ACTRN12616000144415. [V] https://pubmed.ncbi.nlm.nih.gov/27536869/
Gosby AK, Conigrave AD, Raubenheimer D, Simpson SJ. “Protein leverage and energy intake.” Obes Rev 2014;15(3):183-91. Meta-analysis of 38 published ad libitum trials spanning 8-54% protein. Percent dietary protein was negatively associated with total energy intake (F = 6.9, p < 0.0001), and it did not matter whether carbohydrate or fat was the diluent (both p > 0.5). [V] https://pubmed.ncbi.nlm.nih.gov/24588967/

The honest counterweight, which is rarely quoted alongside the above. Martens EA, Lemmens SG, Westerterp-Plantenga MS. “Protein leverage affects energy intake of high-protein diets in humans.” Am J Clin Nutr 2013;97(1):86-93. n = 79, 12-day randomised crossover, 5% / 15% / 30% protein. Intake was significantly lower at 30% protein (7.21 MJ/d) than at 15% (9.62) or 5% (9.33), but the 5% arm did not differ from the 15% arm. The authors state plainly: “No evidence for protein leverage effects from diets containing a lower ratio of protein to carbohydrate + fat was obtained.” [V] https://pubmed.ncbi.nlm.nih.gov/23221572/

So the leverage is asymmetric in the best-powered single trial: raising protein reliably suppresses intake; lowering it below normal did not reliably raise intake at n = 79 over 12 days. Gosby’s 4-day n = 22 study found the increase; Martens’ 12-day n = 79 study did not. Anyone quoting protein leverage as settled is quoting one half of a contradiction.

The bridge. Martínez Steele E, Raubenheimer D, Simpson SJ, Baraldi LG, Monteiro CA. “Ultra-processed foods, protein leverage and energy intake in the USA.” Public Health Nutr 2018;21(1):114-124. NHANES 2009-2010, n = 9,042. Across quintiles of ultra-processed contribution, mean dietary protein density fell from 18.2% to 13.3% of energy, absolute protein intake stayed roughly constant, and total energy intake rose. Cross-sectional, so it is a consistency check rather than a test. [V] https://pubmed.ncbi.nlm.nih.gov/29032787/

This is worth taking seriously as a unifying account of the rubric’s finding #1. The NIH factorial trial decomposed the ultra-processing overeating effect into energy density (+662 kcal/d), hyperpalatability (+158) and processing per se (+128, n.s.). Protein dilution is a fourth candidate that was not a manipulated factor in that design.

[Corrected on review: this file previously said protein dilution “is confounded with energy density in almost every UPF trial, including Hall 2019.” That is wrong for Hall 2019 specifically. The presented diets in Hall 2019 were matched on protein as a percentage of energy, consumed protein came out at 14.0% (ultra-processed) vs 15.6% (unprocessed) and the difference was not significant, and the authors explicitly bounded any protein-leverage contribution at no more than about half of the observed intake excess. https://pmc.ncbi.nlm.nih.gov/articles/PMC7946062/ ]

So the live question is not “is protein dilution the unmeasured factor” but how much of the residual the small realised protein gap explains inside a menu that was already protein-matched on paper. Developed, with that scope, as T2 in Part 3.

9. Sauces and dressings as hidden calorie carriers

Nobody, as far as this search could establish, has published a decomposition of “what share of a restaurant meal’s calories come from sauce.” That specific number appears to be UNKNOWN. What does exist is better than a plausible-sounding estimate: a nationally representative measurement for one dish class, and a randomised trial in which the dressing is the only thing that varies.

9.1 National data: dressing supplies 4.7× the calories of the greens it dresses

Verdict: REAL, and larger than most people would guess.

Sebastian RS, Wilkinson Enns C, Goldman JD, et al. “Salad Consumption in the U.S., What We Eat in America, NHANES 2011-2014.” USDA FSRG Dietary Data Brief No. 19, Feb 2018, Table 1. Nationally representative, individuals aged ≥1 y. [V] https://www.ncbi.nlm.nih.gov/books/NBK589291/table/usdaddb19.tab1/

Mean amount and energy contribution in salads containing the ingredient:

Ingredient	% of salads	Mean amount	Mean kcal	kcal/g [C]
Lettuce / leafy greens	86	70 g (2 C)	22	0.31
Dressing, all types	86	33 g (2 TB)	103	3.12
Tomato	43	62 g	14	0.23
Carrot	34	26 g	14	0.54
Cheese	25	19 g	64	3.37
Meat, poultry, fish	19	67 g	115	1.72
Croutons	10	9 g	41	4.56
Nuts / seeds	8	14 g	84	6.00
Avocado	5	50 g	82	1.64

Derived, all [C] from the verified table:

Dressing carries 4.7× the energy of the greens it is poured on (103 vs 22 kcal).
A greens-plus-dressing salad is 82% dressing calories by energy.
Add the four common vegetables (tomato, carrot, onion, cucumber = 46 kcal) and the whole vegetable base is 68 kcal against 103 kcal of dressing, so a typical garden salad is still 60% dressing calories.
The highest energy densities on the entire salad plate are croutons (4.6 kcal/g) and nuts/seeds (6.0 kcal/g), both above dressing. Croutons are under-suspected: 9 g, a quarter cup, is 41 kcal, nearly twice the entire bed of greens.
The brief also reports mean energy contribution from salads among salad reporters of 234 kcal for adults ≥20 y [V, from the brief’s summary text; the specific table was not separately re-fetched]. On that denominator dressing is ~44% of all salad calories consumed in the United States.

9.2 The randomised version: dressing alone swings total meal intake by 29 percentage points

Verdict: REAL, and this is the cleanest causal demonstration in this section.

Rolls BJ, Roe LS, Meengs JS. “Salad and satiety: energy density and portion size of a first-course salad affect energy intake at lunch.” J Am Diet Assoc 2004;104(10):1570-6. n = 42 women, randomised crossover, lunch in the lab once a week for 7 weeks. Compulsory first-course salad, then ad libitum pasta. Salads varied in energy density (0.33, 0.67 or 1.33 kcal/g) and portion (150 or 300 g). The energy density was manipulated only by changing the amount and type of dressing and cheese. [V] https://pubmed.ncbi.nlm.nih.gov/15389416/

Versus no first course:

First-course salad	Change in total meal energy intake
Low energy density, 150 g	−7%
Low energy density, 300 g	−12%
High energy density, 150 g	+8%
High energy density, 300 g	+17%

The same vegetables, the same eating occasion, the same compulsory first course. Dressing and cheese alone move total meal intake across a 29-percentage-point range, and flip the sign. The larger the salad, the larger the effect in whichever direction the dressing points it. A big salad is a lever with a sign set by the dressing, not a uniformly good idea.

This is also the best available answer to “does a low-calorie starter help”: yes, and it is one of the few interventions where the food you add reduces the total.

9.3 Adding sauce to a meal increases intake, and not through liking

Verdict: REAL in the population studied, with an important caveat about direction of interest.

Appleton KM. “Increases in energy, protein and fat intake following the addition of sauce to an older person’s meal.” Appetite 2009;52(2):340-6. n = 28 older adults, within-subject, two meals with sauce and the same two meals without, on different occasions. [V] https://pubmed.ncbi.nlm.nih.gov/18840490/

Meals with sauce produced greater intake of energy, of energy from protein and of energy from fat (smallest t(27) = 2.13, p = 0.04).
No differences in pre-meal hunger or desire to eat, and none in post-meal pleasantness or familiarity (largest t(27) = 1.47, p = 0.15).
The effect was the same in participants who expected sauce to affect intake and those who did not.

The paper was written as an under-nutrition intervention for older adults, and for that purpose it is a positive result. Read as a lever it says something sharper: sauce increased intake without increasing rated liking. That rules out the obvious hedonic explanation and leaves lubrication, bolus formation and eating rate. n = 28 in older adults, so generalisation to a healthy adult is a hypothesis, not a finding.

9.4 The side dishes you did not count are bigger than the sauce

Verdict: REAL, and this is the highest-yield item in the section.

Urban LE, Dallal GE, Robinson LM, Ausman LM, Saltzman E, Roberts SB. “The accuracy of stated energy contents of reduced-energy, commercially prepared foods.” J Am Diet Assoc 2010;110(1):116-23. Bomb calorimetry. [V] https://pubmed.ncbi.nlm.nih.gov/20102837/

29 quick-serve and sit-down reduced-energy restaurant foods measured 18% above stated values; 10 supermarket frozen meals 8% above. Neither reached significance because of variability, but both exceeded laboratory measurement error.
Some individual restaurant items contained up to 200% of stated values.
Free side dishes raised provided energy to an average of 245% of the stated value for the entrée they accompanied.

That last number is the one to remember. The bread, the chips and salsa, the complimentary side: the uncounted accompaniment was worth more than the entrée itself, on average, in a directly calorimetered sample.

For context on the base rate, the same group’s larger study found stated calories to be accurate in aggregate: Urban LE, et al. “Accuracy of stated energy contents of restaurant foods.” JAMA 2011;306(3):287-93. 269 items from 42 restaurants, mean difference +10 kcal/portion (95% CI −15 to +34, p = 0.52). But 19% of items were understated by ≥100 kcal/portion, and the worst decile was understated by 258-289 kcal/portion on repeat sampling, i.e. the error is reproducible rather than random. [V] https://pubmed.ncbi.nlm.nih.gov/21771989/

And on the absolute scale: Roberts SB, Das SK, Suen VMM, et al. J Acad Nutr Diet 2016;116(4):590-8, bomb calorimetry of frequently ordered meals from non-chain restaurants in three US cities, found 1,205 ± 465 kcal/meal, not significantly different from equivalent large-chain meals (+5.1%, p = 0.41), with 92% of meals exceeding typical single-occasion energy requirements. [V] https://pubmed.ncbi.nlm.nih.gov/26803805/ Their Boston-only predecessor measured 1,327 kcal (95% CI 1,248-1,406) per meal. [V] https://pubmed.ncbi.nlm.nih.gov/23700076/

Practical ordering of the restaurant levers, by measured magnitude: portion size (~1,200-1,300 kcal baseline) > uncounted free sides (+145% of the entrée) > dressing/sauce on a salad (~100 kcal, up to 60-82% of that dish) > menu-stated calorie error (+10 kcal on average, +250 kcal in the worst decile).

10. The person is a bigger variable than the meal, but only for the endpoint nobody measures

Verdict: REAL, and the popular summary of this literature has the two endpoints swapped.

Not on the original topic list, but it changes how every other lever in this file should be interpreted, so it goes in.

Berry SE, Valdes AM, Drew DA, et al. “Human postprandial responses to food and potential for precision nutrition.” Nat Med 2020;26(6):964-973. PREDICT 1, n = 1,002 UK twins and unrelated adults, standardised test meals in clinic and at home, validated in a US cohort of n = 100. NCT03479866. [V] https://pubmed.ncbi.nlm.nih.gov/32528151/

Population coefficient of variation in response to identical meals:

Response	CV
Triglyceride	103%
Glucose	68%
Insulin	59%

Variance decomposition, which is the part that gets misreported:

	Postprandial lipemia	Postprandial glycemia
Meal macronutrients	3.6%	15.4%
Gut microbiome	7.1%	6.0%
Genetic variants	0.8%	9.5%
Model performance	r = 0.47	r = 0.77

For triglycerides the person beats the meal roughly 2:1. For glucose the meal beats the microbiome roughly 2.5:1. The widely circulated takeaway that “identical meals affect people so differently that food composition barely matters” is drawn from the lipemia column and then applied to glucose, where the data say the opposite. Meanwhile the endpoint where personalisation genuinely dominates, triglyceride, is the one no consumer CGM measures and no popular advice discusses.

Zeevi D, Korem T, Zmora N, et al. “Personalized Nutrition by Prediction of Glycemic Responses.” Cell 2015;163(5):1079-1094. n = 800, 46,898 meals under CGM, plus a 100-person validation cohort and a blinded randomised dietary intervention that lowered postprandial responses. [V] https://pubmed.ncbi.nlm.nih.gov/26590418/

Peer-review status is contested and should be stated. This line of work has drawn published editorial criticism twice: “Personalized nutrition by prediction of glycaemic responses: fact or fantasy?” Eur J Clin Nutr 2016 [V] https://pubmed.ncbi.nlm.nih.gov/27050901/ and, more bluntly, “Personalized nutrition by prediction of glycemic responses: garbage in → garbage out.” Am J Clin Nutr 2025 [V] https://pubmed.ncbi.nlm.nih.gov/39755431/. Both are editorials; neither abstract is indexed, so the specific arguments were not verified here.

The best counterweight, and it is recent and positive. Wang et al. “Quantification of personalized glycemic sensitivity to food and its potential for precision nutrition in a series of n-of-1 trials.” Am J Clin Nutr 2025. n = 176 healthy Chinese adults, CGM plus standardised meals, series of n-of-1 trials, plus a 30-person 3-month validation on alternating high- and low-carbohydrate diets. NCT05054153. [V] https://pubmed.ncbi.nlm.nih.gov/40754388/

Large between-person variation with high within-person consistency.
The personalised glycemic sensitivity index was reproducible over a 2-year interval, ICC 0.73, and showed temporal consistency ICC 0.88 in the validation study.

An ICC of 0.73 across two years is the load-bearing number. It means individual glycemic sensitivity is a stable trait, not measurement noise, which is what the “garbage in” critique would predict. So the personalisation signal survives, but its magnitude relative to the meal is endpoint-specific, and for glucose the meal still wins.

Consequence for this whole file. Every glycemic effect size quoted elsewhere in this document is a group mean sitting on top of a 68% between-person CV. A 20% iAUC reduction from food order or vinegar is real on average and can be zero or negative in a specific person. Any of these levers is cheap to test n-of-1 with a CGM, and the 2025 ICC data say such a test would be reproducible.

11. Attention, memory and variety: the cost of a distracted meal lands at the next meal

Not on the original list. Included because it is the largest effect in this file that is not about the food at all, and because the effect is displaced in time in a way that makes it nearly invisible to self-observation.

11.1 Distraction

Verdict: REAL, and the delayed component is roughly twice the immediate one.

Robinson E, Aveyard P, Daley A, et al. “Eating attentively: a systematic review and meta-analysis of the effect of food intake memory and awareness on eating.” Am J Clin Nutr 2013;97(4):728-42. 24 experimental studies, inverse-variance meta-analysis of standardised mean differences. [V] https://pubmed.ncbi.nlm.nih.gov/23446890/

Manipulation	Effect on intake	SMD (95% CI)
Eating while distracted	immediate intake up	0.39 (0.25, 0.53)
Eating while distracted	later intake up	0.76 (0.45, 1.07)
Enhancing memory of what was eaten	later intake down	0.40 (0.12, 0.68)
Removing visual information about amount eaten	immediate intake up	0.48 (0.27, 0.68)
Enhancing awareness of food being eaten	no effect	0.09 (−0.42, 0.35)

Three things worth extracting.

The delayed cost is about double the immediate cost. Eating in front of a screen makes you eat somewhat more now and considerably more at the next eating occasion. Almost nobody attributes an afternoon snack to what they were doing during lunch, because the causal link is separated by hours.
The mechanism is memory, not willpower. Enhancing memory of the prior meal reduces later intake with a comparable effect size, and removing visual information about how much was eaten increases immediate intake. Consistent with the memory account, not with a self-control account.
Simply “being aware” does nothing (CI spans zero). Instructions to “pay attention to your food” without a memory component are the folklore version. Non-trivial: the popular mindful-eating advice is closest to the null arm.

Confirmation in a targeted RCT. Robinson E, Kersbergen I, Higgs S. “Eating ‘attentively’ reduces later energy consumption in overweight and obese females.” Br J Nutr 2014;112(4):657-61. n = 48, between-subjects. Fixed lunch eaten either with audio instructions directing attention to the food or with a neutral audiobook; ad libitum snack intake measured in a later session the same day. Snack intake was approximately 30% lower in the focused-attention condition, statistically significant. Notably the authors found only limited evidence that the effect ran through measured meal memory, so the mechanism is supported at the meta-analytic level but was not cleanly demonstrated in the single trial. [V] https://pubmed.ncbi.nlm.nih.gov/24933322/

11.2 Variety within a meal

Verdict: REAL but SMALL relative to portion size, on the best modern measurement.

The classic result is Rolls BJ, Rowe EA, Rolls ET, et al. “Variety in a meal enhances food intake in man.” Physiol Behav 1981;26(2):215-21, which is where the often-quoted “variety increases intake ~30%” figure comes from. [V] for existence; [R] for the 30% figure, which was not re-verified and should not be quoted without checking. https://pubmed.ncbi.nlm.nih.gov/7232526/

The same lab re-measured it against portion size in a modern design. Roe LS, et al. “Variety and portion size combine to increase food intake at single-course and multi-course meals.” Appetite 2023;191:107089. Two randomised crossover experiments, weekly for four weeks, n = 42 (dishes served simultaneously) and n = 49 (served as three sequential courses). Variety was low (three bowls of the favourite dish) vs high (three different main dishes); portion was 450 vs 600 g. [V] https://pubmed.ncbi.nlm.nih.gov/37844692/

Experiment 1: variety and portion did not interact (p = 0.72); both increased intake independently. Variety +15 ± 7 g. Portion +57 ± 7 g.
Sensory-specific satiety was smaller at high-variety meals in the sequential experiment (p = 0.001), which is the proposed mechanism: variety and large portions produce more intake for the same or a smaller hedonic decline.

Portion size outweighed variety by roughly 3.8:1 in grams here. So “eat a monotonous meal” is a real but minor lever, and it is dominated by simply serving less. Worth stating because the variety effect is often quoted at a magnitude the modern replication does not support.

12. Preloading: water, soup, and protein

Verdict: REAL acutely for low-energy-dense food preloads (~80-130 kcal per meal). SMALL and fragile for water in older adults. For water in young adults, SMALL and timing-dependent, with no weight data at all. [Corrected on review: this previously read “FOLKLORE for water in young adults”, which rested on a single 30-minute-preload study and omitted two positive immediate-preload trials. See §12.2.] REAL and strong for whey on glycemia, NULL for whey on weight at n = 79 over 12 weeks.

12.1 Soup and low-energy-density preloads

Flood JE, Rolls BJ. Appetite 2007;49(3):626-34, PMC2128765. Within-subject crossover, n = 60 normal-weight adults, five lunches (four soup forms plus a no-soup control), compulsory preload, ad libitum meal 15 min later. Soup at 0.33 kcal/g: 129 kcal / 350 mL for women, 172 kcal / 475 mL for men. [V]

Entrée intake 654-704 kcal in the soup conditions versus 936 ± 48 kcal control.
Total lunch including the soup’s own calories: 804-855 vs 936 ± 48 kcal. So intake fell by 81 to 132 kcal (9-14%) even after paying for the soup.
Soup form (chunky vs pureed) had no effect.
Discrepancy flagged: the paper’s headline “20% (134 ± 25 kcal)” does not reconcile with its own Table 3. 134/936 = 14.3%, and the table means give 112.5 kcal = 12.0%. The 20% figure is what gets quoted. Use 9-14%.

The mechanistic key is that it is the water inside the food, not the water in the glass. Rolls BJ, Bell EA, Thorwart ML. AJCN 1999;70(4):448-55, n = 24 lean women, isoenergetic 270 kcal preloads 17 min before lunch: lunch intake 289 kcal after chicken rice soup versus 396 kcal after the same casserole served alongside 356 g of water and 392 kcal after the casserole alone. Water as a beverage did not affect satiety; the identical water incorporated into the food did. PMID 10500012. [V]

The salad-preload trials, which double as the food-order test:

Rolls BJ, Roe LS, Meengs JS. J Am Diet Assoc 2004;104(10):1570-6, n = 42: low-energy-dense salads cut total meal energy 7% (150 g) and 12% (300 g); high-energy-dense salads increased intake 8% and 17%. [V] See §9.2.
Roe LS, Meengs JS, Rolls BJ. Appetite 2012;58(1):242-8, n = 46, the one trial that varies only the order. Ad libitum pasta, accompanied four times by an identical 300 g / 100 kcal / 0.33 kcal/g salad, served 20 min before or with the pasta, each once compulsory and once ad libitum, plus a no-salad control. PMID 22008705, open access PMC3264798. [V]
- Adding a fixed salad cut meal energy 11%, 57 ± 19 kcal, F(4,176) = 3.08, p = 0.018.
- Timing had no significant effect on energy intake across participants (p > 0.05); it interacted only with flexible-restraint score.
- Pasta intake was 378 ± 17 kcal in both ad libitum salad conditions and 439 ± 17 kcal in both compulsory conditions, against 531 ± 24 kcal with no salad. The pairing is by compulsory versus ad libitum, not by before versus with.
- Ad libitum salad consumption was lower than compulsory and did not significantly reduce energy intake.
- Serving the salad first did increase vegetable consumption by 23%.
- Authors’ conclusion: “To moderate energy intake, maximizing the amount of salad eaten may be more important than the timing of consumption.”
Williams RA, Roe LS, Rolls BJ. Obesity 2014;22(2):318-24, n = 46: a salad preload cut intake 123 ± 18 kcal, but raising the main course’s energy density added back 153 ± 19 kcal. The preload effect is fully erasable by the rest of the meal. [V]
Flood-Obbagy JE, Rolls BJ. Appetite 2009;52(2):416-22, n = 58: a whole apple preload cut total lunch energy 15% (187 ± 36 kcal); applesauce and juice did not. [V] Same matrix finding as Haber 1977 in §8.4.
Shafaie 2015, Appetite 89:196-202, n = 75: after a high-fat soup, PROP non-tasters overate by 11 ± 5% while supertasters under-ate 26 ± 10% (p < 0.01). Taste genotype flips the sign. [V]

THE KEY NEGATIVE. Rolls BJ, Roe LS, Beach AM, Kris-Etherton PM. Obes Res 2005;13(6):1052-60. One-year RCT, n = 200 adults with overweight or obesity on an energy-restricted exchange diet, randomised to one soup/day, two soups/day, two high-energy-dense snacks/day, or no special food. PMID 15976148. [V]

Arm	Weight loss at 1 year
No special food	8.1 ± 1.1 kg
Two soups/day	7.2 ± 0.9 kg
One soup/day	6.1 ± 1.1 kg
Two snacks/day	4.8 ± 0.7 kg

p = 0.006 across arms. Two things need saying about how this trial is usually read, including in earlier versions of this file. [Corrected on review.]

“No special food” is not a no-intervention control. All four arms were on the same energy-restricted exchange diet with counselling. The contrast is “diet alone” versus “diet plus a prescribed food”, not “soup versus nothing”.
The comparison the trial was designed to make is soup versus snack, and the omnibus p = 0.006 is what supports it.

The correct summary: adding two low-energy-dense soups per day did not improve on the same diet without them, and adding two energy-dense snacks made it clearly worse. The widely repeated “50% greater weight loss with soup” is stated in the paper relative to the snack arm only. https://pubmed.ncbi.nlm.nih.gov/15976148/

12.2 Water preloading

Dennis EA, et al. Obesity 2010;18(2):300-7, PMC2859815. n = 48, ages 55-75, BMI 25-40, 12 weeks, hypocaloric diet plus 500 mL water before each of three daily meals versus diet alone. [V]

Weight loss about 2 kg greater in the water group; water beta = -0.87 (p < 0.001) vs -0.60 (p < 0.001), a “44% greater rate of decline.”
No between-group 95% CI or p-value appears in the abstract.
Coherence problem inside the paper: the acute mechanism decayed. At baseline the water arm ate 498 ± 25 vs 541 ± 27 kcal (p = 0.009); at 12 weeks 480 ± 25 vs 506 ± 25 kcal, p = 0.069, no longer significant.
Unblinded, no attention control.

Davy BM, et al. J Am Diet Assoc 2008;108(7):1236-9, n = 24, mean age 61.3, BMI 34.3. 500 mL water 30 min before ad libitum breakfast: 500 ± 32 vs 574 ± 38 kcal, p = 0.004 (~74 kcal, ~13%). [V]

The claimed age dependence rests on one study, and it is confounded with timing. Van Walleghen EL, Orr JS, Gentile CL, Davy BM. Obesity 2007;15(1):93-9, PMID 17228036. Same protocol in both age groups, 375 mL (women) / 500 mL (men) 30 min pre-meal. [V, re-verified directly]

Group	With water	Without water	p
Young, n = 29, ages 21-35	913 ± 54 kcal	892 ± 51 kcal	0.65, null and numerically higher with water
Older, n = 21, ages 60-80	624 ± 56 kcal	682 ± 53 kcal	0.02 (~58 kcal, ~8.5%)

Fullness ratings rose with water in all subjects (p = 0.01); only intake diverged. The authors note the older-group effect was “caused primarily by the reduction in meal energy intake after water consumption in older men,” so the effective cell is smaller than 21. n = 50 total, single meal, one lab.

Two positive trials in young adults, omitted from earlier versions of this file. [Corrected on review: the previous FOLKLORE verdict for young adults was reached by citing only the Van Walleghen null.]

Corney RA, Sunderland C, James LJ. Randomised crossover, n = 14 lean young males, 568 mL water immediately before an ad libitum meal versus no water: energy intake 2,551 kJ to 1,967 kJ, about -23%, p < 0.001. https://pubmed.ncbi.nlm.nih.gov/25893719/ [V, review-sourced]
Jeong 2018, Clin Nutr Res 7(4):291-6, n = 15: also positive for a pre-meal water load on intake. https://e-cnr.org/DOIx.php?id=10.7762/cnr.2018.7.4.291 [V, review-sourced]

The moderator may be timing, not age. Van Walleghen and Davy both used a 30-minute gap and found nothing in young adults; Corney used immediately before and found about -20% in the same age band. Plain water empties from the stomach fast enough that a 30-minute gap plausibly discards most of the volume, which is a mechanistic reason to expect exactly this pattern. Nobody has run the 2x2 of age by timing, and sex is a third uncontrolled candidate (Corney was males-only; the Van Walleghen older-group effect was driven by men).

What is still missing is any weight endpoint in young adults. Corney and Jeong are single-meal intake studies. Every water-preload weight trial (Dennis, Parretti, Davy) is in older or middle-aged adults. “Water preloading does not help young adults lose weight” is untested, not refuted.

The only preload RCT with a real attention control. Parretti HM, et al. Obesity 2015;23(9):1785-91, ISRCTN33238158. Two-group RCT, UK primary care, n = 84, 12 weeks. Intervention: 500 mL water 30 min before main meals. Control: asked to imagine their stomach was full. Adherence verified by 24 h total urine collections. [V]

Unadjusted -1.3 kg (95% CI -2.4 to -0.1, p = 0.028).
Adjusted for ethnicity, deprivation, age and gender: -1.2 kg (95% CI -2.4 to +0.07, p = 0.063). Crosses zero. The authors call it “preliminary.”

Pooled: null. Chen QY, Khil J, Keum N. Nutrients 2024;16(7):963, PMID 38612997, systematic review and meta-analysis of 8 RCTs in adults with overweight or obesity: [V]

Contrast	Effect	I²
Water intake, body weight	-0.33 kg (95% CI -1.75 to +1.08)	78%
BMI	-0.23 (-0.55 to +0.09)	0%
Waist	+0.05 cm (-1.20 to +1.30)
Water replacing sugar-sweetened beverages	-0.81 kg (-1.66 to +0.03)	2%
Water replacing artificially sweetened beverages	+1.82 kg (0.97 to 2.67)

Authors: “water intake may not significantly impact adiposity.” Note the last row: swapping diet drinks for water made weight go up, which is the mirror image of the usual advice and worth a separate look.

Bracamontes-Castelo 2019, Nutr Hosp 36(6):1424-9, narrative SR of 6 RCTs ≥12 wk: the most effective strategy was replacing caloric beverages with water, not preloading; evidence for weight loss rated low to moderate, explicitly insufficient for a recommendation. [V] And a 2025 review by the Davy group itself concedes “the optimal timing and volume of water intake remains unknown.”

12.3 Whey and protein preloading

Verdict: the single best-supported glycemic intervention in this entire file, and null for weight.

Ma J, et al. Diabetes Care 2009;32(9):1600-2, PMC2732158. Crossover, n = 8 diet-controlled T2D. 55 g whey in 350 mL soup 30 min before a potato meal, versus whey in the meal, versus no whey. [V]

Glucose iAUC 363.7 ± 64.5 (preload) vs 406.3 ± 85.9 (in meal) vs 734.9 ± 98.9 (no whey), p < 0.005. About -50% for the preload.
Gastric emptying T50: 87.3 ± 5.4 vs 53.0 ± 8.3 vs 39.0 ± 6.2 min, p < 0.0005. More than doubled.
n = 8: treat -50% as an upper bound.

Jakubowicz D, et al. Diabetologia 2014;57(9):1807-11. Randomised open-label crossover, n = 15, randomised by coin flip, 50 g whey in 250 mL water immediately before a high-GI breakfast: glucose -28% across 180 min; insulin +105%, C-peptide +43%, total GLP-1 +141%, intact GLP-1 +298%. Funded in part by the Milk Council. NCT01571622. [V]

Meta-analyses:

Smedegaard S, et al. AJCN 2023;118(2):391-405. 16 randomised crossover trials, n = 244. Whey premeal vs non-active comparator: peak glucose -1.4 mmol/L (95% CI -1.9 to -0.9); glucose AUC -0.9 SD (-1.2 to -0.6). High certainty. Slowed gastric emptying and raised peak insulin, high certainty; raised GLP-1, low certainty. Dose 4-55 g, with meta-regression showing dose correlated with the effect. Authors: “long-term effects await future clinical trials.” [V]
Chiang SW, et al. Nutr Res 2022;104:44-54, T2D only, 5 RCTs, n = 134: glucose at 60 min -2.67 mmol/L (-3.62 to -1.72); at 120 min -1.59 (-2.91 to -0.28). Authors: “the level of certainty of current evidence is not high enough.” [V]

THE KEY NEGATIVE. Watson LE, et al. Diabetes Obes Metab 2019;21(4):930-8, PMID 30520216. 12-week, single-blind, randomised, placebo-controlled, n = 79 T2D (baseline HbA1c 6.6%). 17 g whey + 5 g guar in 150 mL (n = 37) versus flavoured placebo (n = 42), 15 min before two meals daily. [V, re-verified directly]

Gastric emptying slower (p < 0.01) and postprandial glucose lower (p < 0.05), sustained at 12 weeks.
HbA1c -1 mmol/mol (-0.1%) versus placebo (p < 0.05). Statistically significant, clinically trivial.
Verbatim: “There were no differences in energy intake, body weight, or lean or fat mass between the groups.”

Do not confuse this with Jakubowicz 2017, J Nutr Biochem 49:1-7 (n = 56, 12 weeks, three isocaloric whole-diet arms differing in breakfast composition), which reported HbA1c -0.89 ± 0.05% and weight loss 7.6 ± 0.3 kg in the whey breakfast arm. That is a whole-diet manipulation, not a preload added to an existing diet, and the standard errors (±0.3 kg across 19 people over 12 weeks) are implausibly tight. The design-matched preload test is Watson 2019, and it found zero weight effect.

12.4 The unifying finding across §§1, 3 and 12

Postprandial glycemia moves reliably. A weight effect has not been detected in trials that were mostly not powered to detect one. [Corrected on review: this section previously read “Body weight does not [move]” and claimed that “every design-matched randomised trial … is null on weight”. The first overstated the evidence and the second is false. See the positive-trial list below.]

Intervention	Trial	Weight or intake result	Power for 1-2 kg
Food order	Tricò 2016, 8 wk, n = 17	-1.9 vs -2.0 kg, no between-arm difference	none
Food order	Shukla 2023, 16 wk, n = 39	between-group weight p = 0.625, intake change p = 0.205	none, pilot
Soup preload	Rolls 2005, 1 yr, n = 200	soup did not beat the same energy-restricted diet without it	adequate for this contrast
Water preload	Parretti 2015, 12 wk, n = 84, attention control	-1.2 kg, CI -2.4 to +0.07 after adjustment	borderline
Water, pooled	Chen 2024, 8 RCTs	-0.33 kg (-1.75 to +1.08)	CI still admits -1.75 kg
Whey preload	Watson 2019, 12 wk, n = 79	no difference in energy intake, weight, lean or fat mass	marginal
Salad order	Roe 2012, n = 46	timing null; only the amount of salad mattered	single meal, no weight endpoint

Touhamy 2025 has been removed from this table. [Corrected on review.] It was a 12-day eucaloric provided-meal crossover in which both arms lost 1.97 kg because both were fed the same energy. A eucaloric design cannot produce a between-arm weight difference, so “identical weight change in both arms” is a property of the design and carries no information about food order.

The positive trials, which earlier versions of this file did not list at all.

Protein plus fibre preload, 12 weeks, double-blind, n = 206: -3.3 kg vs -1.8 kg, p < 0.05. Funded by Beachbody, which sells the product. https://pmc.ncbi.nlm.nih.gov/articles/PMC9178960/
Khezri 2018, apple cider vinegar added to energy restriction, n = 39, 12 weeks: roughly 2 kg additional loss. Open-label. https://www.sciencedirect.com/science/article/abs/pii/S1756464618300483
Yabe 2019, meal-sequence cluster-randomised, 6 months, n = 42 across three clusters: weight fell significantly more in the meal-sequence cluster. Industry co-authors (Kao), open-label, cluster design, kg values not retrievable.
Kondo 2009, vinegar, n = 155 analysed: -1.2 and -1.9 kg, dose-ordered, nominally double-blind. All authors employed by the vinegar manufacturer, no CIs anywhere, energetically incoherent, fully rebounded within 4 weeks of stopping.

The defensible generalisation, replacing the old one: every well-controlled, non-industry, attention-controlled trial in this space is null or borderline; positive trials do exist, and each one carries a funding conflict, an open-label design, or a clustering or blinding defect. That is weaker than “it never works”, and unlike the old claim it is falsifiable: one adequately powered, independent, blinded replication of any of the four above would overturn it.

13. Sleep, stress and exercise interactions

Only randomised or controlled-laboratory evidence is included.

13.1 Sleep extension: the largest well-measured behavioural effect in this file

Verdict: REAL and causal, and it is bigger than everything else here.

Tasali E, Wroblewski K, Kahn E, Kilkus J, Schoeller DA. “Effect of Sleep Extension on Objectively Assessed Energy Intake Among Adults With Overweight in Real-life Settings.” JAMA Intern Med 2022;182(4):365-74, NCT02253368. Single-centre RCT, intention to treat. n = 80 adults aged 21-40, BMI 25.0-29.9, habitual sleep under 6.5 h/night. Two weeks. The intervention was one individualised sleep-hygiene counselling session targeting 8.5 h in bed, with no diet or activity prescription, free-living. Energy intake was measured objectively as total energy expenditure by doubly labelled water plus change in body energy stores (home weights plus DXA), with actigraphy verification. [V]

Sleep duration +1.2 h/night (95% CI 1.0 to 1.4, p < 0.001)
Energy intake -270 kcal/day (95% CI -393 to -147, p < 0.001)
Sleep change correlated with intake change, r = -0.41 (-0.59 to -0.20)
No treatment effect on total energy expenditure, so the deficit went to weight

-270 kcal/day from a single counselling session, derived from doubly labelled water rather than food diaries. Caveats: two weeks, n = 80, single centre, only in short sleepers with overweight, no independent replication located.

Apply the same discount this file applies to Kondo. [Added on review.] Kondo is discounted here for being single-source, unreplicated and internally incoherent. Tasali is n = 80, two weeks, one centre, unreplicated, and its headline number is not a measurement of intake: energy intake is inferred as doubly-labelled-water expenditure minus the change in body energy stores from DXA and home scales, so it inherits the error of both components over a 14-day window in which body-composition change is small relative to measurement noise. Tasali is much better designed than Kondo and has no product attached, but “-270 kcal/day” is a single unreplicated estimate from an inferred endpoint and should not sit at the top of the summary table as though it were settled.

The sleep-restriction side, for direction consistency:

St-Onge MP, et al. AJCN 2011;94(2):410-6, randomised-order inpatient crossover, n = 30, 4 h vs 9 h time in bed for 5 nights: day-5 intake +296 kcal (p = 0.023), driven by fat (+20.7 g) and saturated fat (+8.7 g). RMR unchanged (p = 0.136), TDEE unchanged (p = 0.832). [V]
Markwald RR, et al. PNAS 2013;110(14):5695-700, 14-15 day inpatient whole-room calorimetry: TDEE rose ~5% with insufficient sleep but intake exceeded it, especially after dinner, producing +0.82 ± 0.47 kg in 5 days. Ghrelin, leptin and PYY signalled excess stores, so the intake was not homeostatically driven. Recovery sleep: -0.03 ± 0.50 kg. [V]
Spaeth AM, et al. AJCN 2014;100(2):559-66, n = 44, 5 nights at 4 h: 532.6 ± 295.6 kcal consumed in the 22:00-03:59 window; men increased more than women (d = 0.62). [V]

13.2 Acute exercise does not trigger compensatory eating

Verdict: REAL for the acute window.

Schubert MM, Desbrow B, Sabapathy S, Leveritt M. Appetite 2013;63:92-104. Meta-analysis, 29 studies / 51 trials, 30-120 min at 36-81% VO2max, ad libitum meals 0-2 h post-exercise. [V]

Absolute energy intake: ES 0.14 (95% CI -0.005 to 0.29). Trivial, CI crosses zero.
Relative energy intake (intake minus exercise cost): ES -1.35 (95% CI -1.64 to -1.05). Large deficit.

People do not acutely eat back the calories. Dorling J, et al. Nutrients 2018;10(9):1140, narrative review, confirms the short-term deficit without compensatory appetite and reports that adiposity and sex do not modify the response, with hormone responses “equivocal”. [V] Scope caveat: these are hours-long windows. Neither source establishes absence of compensation over weeks.

13.3 Stress

Verdict: SMALL and suggestive; not a randomised stress manipulation.

Epel E, Lapidus R, McEwen B, Brownell K. Psychoneuroendocrinology 2001;26(1):37-49. n = 59 healthy premenopausal women, each attending a stress session and a control session on separate days. High cortisol reactors consumed more calories on the stress day than low reactors, but ate similar amounts on the control day. High reactors ate more sweet food across both days; increases in negative mood correlated with greater consumption. [V]

Effect size in kcal is not in the abstract. Cortisol-reactor status is measured, not randomised, so this is a within-subject moderator analysis rather than a randomised stress trial with a primary intake endpoint. No direct replication was located.

Part 2. The most surprising results

Five, ordered by how much they should change what you do.

S1. Across five different “eat this first” interventions, glucose moves reliably and no trial has been built that could see a 1 kg weight effect

[Corrected on review. The previous headline was “glucose moves and weight never does”, supported by the claim that “every design-matched randomised trial with a weight or energy-intake endpoint is null on weight”. That claim is false as stated and the trials behind the true part of it were not powered for the effect they are being said to rule out.]

Food order, vinegar, soup preload, water preload and whey preload are marketed on the same premise. What the literature actually supports:

The glycemic effects are real and replicated. That part is unchanged.
The weight evidence is thin in both directions. The design-matched trials are Tricò 2016 (n = 17), Shukla 2023 (n = 39) and Watson 2019 (n = 79). Detecting 1 kg needs on the order of 400 per arm [C]. Not detected is not the same as null.
Touhamy 2025 does not belong in this argument at all: it was eucaloric with provided meals, so it could not have produced a between-arm weight difference.
The between-group numbers, not the within-arm ones. Shukla 2023’s between-group tests are weight p = 0.625 and intake change p = 0.205. The p = 0.649 figure quoted for years in this file, including in earlier versions of this section, is a within-arm p. The satiety p = 0.65 is from Mishra 2023, a kiwifruit-for-cereal substitution study, where it means satiety was preserved.
The one-year soup RCT is not “no special food beats soup”. All four arms were on an energy-restricted counselled diet, and the comparison the trial was built for was soup versus snack (omnibus p = 0.006).
Positive trials exist: a double-blind protein plus fibre preload (n = 206, -3.3 vs -1.8 kg) that is Beachbody-funded, Khezri 2018 (n = 39, ~2 kg, open-label), Yabe 2019 (cluster, industry co-authors) and Kondo 2009 (manufacturer-run).

The honest headline: these are well-evidenced glycemic-variability tools; as weight tools they are unproven in either direction, and every trial that reports a weight benefit has a funding, blinding or clustering defect. See §12.4.

S2. In a real mixed meal, glycemic index is a readout of the meal’s energy, and a 2-hour glucose trace cannot see whatever happens afterwards

[Corrected on review on two points: the old headline led with fat rather than energy, and it asserted a late glucose cost in people without diabetes that no source here supports.]

Flint 2004, n = 28, thirteen breakfasts all standardised to 50 g available carbohydrate: no association at all between GI predicted from tables and GI measured. The best univariate predictor of measured GI was meal energy content, R² = 0.93, with fat R² = 0.88 as the main contributor to that energy and a fat plus protein multivariate model reaching the same R² = 0.93 that energy alone reaches. So the finding is “GI tracks how much energy is on the plate”, and fat is the largest component of that, not an independent masking agent.

Wolpert 2013, closed-loop insulin delivery over 18 hours, n = 7 with type 1 diabetes, dinners identical in carbohydrate and protein and differing only in fat (60 g vs 10 g): the high-fat dinner needed 40% more insulin and, despite that, produced roughly double the hyperglycemia area. That is a demonstration that carbohydrate-indexed exogenous insulin cannot track delayed fat-modified absorption. [Withdrawn on review: the previous sentence “a CGM-guided eater is being systematically rewarded for adding fat” generalised an n = 7 type 1 result to people with endogenous insulin, for which no data is cited anywhere in this file.]

What survives: adding fat lowers the 0-2 h curve, and the 2-hour metric by itself cannot see a late cost if there is one. Whether a late cost exists in non-diabetics is UNKNOWN and directly testable (T1). Partial prior art: a 2025 meta-analysis found CGM users raised fat and protein share within arms but not versus controls. https://pubmed.ncbi.nlm.nih.gov/41015297/

Haber 1977 adds the third leg: whole apples, purée and juice produced the same peak glucose and completely different insulin and rebound behaviour, so peak height would have scored three very different foods as identical. See §8.1, §8.2, §8.4.

S3. The calorie content of a nut is a property of the nut and your molars jointly, and the label is wrong by 24% of the label value

Four USDA controlled-feeding studies with complete faecal and urine collection: almonds deliver 4.6 kcal/g against an Atwater-predicted 6.05, so 28 g is 129 kcal, not 170. State the denominator: that is -24% of the label, or equivalently Atwater over-predicting by 32% of the measured value, which is Novotny’s own framing. Same gap, two denominators. https://pubmed.ncbi.nlm.nih.gov/22760558/ Walnuts -21%, cashews -16%, pistachios -5%. Then the mechanism is demonstrated directly: grinding the almond into butter restores the full Atwater value (6.53 kcal/g, indistinguishable from prediction, p = 0.08), and roasting alone moves it partway because roasted almonds fracture at lower force. Chewing 10 times instead of 40 measurably raises faecal fat excretion.

This is one of the very few places where a nutrition label is systematically and knowably wrong in a direction you can exploit, and where “chew more” has a measured energy consequence rather than a satiety story. It is also a warning: the same logic says the calorie content of anything you pulverise, blend or purée is higher than the intact version. See §4.1.

S4. Distracted eating costs about twice as much at the next meal as at the one you are eating

Robinson 2013, meta-analysis of 24 experimental studies: distraction raises immediate intake at SMD 0.39 (0.25, 0.53) and later intake at SMD 0.76 (0.45, 1.07). Enhancing memory of the prior meal reduces later intake at SMD 0.40; removing visual information about how much was eaten raises immediate intake at SMD 0.48. And simply “being aware” while eating does nothing (SMD 0.09, CI spans zero), which is the arm closest to popular mindful-eating advice.

The mechanism is memory, not self-control, and the cost is displaced by hours, which is precisely why nobody attributes an afternoon snack to what they were watching during lunch. See §11.1.

S5. Two of the most-repeated “small tricks” are measurement artifacts, and their real drivers are boring

“A calorie in the morning burns hotter.” The thermic effect of food genuinely looks 1.6x higher at breakfast than lunch and 2.4x higher than dinner (p = 0.022). Adjust for the underlying circadian rhythm in resting metabolic rate and it is flat: 54.1 vs 49.5 vs 49.1 kcal, p = 0.680. The companion 4-week isocaloric crossover with doubly-labelled-water expenditure found identical weight loss (3.33 vs 3.38 kg, p = 0.848) and identical total expenditure (p = 0.184) [R, not [V]: the abstract carries no p-values, so these two figures were never confirmed against a fetched source; downgraded on review]. Front-loading works, and it works entirely through appetite. See §5.3.

“Use a smaller plate.” A pre-registered RCT is null (d = 0.07, mean difference 19.2 kcal, CI -76.5 to 115.0), the independent meta-analysis gives SMD -0.18 with a CI touching zero, and the decomposition is explicit: d = 0.70 when food is self-served, d = 0.03 when the portion is held constant. Plate size is the portion-size effect wearing a costume, the underlying literature has 12 PubMed-indexed retractions from one lab, and the Cochrane SMD of 0.38 that people cite pools portion and package size with tableware whose own strands are rated low and very low quality. Meanwhile the thing that is real, portion size itself, is worth +423 kcal/day with no habituation over 11 days. See §4.4, §4.5.

Honourable mentions. The largest PAH exposure from grilling is inhaled, not eaten (benzo[a]pyrene 1,431 ng/Sm³ in charcoal smoke against a 1 ng/Sm³ air guideline, roughly 1,400x, versus 0.62-1.25 µg/kg in the meat, within EU limits). Fat-free dressing on a salad produces negligible carotenoid absorption while adding avocado gives 15.3x for beta-carotene, and the dose-response to oil is still linear at 32 g. Cooling rice saves about 1% of the meal’s calories, against a measured microbiome-associated swing in energy harvest of ~150 kcal/day, 40 to 60 times larger and of uncertain sign. And swapping artificially sweetened drinks for water made pooled weight go up 1.82 kg (0.97, 2.67).

Part 3. Novel theory candidates

Falsifiable hypotheses that combine levers, each with the experiment that would kill it. None of these is established; they are the interesting shapes left after the above.

Each now carries an explicit prior and a prior-art line, added on review. The original versions were presented without stated confidences and without checking whether the experiment had already been done or half-done. Several had, and two of the eight are not really theories at all.

	Claim in one line	Confidence it survives its own kill condition	Novel?
T1	CGM feedback on 0-2 h glucose selects for higher-fat meals that are worse over 24 h	~25%	partly, see the 2025 CGM meta-analysis
T2	Protein dilution is the missing factor in the UPF decomposition	~20%	no, Hall 2019 already bounded it
T3	Order effects are volume effects, separable in one 2x2	~65%	design is novel, result is half-known
T4	Metabolizable energy is a function of particle size at swallow, for whole meals	~15%	novel, and contradicted for cereals
T5	Sleep restriction attenuates every small behavioural lever	~50%	novel as an interaction test
T6	There is a computable optimum fat dose on a salad, 6-12 g	~55%	dose-finding, not a theory
T7	Single-meal reductions below ~150 kcal do not survive 24 h	~45% as a dose-response, much lower as a floor	partly known
T8	Trait glycemic sensitivity predicts who responds to sequencing	~60% that r > 0.4 appears, ~30% that it means anything beyond baseline dependence	weak

T1. The fat-masking hypothesis: CGM-guided eating selects for meals that are worse over 24 hours

Confidence: ~25%. [Revised down on review.] Two of the three legs are weaker than the original write-up implied. Flint’s fat term (R² = 0.88) sits under the energy term (R² = 0.93), so “fat masks glucose” is better read as “energy predicts glucose”. And Wolpert is n = 7 with type 1 diabetes on closed-loop insulin, so the late-glucose leg is a statement about exogenous dosing algorithms; the non-diabetic 0-18 h data does not exist. What remains is a plausible metric-versus-goal divergence with one weak empirical leg.

Prior art, partial. A 2025 meta-analysis of CGM use in non-diabetic and prediabetic adults found users raised fat and protein share within arms but not versus controls. https://pubmed.ncbi.nlm.nih.gov/41015297/ So the behaviour half of T1 has been looked at once and came out null on the controlled comparison. The relevant T1D review of fat and protein effects on postprandial glucose is Diabetes Care 2015, https://pubmed.ncbi.nlm.nih.gov/25998293/, which is again a T1D-dosing literature.

Claim. Because mixed-meal glycemic response falls with meal energy, of which fat is the largest component (Flint), and because fat delays absorption enough to defeat carbohydrate-indexed insulin in T1D (Wolpert), a feedback loop that scores meals on 0-2 h glucose may progressively select higher-fat, higher-energy-density meals. The metric and the goal would then diverge.

Prediction. In a randomised crossover at fixed available carbohydrate with fat at 10, 30 and 60 g, 0-2 h iAUC will fall monotonically with fat while 0-8 h and 0-18 h iAUC rise, producing a crossover point somewhere between 3 and 5 hours.

Test. n ≈ 20 healthy adults, three arms, CGM plus venous sampling, 18 h observation. Pre-register the crossover time. Kill condition: if 0-18 h iAUC does not rise with fat in non-diabetic subjects, the Wolpert result is specific to exogenous insulin dosing and the hypothesis fails.

Second-order test, which is the interesting one. Randomise n ≈ 60 CGM users to be shown either 2-hour or 8-hour post-meal curves for 8 weeks, and measure the energy density and fat content of the meals they converge on. Prediction: the 2-hour group’s self-selected diet drifts higher in fat and energy density. Nobody appears to have asked whether the feedback signal changes the diet in the direction of the metric rather than the outcome.

T2. How much of the ultra-processing residual survives protein-matching on paper

Confidence: ~20%, and this is NOT novel. [Substantially rewritten on review.] The original version said “protein dilution is confounded with energy density in essentially every ultra-processing trial, including Hall 2019.” That is wrong about Hall 2019. Hall’s presented diets were matched for protein as a percentage of energy; consumed protein came out at 14.0% (ultra-processed) vs 15.6% (unprocessed), a non-significant difference, and the authors explicitly bounded protein leverage at no more than about 50% of the observed effect. https://pmc.ncbi.nlm.nih.gov/articles/PMC7946062/ So the question has already been asked and partially answered by the trial the hypothesis proposed to replicate.

What is left is a test of the residual, not of an unmeasured factor. The realised 1.6-point protein gap in Hall 2019 arose from what people chose to eat off a protein-matched menu. The live question is whether enforcing equal consumed protein density (rather than equal presented protein density) removes a detectable share of the intake excess, and whether it lands anywhere near Hall’s own 50% upper bound.

Claim. Enforcing equal consumed protein density between an ultra-processed and an unprocessed arm removes a measurable but minority fraction of the intake excess.

Prediction. An ultra-processed arm with consumed protein density held at the unprocessed arm’s level will show an intake excess between 50% and 100% of Hall’s +508 kcal/day, i.e. most of the effect is not protein.

Test. Replicate Hall 2019’s inpatient crossover with a third arm: ultra-processed, energy-density-matched, and protein-density-matched. n ≈ 20, 14 days per arm. Kill condition: if the protein-matched ultra-processed arm still overeats by ≥400 kcal/day, protein leverage is not the mechanism and the residual belongs to eating rate or hyperpalatability.

The honest weakness up front: protein leverage is itself asymmetric in the best-powered trial. Martens 2013 (n = 79, 12 days) found raising protein suppressed intake but lowering it below normal did not raise intake. https://pubmed.ncbi.nlm.nih.gov/23221572/ Since the ultra-processing case is dilution downward from ~15.6% to ~14.0%, which is exactly the arm Martens found inert and a far smaller step than Martens’ 15% to 5%, the prior on T2 should be low. Between Hall’s ≤50% bound and Martens’ null on downward dilution, this is a narrow confirmatory experiment, not a new mechanism.

T3. Order effects are volume effects, and the two can be separated in one 2x2

Confidence: ~65%. [Added on review.] This is the best-supported candidate in Part 3 because three of the four cells already exist and point the right way. Roe/Rolls 2012 varied only order at fixed salad and found no order effect on energy intake while salad quantity did move it. https://pubmed.ncbi.nlm.nih.gov/22008705/ Kuwata 2016 supplies the other half of the dissociation: order moves gastric emptying and glycemia without an intake endpoint. The novelty is therefore the experiment, not the claim; the claim is already the most likely reading of the existing data.

Claim. Vegetables-first does not work through sequence. It works because putting vegetables first reliably causes more vegetables to be eaten (+23%, Roe 2012), and low-energy-dense volume is what displaces energy. Sequence per se moves glycemia, via gastric emptying, and nothing else.

Prediction. A 2x2 factorial of order (before vs with) by vegetable quantity (150 vs 300 g), with the quantity consumed enforced, will show a main effect of quantity on energy intake with no main effect of order, and simultaneously a main effect of order on glucose iAUC with no main effect of quantity. A double dissociation in one experiment.

Test. n ≈ 48, within-subject, four lab lunches, CGM plus ad libitum main course weighed. Kill condition: any significant order x intake main effect, or any significant quantity x glycemia effect at matched order, falsifies the clean dissociation. Roe 2012 already provides three of the four cells with a null on order; nobody has run it with both endpoints together.

T4. The oral-processing ledger: metabolizable energy is a function of particle size at swallow, for whole meals and not just nuts

Confidence: ~15%, and there is a direct contradiction. [Revised sharply down on review.] Wisker 1996 compared coarse versus finely ground rye in a controlled human balance study and found energy digestibility of 91.2% vs 91.6%, a difference of 0.4 percentage points. https://pubmed.ncbi.nlm.nih.gov/8632222/ T4’s prediction is a 3-8% of intake difference from a chewing manipulation. Wisker is roughly an order of magnitude below that, on the exact food class (intact cereal grain) T4 generalises to.

Nuts are plausibly special. The Grundy and Ellis work on almond cell walls attributes the nut effect to lipid encapsulation: almond lipid is stored inside intact parenchyma cells and simply is not released unless the cell wall is ruptured. Cereal and legume starch is not encapsulated the same way, is far more susceptible to amylase once hydrated and cooked, and has much less energy per gram locked behind a wall. The mechanism that makes the almond result large is not obviously present in a mixed meal, which is why T4 is now the lowest-confidence entry here.

Claim. The nut result (almonds -24%, restored to label value by grinding, faecal fat rising when chews drop from 40 to 10) is a special case of a general law: measured metabolizable energy of a food is a decreasing function of the particle size distribution at the moment of swallowing, mediated by intact plant cell walls. If so, the Atwater system misestimates every structurally intact food, and the error is a property of the eater as much as the food.

Prediction. In a controlled-feeding study of a whole mixed meal built around intact plant tissue (legumes, whole grains, seeds), a randomised chewing manipulation (fixed low vs high chew count) will produce a measurable difference in faecal energy of the order of 3-8% of intake, and the difference will correlate with the D50 particle size of the expectorated bolus.

Test. Beltsville-style crossover, n ≈ 18, complete faecal and urine collection, bomb calorimetry, plus laser diffraction on expectorated boluses. Kill condition: no dose-response between bolus particle size and faecal energy. Wisker 1996 arguably already fired the kill condition for cereals (0.4 percentage points from coarse versus fine grinding), so the strongest remaining version of T4 is restricted to lipid-encapsulating foods: nuts, seeds and possibly whole legumes, not “whole meals” in general.

Why it matters if true. It would mean the ~7% of intake already lost in stool (Basolo 2020) is partly under behavioural control, on the same order as the entire measured effect of every “trick” in Part 1 combined.

T5. Sleep is the upstream gate, and every behavioural lever should be tested under sleep restriction

Confidence: ~50%. [Added on review.] The main-effect legs are solid: Nedeltcheva 2010 (n = 11 crossover, 14 days at 5.5 h vs 8.5 h in bed) showed sleep restriction shifts the composition of weight loss toward lean tissue and raises hunger, https://pubmed.ncbi.nlm.nih.gov/20921542/, and Markwald 2013 showed the excess intake under restriction is not homeostatically driven. What is untested is the interaction, which is what T5 actually claims.

Powering warning. Detecting an interaction of the size proposed (a 50% attenuation of an SMD ~0.45 main effect) needs roughly four times the sample of the main effect. The n ≈ 24 originally proposed here is not enough; n ≈ 48 or more completers is the realistic floor for a 2x2 within-subject design, and more if the preload effect is at the lower end of its CI.

Claim. Sleep extension produced -270 kcal/day measured by doubly labelled water from a single counselling session, which is larger than portion-size manipulation, chewing, eating rate, plate size, capsaicin and every preload in this file. Markwald showed the excess intake under sleep restriction is not homeostatically driven (ghrelin, leptin and PYY all signalled excess stores). If sleep restriction disables the satiety signal that the small levers act through, then those levers should be weaker, not stronger, in the sleep-deprived, which is the population most likely to try them.

Prediction. The eating-rate effect (SMD 0.45) and the low-energy-dense preload effect (~11%) will both be attenuated by 50% or more under 5 nights of 4-5 h sleep versus 8.5 h, at matched food.

Test. Inpatient crossover, n ≈ 48+ completers (see the powering warning above; the original n ≈ 24 was for a main effect, not an interaction), 2x2: sleep (restricted vs extended) by preload (300 g salad vs none), ad libitum main course. Kill condition: no interaction. This has never been run, and it is the single most decision-relevant experiment in this file, because it would tell you whether to spend effort on the mechanics at all before fixing sleep.

T6. There is a computable optimum fat dose on a salad, and it is around 6-12 g

Confidence: ~55%, and this is a dose-finding study, not a theory. [Reclassified on review.] That an interior optimum exists follows almost trivially from two verified monotone curves running in opposite directions on the same food; the only genuinely uncertain part is where it sits, and 6-12 g is a guess dressed as a prediction. Nothing here would be falsified by an optimum at 4 g or at 20 g. Keep it as the cheapest experiment in Part 3, not as a hypothesis.

Claim. Two verified curves point in opposite directions on the same food. Carotenoid and fat-soluble vitamin absorption rises linearly with oil to 32 g with no saturation, and fat-free dressing gives negligible absorption. Meanwhile the same dressing determines whether a first-course salad cuts total meal intake 12% or raises it 17%. The net-benefit function therefore has an interior maximum.

Prediction. A dose curve at 0, 6, 12 and 32 g of oil on a fixed 300 g salad, measuring both chylomicron carotenoid AUC and ad libitum intake at the same meal, will show most of the absorption benefit captured by 12 g while the intake penalty is still small, with net utility peaking between 6 and 12 g.

Test. n ≈ 20, four-arm crossover, dual endpoints in the same session. Kill condition: if intake compensation is already complete at 12 g, the optimum moves down and low-fat dressing plus a separate fat source elsewhere in the meal wins. The two literatures have never been run in the same experiment.

T7. Compensation of single-meal intake reductions is a dose-response, not a floor

Confidence: ~45% for the dose-response version, much lower for the floor version. [Reframed on review: the original stated a 150 kcal threshold, which is a sharp discontinuity nobody has evidence for, inferred from one trial (Williams 2014) in which the compensation happened within the same meal.]

Prior art. Almiron-Roig 2013 systematically reviewed preload compensation and found it is partial, variable and continuous, not switch-like: compensation indices scatter across the range rather than clustering above and below a threshold. https://pubmed.ncbi.nlm.nih.gov/23815144/ A floor at a specific kilocalorie value is therefore the less likely shape.

Reframed claim. The proportion of a single-meal energy reduction that survives 24 hours increases continuously with the size of the reduction and with whether the manipulation changed the food matrix, rather than switching on above some kilocalorie threshold. Behavioural nudges sit at the low-persistence end because they are small, not because they are behavioural.

Original claim, retained for the record. Williams 2014 is the tell: a salad preload cut intake 123 ± 18 kcal, and raising the main course’s energy density added back 153 ± 19 kcal in the same meal. Nearly every “trick” in Part 1 lands in the 60-130 kcal per meal range, and almost none has been measured beyond the test meal. The hypothesis is that single-meal reductions below roughly 150 kcal are fully compensated within 24 hours unless the intervention changes the food matrix itself (energy density, particle size), because matrix changes persist into subsequent meals while behavioural nudges do not.

Prediction, revised. In a 24-hour ad libitum residential design with the reduction dosed at several magnitudes, the fraction of the deficit surviving to 24 h will rise monotonically with the size of the initial reduction, with no detectable breakpoint at 150 kcal, and matrix manipulations will sit above behavioural ones at matched initial deficit.

Test. n ≈ 30, residential, 24-h weighed intake, at least four deficit magnitudes rather than the original four conditions, fitted as a dose-response. Kill condition, revised: a genuine segmented-regression breakpoint would support the original floor version; a flat or noisy relationship between initial deficit and surviving deficit kills both versions. Portion size already provides the positive control from the other direction: Rolls 2007 showed no compensation at all over 11 days for a +50% portion, which is a matrix-scale manipulation.

T8. Personal glycemic sensitivity predicts who responds to the sequencing levers

Confidence: ~60% that r > 0.4 shows up, ~30% that it means anything beyond baseline dependence. [Split on review.] The correlation is close to guaranteed for an uninteresting reason: regression to the mean. People with higher baseline postprandial excursions have more room to fall, so any absolute reduction will correlate with baseline by construction, and “personal glycemic sensitivity” is largely a restatement of baseline excursion size. The file already notes the food-order effect is large in T2D and not significant in healthy people, which is the same axis.

Required design fix: the outcome must be the relative reduction (percentage of each person’s own baseline iAUC), and the analysis must adjust for baseline or use a change-from-baseline model that is not mechanically coupled to it. Otherwise the study cannot distinguish a trait moderator from arithmetic.

Claim. Postprandial glucose to identical meals has a CV of 68% between people, and personal glycemic sensitivity is a stable trait, ICC 0.73 over two years. Every effect size in §§1, 2 and 3 is a group mean sitting on top of that. The food-order effect scales with baseline hyperglycemia (large in T2D, not significant in Kuwata’s healthy arm), which is the same axis.

Prediction, revised. Personal glycemic sensitivity measured at baseline will predict the relative (percentage-of-baseline) magnitude of the food-order and vinegar effects with r > 0.4 after adjustment for baseline iAUC, and the lowest-sensitivity tertile will show no significant effect. An unadjusted correlation on absolute reductions does not count as support.

Test. n ≈ 60, each subject performs the standardised sensitivity protocol, then four crossover meals (order and vinegar, each with and without). Kill condition: no correlation between trait sensitivity and intervention response, which would mean the responder variance is occasion-level noise and n-of-1 self-testing is useless for these levers.

Part 4. Open questions

Ordered by how cheap the answer would be relative to how much it would change behaviour.

Does eating with hands versus utensils change intake? One n = 11 study exists and it measured glycemic index, not intake. Given that bite size (-8%) and eating rate (SMD 0.45) both move intake, and implements plainly determine bite size, this is a trivially cheap experiment that nobody has run.
Does food temperature change intake or satiety? No human trial manipulating meal temperature and measuring intake was located under five query formulations. The adjacent finding is that ambient temperature moves food choice without moving amount (p = 0.120 for intake). A genuine hole.
What share of a restaurant meal’s calories is sauce? Nobody has published the decomposition. The closest measurements are a national dataset for one dish class (dressing = 103 kcal, 4.7x the greens it dresses, ~44% of all salad calories) and a calorimetry finding that free side dishes bring provided energy to 245% of the entrée’s stated value. A bomb-calorimetry study that separated sauce from substrate across restaurant categories would be a straightforward extension of work Roberts’ group has already done.
Is the whole-meal analogue of the nut metabolizable-energy effect real? See T4. This is the largest potential unmeasured quantity in the file.
Does the second-meal effect survive to any clinical endpoint? Every trial in §2.5 measured glucose, insulin, short-chain fatty acids or appetite over 2 to 12 hours. No trial has followed a second-meal design to weight, HbA1c or anything else. And the one trial that tested cold-stored grains in a second-meal design found no second-meal effect, which is a direct negative for the popular version.
Is retrograded RS3 from domestic cooling equivalent to supplemental RS2? All the supportive meta-analytic evidence uses 10-45 g/day of supplemental resistant starch. Cooling one serving of rice yields 1-2 g. The scale gap is 10-40x and nobody has bridged it.
Does the microbiome-associated energy-harvest swing respond to fibre in either direction? Jumpertz measured ~150 kcal/day associated with a 20% Firmicutes shift, but caloric load was manipulated and the microbiome was only observed. This is the largest number anywhere in this file and its sign is unknown.
Why did water replacing artificially sweetened beverages produce weight gain (+1.82 kg, 0.97 to 2.67)? That result sits inside an otherwise-null meta-analysis and is the opposite of standard advice. It could be confounding by indication, reverse causation, or real. It is unexplained.
Does the TRE blood-pressure effect exist at anything like the reported size? Sutton 2018’s -11/-10 mmHg at n = 8 is implausibly large and the best-powered follow-up found -4 mmHg diastolic only. The eucaloric design is the strongest in chrononutrition and the effect size is the least believable.
Is the HbA1c effect of vinegar anywhere near the pooled estimates? The originating lab measured -0.16%; pooled estimates run -0.70% to -1.53% with I² of 96-99% and detected publication bias (Egger p = 0.037). A single adequately powered, properly blinded 12-week trial would settle a literature that currently contains 38 studies of which 32 are rated low quality.
Marinade chemistry moves different heterocyclic amines in opposite directions. Salmon 1997 cut PhIP by 92-99% while raising MeIQx more than tenfold, with total Ames mutagenicity higher in marinated samples at 40 min. “Marinate to reduce carcinogens” is not internally consistent even at the chemistry level, and no human outcome has ever been measured.
Is dietary AGE exposure measurable on a common scale at all? The database everything rests on reports arbitrary ELISA units and contains an internal anomaly (raw chicken reading only ~30% below boiled). Modern LC-MS/MS values are not convertible. Until that is fixed, “low-AGE diet” is not a well-defined exposure, and the largest prospective test (1.22 million people, 23,229 cancers) is already null.

Hardest open questions (from review)

These are the questions an adversarial read of this file could not answer from the file itself. They are listed because each one, if answered, would change a verdict above rather than add a footnote.

What sample size would any of these levers need to demonstrate a 1 kg weight effect, and has any trial in this file ever come close? On a ~5 kg standard deviation of weight change the answer is on the order of 400 per arm [C]. The largest design-matched trial here is n = 200. Every “null on weight” verdict in this file is really “underpowered”.
What does T2 add beyond Hall 2019? Hall matched presented protein, realised 14.0% vs 15.6% (NS), and bounded protein leverage at ≤50% of the effect. If the answer is “it enforces consumed rather than presented protein density”, say so; if there is no other answer, T2 is a confirmatory replication and should not be listed as a novel theory.
Why was Corney 2016 absent from the water-preload section? A 14-person crossover showing -23% single-meal intake in lean young males is the single most directly contradictory result to a FOLKLORE verdict on young adults, and it is on PubMed. What else did the search strategy that missed it also miss?
Why were within-arm p-values quoted and between-arm p-values never? Shukla 2023’s -64.8 kcal/d, p = 0.649 was presented as “the decisive number” for four sections. The between-group figure is p = 0.205. Is this an isolated slip or a systematic preference for whichever number supports the section’s verdict?
Is there any 0-18 h glucose data on added fat in people with endogenous insulin? The whole fat-masking argument rests on a single n = 7 type 1 study with closed-loop insulin. If no non-diabetic dataset exists, T1’s first leg is unsupported and should be labelled as such.
Why lead with fat’s R² = 0.88 when energy’s R² = 0.93 is in the same table? The stronger predictor was demoted in favour of the more surprising one. How often does that happen elsewhere in this file?
Why does Tasali get a top-of-table headline while Kondo gets a discount? Both are single, unreplicated, single-centre. Tasali’s intake figure is inferred from DLW minus DXA-derived energy stores, not measured. The differences (no product, better design) are real, but the discount should be applied to both or argued explicitly.
Was “the pooled vinegar weight effect is carried by a single trial” ever verified? A 2025 pool contains 10 RCTs, 8 nominally positive. The claim appears to have been inferred rather than checked. What other single-sentence attributions in this file were never sourced?
How does T4 survive Wisker 1996? Coarse versus finely ground rye differed by 0.4 percentage points in energy digestibility (91.2% vs 91.6%) against T4’s predicted 3-8% of intake. Either T4 is restricted to lipid-encapsulating foods, or it is falsified for cereals already.
Is “no special food beats soup” the comparison Rolls 2005 was built to make? All four arms were energy-restricted with counselling, and the designed contrast was soup versus snack (omnibus p = 0.006). The “surprise” as originally written may be an artifact of reading the wrong column.
Is the water-preload moderator age, timing, or sex? Van Walleghen and Davy used a 30-minute gap; Corney used immediately-before. Corney was males only. The 2x2x2 has never been run, and the current verdict picks age by default.
Is Wolpert a fact about physiology or about an algorithm? Closed-loop insulin on a carbohydrate-indexed dosing rule failing to track fat-delayed absorption is an engineering result. What would it take to separate the two?
Can a eucaloric provided-meal design contribute anything to a weight question? Touhamy 2025 was listed as evidence of no weight effect. It cannot be. Are there other design-impossible results being counted as null findings?
Which standard applies to pooled meta-analyses, Hollands or Robinson 2023? Hollands’ SMD 0.38 is dismissed for pooling across manipulation types; Robinson 2023’s SMD -0.709 across 14 heterogeneous studies is quoted as -235 kcal/day without the same caveat. One standard has to be picked.
Which decision does any of T1 through T8 actually change? If T3 resolves, you serve more salad and stop caring about order, which the file already recommends. If T5 resolves, you fix sleep first, which the file already recommends. T5 is the only one whose answer would reallocate effort, and only if the interaction is large. The rest are interesting, not decision-relevant.

Provenance and reliability

Method. Five parallel research agents plus direct retrieval, 2026-09-06. The session’s WebSearch budget (200 calls) was exhausted partway through, so the majority of primary-source retrieval ran through the Europe PMC REST API, the PubMed E-utilities API, PMC, J-Stage and publisher pages fetched directly. That is why abstract-level coverage is strong and paywalled full-text point estimates are sometimes missing. Publisher sites for NEJM, Cell Press, Wiley, JAMA Network, ScienceDirect, AJCN and Cochrane returned HTTP 403 to automated fetching; every number attributed to those journals came from an indexed abstract, a PMC deposit or an author manuscript.

Citation marks used throughout. [V] = fetched or searched in this session. [R] = recalled, not re-verified. [C] = arithmetic performed here on [V] numbers. Explicit UNVERIFIED tags mark claims that could not be checked.

Second-hand provenance, disclosed. One research agent delegated part of its work further and initially over-stated its own verification. It subsequently re-pulled the five most load-bearing previously-unchecked claims directly from PubMed (Van Walleghen 2007, Rolls 2005, Darzi 2014, Chen 2024, Watson 2019) and all five matched exactly, including verbatim quotes. The following items in §§3 and 12 remain agent-sourced without a second check: Johnston 2004/2005/2008/2009, Brighenti 1995, Freitas 2021/2022, Cheng 2020, the Shahmohammadi 2026 umbrella table, Kondo’s group-by-group kilogram values (the n per arm, the “1-2 kg” text, Mizkan employment and the lactate placebo were verified from the PDF), the Flood & Rolls table-versus-abstract discrepancy, Rolls 1999/2004, Williams 2014, Flood-Obbagy 2009, Shafaie 2015, Ma 2009/2015, Chiang 2022, Jakubowicz 2017 and Bracamontes-Castelo 2019. Every spot-check came back exact, so confidence is reasonably high, but these were not witnessed.

Known unretrieved values, listed so they are not silently treated as verified: the Acta Diabetol 2026 food-order meta-analysis I² values, publication-bias assessment and the content of its August 2026 Correction; Shishehbor 2017’s I² and Egger tests; the Robinson 2014 eating-rate pooled effect in kcal; Holden 2016’s d = 0.70 / d = 0.03 as printed in the original rather than as quoted by Kosīte; Raatz 2016’s potato resistant-starch values (press release, not the paper); the Jakubowicz 2013 kilogram figures; per-method folate retention percentages; Felton 1994’s microwave pre-treatment fold reductions; Dennis 2010’s between-group CI; the exact CIs for Qin 2021’s per-114 g dose-response; the chopped-almond metabolizable energy value from Gebauer 2016.

Two corrections to commonly circulated framings, made here:

The canonical “lente carbohydrate” second-meal paper is Jenkins 1982, Am J Clin Nutr 35(6):1339-46, not BMJ 1982.
Koschinsky 1997 reports ~10% of ingested advanced glycation end products absorbed and only ~30% of the absorbed fraction excreted, i.e. roughly two-thirds of what is absorbed is retained, not one-third as the claim is usually restated.

Relationship to the rubric. Nothing here contradicts /research/health/rubric/. Two things extend it. First, energy density remains the master variable, and several levers in this file (dressing, preload, portion, plate size) turn out on inspection to be energy density or portion size in disguise. Second, the rubric’s finding that ultra-processing per se was not significant (+128 kcal/day, p = 0.065) in the NIH factorial acquires a candidate fourth explanatory factor here: protein dilution, which was not a manipulated factor in that design and falls from 18.2% to 13.3% of energy across ultra-processed quintiles in NHANES. [Corrected on review: Hall 2019 itself matched presented protein density, realised 14.0% vs 15.6% (NS), and bounded protein leverage at no more than about half the effect. https://pmc.ncbi.nlm.nih.gov/articles/PMC7946062/ So this is a test of a residual that is already partly bounded, not a new factor.] Developed with that scope as T2 in Part 3.

/research/health/rubric/, what is in the food
/research/health/composition/, what separates two meals with identical panels
/research/health/patterns/, ultra-processing and diet-quality scoring