In plain terms

The claim being assessed: In common human obesity, a diet high in glycemic load is the primary modifiable upstream cause: its hormonal response, centred on a raised insulin-to-glucagon ratio, shifts fuel partitioning toward adipose storage and lowers fuel availability to other tissues, which increases hunger and lowers energy expenditure, so that fat storage drives positive energy balance rather than following from it; after an adaptation period of weeks, a lower-glycemic-load diet should therefore raise energy expenditure at equal energy and protein intake and reduce ad libitum intake.

Where the evidence stands: Γ — Plausible mechanism. There is a believable biological story and partial support, but the causal step has not been demonstrated.

What this mechanism does: This makes the process worse or keeps it going, but something else starts it.

Strongest evidence for it: The endocrine link (carbohydrate/glycemic load → insulin and insulin-to-glucagon ratio) reaches Η on two independent causal lines. The downstream chain remains Γ: isocaloric fat-balance predictions are negative in short studies and untested at long duration; expenditure gains after adaptation are contested and lack a fat-loss correlate; GI effects on hunger, intake and regain are positive in defined contexts but insulin mediation failed its direct test; the responder rule failed its prespecified replication; and primacy has no necessity or rate-limitation evidence. Weakest-link, primary-subclaim and scope-matched aggregation all return Γ.

What decisive evidence would look like

In plain terms, researchers would change the proposed cause, confirm it changed as intended, and measure a meaningful health outcome while checking other explanations. The result should be able to support or count against the claim.

Relative explanatory fit: Alternative ranked ahead

What this means for carbohydrates, insulin, and obesity

Meals with different carbohydrate or glycemic loads clearly change insulin and related hormones. The stronger popular claim is that this hormonal change is the primary cause of common obesity because it makes fat storage drive hunger and positive energy balance. The review finds that the first hormonal link is real, but the full causal chain and its claimed primacy have not been demonstrated in people.

Implications for health and treatment: This leaves room for different dietary approaches to help different people, especially when they are nutritionally adequate and sustainable. It does not make carbohydrate restriction a universally superior treatment, nor does it show that body weight is simply the result of willpower or one hormone. In practice, obesity treatment can involve nutrition, physical activity, sleep, medicines, mental-health support, management of related disease, and, for some people, bariatric procedures—choices that need an individual assessment.

Limits for individual care: Short feeding studies, changes in appetite, and an insulin response after a meal cannot by themselves establish a long-term, universal mechanism. The evidence does not identify a validated 'insulin responder' test that can prescribe one diet for an individual. People using glucose-lowering drugs, insulin, or medicines that affect appetite should not make large dietary changes from this page without discussing safety and dose adjustment with the clinician managing their care.

Medical information notice: This review provides general health information, not medical advice. It cannot diagnose a condition or replace an individual discussion with a qualified clinician.

This assessment rests on 44 sources and 31 individually checked claims. 31 of those claims were verified against the original paper.

A low grade is not an accusation. It means the causal work has not been done yet, which is the ordinary state of most actively studied mechanisms in biology.

At a glance

Sub-claims assessed separately

ID	Sub-claim	Type	Level	Pivotal card
SC1	Increasing digestible carbohydrate or glycemic load relative to an appropriate comparator increases postprandial insulin secretion and the insulin-to-glucagon ratio in humans.	mechanism	Η	—
SC2	The carbohydrate-driven insulin response shifts fuel partitioning toward adipose storage enough to cause net fat accumulation in humans, so that storage precedes and drives positive energy balance.	mechanism	Γ	Card ↗
SC3	Carbohydrate-driven insulin responses lower late-postprandial fuel availability, which increases hunger and subsequent voluntary energy intake.	mechanism	Γ	—
SC4	After an adaptation period of weeks, lowering carbohydrate (glycemic load) at equal energy and protein intake produces a sustained, physiologically important increase in energy expenditure or greater fat loss.	sufficiency	Γ	—
SC5	Baseline insulin secretion identifies people who benefit preferentially from lowering glycemic load.	generalization	Γ	—
SC6	The carbohydrate-insulin pathway is the primary, necessary, or rate-limiting cause of common human obesity rather than one contributor among multiple regulators of energy intake, expenditure, and partitioning.	primacy	Γ	Card ↗

Visual Evidence Cards · Pivotal Subclaims

Pivotal subclaim illustrations

Dark-theme simulated-micrograph renders and concise causal evidence summaries for the pivotal subclaims that govern this review's grade. Click any card to inspect the master 2400×2400 render.

Master 2400×2400 render ↗

Master 2400×2400 render ↗

Figures

Explanatory figures summarizing this review. They carry no evidential weight of their own; where a figure and the review differ, the review governs.

Full assessment

From here the review is written for a technical reader: caps and thresholds, competing models, the independence basis, and the citation ledger.

HSEF v5.4 evidence review · original 2026-09-28 · red-team synthesis 2026-09-30

Provenance: 44 source records · 31 registered load-bearing claims · 31 verified · 0 unverified · 0 contradicted · 30 Tier A sources · gate passed

Adversarial review: Generated by Sol 6.1; adversarially reviewed by Claude Opus 5.5 (Anthropic) on 2026-10-14. 3 corrections applied (see ADVERSARIAL REVIEW AND APPROVAL). Status: Approved as revised.

AT A GLANCE

The endocrine premise is solid. Carbohydrate raises insulin, insulin coordinates postprandial storage, and chronic hyperinsulinemia can drive adiposity in engineered mice. The grade is Γ because the hypothesis claims much more than that. It claims that this pathway, driven by high-glycemic-load diets, is the primary cause of common obesity, through fat storage that drives hunger and lowers energy expenditure. The human tests of that chain are negative, mixed, or disputed on analysis. Some positive signals do exist, notably for energy expenditure in longer feeding studies and for glycemic index effects on late-postprandial hunger and weight regain. Their proponents report them, and this edition gives them more space than the original did.

LAY SUMMARY

The claim is that eating a lot of quickly digested carbohydrate raises insulin, insulin pushes fuel into fat cells, the rest of the body is left short of fuel, hunger rises and metabolism slows, and that this chain, rather than simply eating too much, is the main cause of common obesity.

This review assigns Γ, meaning mechanistically plausible, partial system support. The first step is firmly established: carbohydrate raises insulin. Experiments in mice and human genetic studies suggest that high insulin can add to fat gain in some settings. Some human studies found that low-glycemic-index meals left people less hungry hours later, and one large European trial found slightly less weight regain on low-glycemic-index diets. But carefully controlled human diets that sharply lower insulin have not produced the extra fat loss the model predicts. Studies of energy use disagree with one another, and in one inpatient study people ate less on a high-carbohydrate diet that raised their insulin more.

Both sides of this debate accept that stored fat can only grow if more energy comes in than goes out. What they disagree about is what drives that imbalance. Current evidence supports carbohydrate-driven insulin as one factor that can amplify fat storage or appetite in some people and situations. It does not show that this factor is the primary cause of common obesity. A large feeding trial designed to test one key prediction has been completed but its main results are not yet published.

CLAIM RESTATEMENT AND SCOPE

Wording submitted

Dietary carbohydrate drives insulin secretion that promotes fat storage and increases hunger, and this mechanism is the primary cause of obesity rather than excess energy intake as such.

Gradeable strong proposition (revised in this edition to match the currently defended model)

In common human obesity, a diet high in glycemic load is the primary modifiable upstream cause: its hormonal response, centred on a raised insulin-to-glucagon ratio, shifts fuel partitioning toward adipose storage and lowers fuel availability to other tissues, which increases hunger and lowers energy expenditure, so that fat storage drives positive energy balance rather than following from it; after an adaptation period of weeks, a lower-glycemic-load diet should therefore raise energy expenditure at equal energy and protein intake and reduce ad libitum intake.

This follows the proponents' current statements [19,38]. The exposure is glycemic load: "Among modifiable factors, dietary glycemic load (GL) has central importance" [19]. The endocrine signal is the insulin-to-glucagon ratio together with incretins, not insulin alone. The causal claim is a reversal of direction: "a positive energy balance does not cause increasing adiposity; rather, a shift in substrate partitioning favoring fat storage drives a positive energy balance" [19]. Chronic effects are said to need weeks of adaptation [19,22]. Proponents also write that "many dietary and nondietary factors undoubtedly contribute" [19]. Opponents describe the same formulation as treating "diets high in glycemic load as the primary factor responsible for common obesity" [20]. Grading the primary-cause wording is therefore fair to the model as defended.

The phrase "rather than excess energy intake as such" sets a mechanism against a conservation law. Energy balance is shared by every viable model. The dispute is about which processes upstream of that balance regulate intake, expenditure and storage [19,20]. The original edition phrased SC2 as storage "independently of total energy balance". No proponent defends that, and it is corrected below.

Scopes were fixed before outcomes. "Common human obesity" excludes a post hoc "carbohydrate-sensitive responder" category unless the susceptibility is measured before randomization and reproduced independently. "At equal energy and protein intake" makes the isocaloric prediction load-bearing, and "after an adaptation period of weeks" makes study duration a stated boundary rather than a later excuse.

Prespecified subclaims

The preregistered aggregation rule is weakest-link, because the statement is a serial causal chain ending in primacy. With SC1 = Η and SC2–SC6 = Γ, the composite is Γ. The primary-subclaim rule (primary = SC6) returns Γ. Scope-matched aggregation returns Γ as well, since SC6 is the only subclaim whose scope matches the full claim. No rule returns anything else. The hypothesis index has no entry for this hypothesis, so no index rule is overridden.

Rubric routing. All seven rubrics were checked. The closest fit is 06-metabolism-aging.md, which covers "metabolic master-regulator models". Its thresholds are written for lifespan and healthspan ("Requires at least one intervention changing lifespan or healthspan in a way consistent with the hypothesis"). In this review they are applied with adiposity and weight regulation as the organismal phenotype. That is an adaptation, declared here as SKILL.md Step 5 requires, and flagged as a gap: no rubric in the packet covers body-weight or energy regulation. The rubric's Γ hard rules apply without adaptation: "Biomarker-only improvement without organismal benefit" and "Single-species or single-strain evidence". So do its confounder "Caloric intake confounding" and universal cap 4, "Biomarker change without functional/phenotypic change caps at Γ".

COMPETING MECHANISTIC MODELS

The tournament does not set the HSEF letter or the mechanistic role. It shows how a real endocrine response can sit alongside a low grade for the stronger causal chain. Models 1 and 3 are not mutually exclusive: model 3 is effectively a special case of model 1, a reading that one side of the debate has itself put forward [20]. The experiment that would separate an insulin-specific route from correlated properties of high- and low-glycemic-load foods is a long factorial feeding trial. It would vary glycemic load and food architecture independently, clamp protein and offered energy density, and measure insulin, fuel flux, hunger, ad libitum intake, expenditure and fat mass.

FULL REVIEW

1. What the model gets right

Digestible carbohydrate raises glucose and insulin more than dietary fat, and it raises the insulin-to-glucagon ratio. The controlled human interventions reviewed here all produced the expected endocrine separation. Carbohydrate restriction lowered 24-hour urinary C-peptide by 22.3% ± 7.0% [1]. An isocaloric ketogenic diet lowered it by 47% ± 3%, "rapidly and persistently", while glucagon rose [2]. A high-glycemic-index meal produced higher insulin and lower glucagon than a low-glycemic-index meal [8]. Glycemic-index-graded meals moved glucose and insulin in the predicted pattern [7].

A second, independent line comes from human genetics. Inactivating glucokinase mutations shift the glucose–insulin-secretion dose–response curve to the right, with secretion 61% lower between 5 and 9 mM glucose [42]. Activating KCNJ11 mutations uncouple secretion from glucose metabolism, and closing the channel with a sulfonylurea restores it, "more highly stimulated by oral glucose or a mixed meal than by intravenous glucose" [43]. A food-matrix or assay artifact in feeding studies could not produce the genetic dose–response shift, and a developmental artifact of the mutations could not produce within-person dietary responses. SC1 therefore reaches Η.

It stops short of Θ for a reason that matters to the CIM. In a comparison of 38 foods, protein-rich foods and fat-rich refined bakery products "elicited insulin responses that were disproportionately higher than their glycemic responses", and total carbohydrate content correlated only r = 0.39 with the insulin score (36 foods with composition data) [41]. Carbohydrate is a major dietary insulin secretagogue but not the only one.

Insulin also has causal capacity beyond acting as a marker. In engineered mice fed a high-fat diet, lowering Ins1 gene dosage on an Ins2-null background prevented hyperinsulinemia and "completely protected" against diet-induced obesity, partly by raising energy expenditure in white adipose tissue [15]. In a bidirectional Mendelian-randomization analysis, genetically higher 30-minute insulin was associated with higher BMI, while genetically higher BMI was not associated with insulin-30 [16]. Both findings are bounded. The mouse exposure was a high-fat diet in a modified genotype. The genetic instruments explained about 0.9% of insulin-30 variance in validation, and their heterogeneity was driven by a variant near QPCTL/GIPR; removing it attenuated the estimate [16]. That points to a possible incretin route rather than a dietary-carbohydrate route.

2. Partitioning and net fat gain

Insulin suppresses adipose lipolysis after eating and favours glucose use and storage. The CIM goes further: partitioning driven by glycemic load should be large enough that, at equal intake, a lower-glycemic-load diet produces more fat loss or higher expenditure once adaptation has occurred [19,38].

The 2015 NIH ward study tested the short-term form of this prediction [1]. Nineteen adults with obesity each received two six-day diets, each cutting energy by 30%, in randomized crossover order. Carbohydrate restriction cut insulin secretion but raised net fat oxidation. Fat restriction left fat oxidation unchanged. Fat loss, calculated from calorimetry-measured fat balance (not from a model), was 89 g/day on fat restriction versus 53 g/day on carbohydrate restriction, cumulatively 463 ± 37 g versus 245 ± 21 g. DXA "did not detect a significant difference", and the carbohydrate-restricted diet produced more weight loss [1]. The 2016 ward study used stronger restriction for four weeks in 17 men, after a four-week high-carbohydrate baseline. Chamber expenditure rose 57 ± 13 kcal/day and sleeping expenditure rose in the first week before it "waned". Doubly labelled water gave +151 ± 63 kcal/day. Body-fat loss slowed [2].

The two pooled analyses of isocaloric controlled-feeding studies come from opposite sides of the debate, and on duration they do not conflict. One pooled 32 studies (563 subjects) with all food provided. It found that both expenditure and fat loss were modestly greater with lower-fat diets [21]. The other reanalysed and updated the set to 29 studies (617 participants; median duration 4 days). It found that duration modified the effect. Lower-carbohydrate diets reduced expenditure in the 23 shorter trials and raised it in the 6 trials lasting 2.5 weeks or more [22]. Both analyses are syntheses, cited here for their direction; they are Tier B and carry no experimental number in this review. Together they show two things. The short-term isocaloric data favour lower-fat diets. The longer-term data are too few to settle the adaptation question: six trials, three of which measured expenditure by doubly labelled water, a method whose validity under changing macronutrient composition is itself disputed [10,22,24].

Controlled overfeeding shows that imposed surplus energy is stored whatever its source. When excess carbohydrate and excess fat were each imposed for 14 days, 75–85% of the carbohydrate surplus and 90–95% of the fat surplus was stored [17]. Across 5%, 15% and 25% protein diets at about 40% excess energy for eight weeks, body fat rose similarly [18]. Because these designs impose the surplus, they cannot test the CIM's upstream claim that glycemic load drives the surplus. What they show is that excess dietary fat is stored efficiently without a carbohydrate-insulin trigger.

SC2 remains Γ (Moderate confidence). Mouse genetics crosses Δ within the mouse system [15]. Under the framework's precedence rule, the inference from that system to human diet-driven obesity is a separate step and is not established. The direct human prediction, less fat accumulation at equal intake on a lower-glycemic-load diet, is negative in short studies. No long-duration test of fat balance exists. The rubric's Γ hard rule "Biomarker-only improvement without organismal benefit" describes this pattern: insulin changes substantially and the adiposity outcome does not follow.

3. Hunger and voluntary intake

The positive evidence is more substantial than the original edition reported. - In 12 adolescent boys with obesity, voluntary intake after the high-glycemic-index meal was 5.8 MJ. That was 53% more than after the medium-GI meal and 81% more than after the low-GI meal [8]. Only the high- and medium-GI meals "were designed to have similar macronutrient composition, fiber content, and palatability", so the 53% contrast is the matched comparison. - In 12 overweight or obese men, a blinded crossover used meals "controlled for calories, macronutrients, and palatability". Four hours after the high-GI meal, plasma glucose was lower (4.7 vs 5.3 mmol/L) and hunger was greater (P = .04), with greater nucleus accumbens activity [32]. - In DIOGENES, 773 adults who had lost weight were randomized to ad libitum maintenance diets in a protein × glycemic-index design. By intention-to-treat, regain was 0.95 kg less (95% CI 0.33 to 1.57) on low- than on high-GI diets [33]. - In the 2025 randomized meal trial, at group level, low-GI meals led to lower changes in energy intake at the next meal [7]. That is the direction the CIM predicts.

The negative and mediator evidence: - In that same 2025 trial, subjective hunger did not follow the glucose and insulin pattern. At the individual level, intake changes were unrelated to glucose, insulin, free fatty acids, β-hydroxybutyrate and several hormones. There was "a weak negative association … between energy intake changes and insulin or insulin-glucagon ratio at 300 min, opposite to the model's prediction" [7]. The trial was registered for healthy adults aged 18–40 with BMI 18.5 to under 24 kg/m², each randomized to one of three breakfasts (ClinicalTrials.gov NCT05804942; registry enrollment 135). A published erratum corrects Figure 2F (percentage energy from fat) only. The two sides disagree in print about whether the data support the model [30,31]. - In the 2021 NIH inpatient crossover, 20 adults (one withdrew) ate a minimally processed, plant-based low-fat diet (75.2% carbohydrate; glycemic load 85 g per 1,000 kcal) and an animal-based ketogenic diet (10.0% carbohydrate; 6 g per 1,000 kcal), each ad libitum for two weeks. Intake was 689 ± 73 kcal/day lower on the low-fat diet over two weeks, and 544 ± 68 kcal/day lower in the final week [3]. Twenty-four-hour urinary C-peptide was greater on the low-fat diet in both diet-order groups (P < 0.01) [6]. Self-reported appetite did not differ significantly [27]. The diets also differed in food source, fibre, food mass and energy density, so no single mediator can be isolated. What the trial does show directly is that greater carbohydrate-driven insulin secretion need not produce greater intake.

The 2021 trial had no washout, which created a real and still-live dispute. A reanalysis of the public data (16 of 20 participants) found a large carry-over interaction (P = .0004) and a non-significant net dietary effect (P = .4) [4]. The originating group called the reanalysis flawed [5]. The reanalysis authors replied [26]. A later analysis by the originating group found diet order affected 28-day energy balance, but modelling order did not significantly change the within-subject intake difference (interaction 143 ± 140 kcal/day, P = .32) [6]. The exchange continued in 2025 [28,29]. All of these papers analyse the same NCT03878108 participants. They qualify how the trial should be read, but they add no independent causal line.

SC3 remains Γ (Moderate confidence). Glycemic-index manipulations change late-postprandial glucose, hunger, next-meal intake and weight regain in several defined contexts [7,8,32,33]. SC3, however, claims a mediated chain, insulin-driven fuel reduction leading to hunger and then intake. Its one direct mediator test failed [7], and the largest insulin contrast went the opposite way for intake [3]. Under Step 9, the answer to "is the perturbation's result positive?" is therefore unclear for the mechanism as stated. SC3 is one of the two subclaims closest to Δ. A trial that isolates glycemic load from food form, and tests mediation prospectively, would decide it.

4. Energy expenditure: the longer-duration evidence and the Framingham dispute

The strongest human results for the model concern energy expenditure after weight loss. In a crossover of three 4-week isocaloric diets after 10–15% weight loss, total energy expenditure fell by 423 kcal/day on low-fat, 297 on low-glycemic-index and 97 on very-low-carbohydrate, relative to the pre-weight-loss baseline (P for trend by glycemic load < .001) [23]. In the Framingham State Food Study, 164 adults who had lost 12% (within 2%) of body weight were randomized for 20 weeks to 60%, 40% or 20% carbohydrate, with protein controlled and energy titrated to hold weight [9]. By the published analysis, which used the post-weight-loss baseline, expenditure rose 52 kcal/day (95% CI 23 to 82) per 10-percentage-point fall in carbohydrate. The difference was 209 kcal/day (91 to 326) for low- versus high-carbohydrate by intention-to-treat, and 308 kcal/day in the highest third of pre-weight-loss insulin secretion [9]. In a secondary analysis, the energy intake needed to hold weight in weeks 10–20 was higher on the low-carbohydrate diet, by 181 (95% CI 8 to 353) to 246 (64 to 427) kcal/day across covariate models [24].

This is a randomized, provided-food trial with a meaningful diet contrast, and it deserves weight. It is also analysis-sensitive. Using the registered pre-weight-loss baseline, a reanalysis found no diet effect (P = 0.43). It estimated mean unaccounted energy at 483 ± 46 kcal/day, and found that the revised-analysis effect "depended on including subjects with excessive amounts of unaccounted energy" [10]. The investigators replied on three points [25]. They argued that the post-weight-loss baseline is statistically preferable. They reported that an ANCOVA using neither baseline gave a similar effect. And they argued that excluding participants for unaccounted energy introduces post-randomization bias. The energy-requirement analysis [24] partly answers the concern about doubly labelled water. All of these reports describe NCT02068885, so they count as one contested study.

The independent-laboratory data point in both directions. In the 2016 ward study, the chamber-measured rise was small and waned, the doubly-labelled-water rise was +151 kcal/day, and fat loss slowed [2]. The short-trial pool favours lower-fat diets [21], and the longer-trial pool, only six trials, favours lower-carbohydrate diets [22]. No study shows the fat-loss consequence that a sustained expenditure gain of this size should produce.

SC4 remains Γ (Moderate confidence). Under Step 9, a perturbation crosses Δ only when its result is positive. Here the only randomized parallel test is positive by the investigators' revised analysis plan and null by the registered one. The expenditure signal also lacks a measured body-composition correlate. So the answer is contested, and the ceiling stays at Γ. SC4 is the other subclaim closest to Δ. A residential isocaloric trial has been completed that compares very-low-carbohydrate, high-starch and high-sugar diets for 10 weeks after weight loss, with expenditure and body-composition outcomes. Seventy-seven participants were randomized and 70 completed, which is 65% of the planned number [37]. Its primary results were not found in PubMed as of 2026-09-30, and they bear directly on SC4.

5. Long-term diet trials and susceptibility

In DIETFITS, 609 adults were randomized to healthy low-fat or healthy low-carbohydrate advice for 12 months without calorie targets [11]. Weight change was −5.3 versus −6.0 kg, a difference of 0.7 kg (95% CI −0.2 to 1.6). At 12 months, carbohydrate provided 48% versus 30% of energy. Baseline 30-minute insulin (INS-30) did not modify the diet effect (P = .47), and neither did the prespecified genotype pattern (P = .20).

POUNDS LOST assigned 811 adults to four fat–protein–carbohydrate combinations for two years [12]. Weight loss was similar across diets (P > 0.20), and "Satiety, hunger, satisfaction with the diet … were similar". Foster et al. compared low-carbohydrate and low-fat programmes in 307 adults over two years [13]. The low-carbohydrate diet reintroduced carbohydrate by protocol after three months. Weight, body composition and bone density did not differ, and "attrition at 2 years was high".

A meta-analysis of 53 trials lasting at least one year found a small, consistent advantage for low-carbohydrate over low-fat weight-loss interventions (18 comparisons; I² = 10%) [36]. That advantage is compatible with several mechanisms, including protein content, ketosis-related appetite effects and adherence, and it is far smaller than a primary-cause model implies. The accurate summary is therefore that no large macronutrient-specific advantage survives ordinary long-term adherence conditions. Two limits apply: in free-living trials the diets converge, and achieved separation is modest.

On susceptibility, the evidence is split along lines of design. - In a trial of 73 obese young adults over 18 months, INS-30 modified the response to a low-glycemic-load versus low-fat diet (P = .02). Above the median, weight change was −5.8 versus −1.2 kg [34]. - In the Framingham trial, the expenditure contrast was larger in the high-insulin-secretion third [9]. - An exploratory reanalysis of DIETFITS modelled achieved glycemic-load reduction, which was not the randomized exposure. It found effect modification by baseline insulin secretion at 3 and 6 months, weakening by 12 months [35]. - The one large, prespecified, independent test was null (DIETFITS, P = .47) [11].

Because the positive signals are small, come from one research group, or are post hoc, and the prespecified replication failed, SC5 remains Γ (Moderate confidence). A true responder stratum remains possible. To establish it, the stratum must be defined before outcomes are known and then reproduced in a new trial.

6. Alternative routes to higher intake and weight

In an inpatient crossover, ultra-processed and unprocessed diets were matched for presented energy, energy density, macronutrients, sugar, sodium and fibre. Participants ate 508 ± 106 kcal/day more on the ultra-processed diet, gaining 0.9 kg, and lost 0.9 kg on the unprocessed diet [14]. An independent 8-week free-living crossover in England found greater weight loss on minimally processed than on ultra-processed diets that both followed national guidelines (difference −1.01%, 95% CI −1.87 to −0.14) [44]. Together these establish a food-architecture route to intake and weight that does not depend on macronutrient proportions.

They do not exclude a glycemic-load route. Glycemic index and glycemic load were not matched or reported in the inpatient trial. Urinary C-peptide was somewhat higher on the ultra-processed diet during chamber days (38.9 ± 2.8 vs 30.9 ± 2.8 nmol/day; p = 0.052), and the extra intake came from both carbohydrate (280 kcal/day) and fat (230 kcal/day) [14]. The current CIM names processed, high-glycemic-load carbohydrate as its exposure [19]. The ultra-processed trials therefore strengthen the food-architecture model (Model 2). They do not, on their own, falsify carbohydrate–insulin primacy.

None of this establishes that all calories or foods act alike. Protein changes expenditure and lean-mass gain [18]. Carbohydrate and fat differ in oxidation and storage kinetics [17]. Food form affects eating behaviour [14,44]. Glycemic index affects late-postprandial hunger in some settings [32].

7. Primacy adjudication

The HSEF direction-of-causation procedure reserves primary driver for a mechanism whose engagement initiates the phenotype, whose removal prevents or materially delays it, and which is rate-limiting. At the scope of common human obesity, the carbohydrate–insulin pathway meets none of the three: - Removal. Lowering glycemic load or insulin exposure has not produced greater fat loss at equal intake [1,2,21], and in long-term trials it produces at most a small advantage [11–13,36]. - Engagement. A higher-glycemic-load, higher-insulin diet produced lower ad libitum intake than a ketogenic comparator in the only inpatient test [3], and the direct mediator test failed [7]. - Rate limitation. Excess dietary fat is stored efficiently without a carbohydrate trigger [17], and food architecture moves intake at matched macronutrient proportions [14,44]. The positive findings (mouse genetics, MR, glycemic index effects on hunger and regain, and the longer-duration expenditure signals) show that the pathway can contribute in defined contexts. They do not show that it initiates common obesity or controls its rate.

Role-of-mechanism derivation (framework five-step procedure; stop at first match): 1. Downstream consequence? No. The dietary glycemic-load→insulin response exists independently of adiposity, and bidirectional MR did not support BMI→insulin-30 [16]. 2. Epiphenomenon? No. Causal dependency is shown for the endocrine link (SC1) and, in mice, for adiposity [15]. 3. Primary driver? No. Removal, initiation and rate limitation are not shown in humans (above). 4. Permissive factor? No at the claimed scope. Insulin signalling per se is required for adipose storage, but carbohydrate-driven insulin has not been shown to be required for common obesity, and fat surplus is stored efficiently [17]. 5. Otherwise it amplifies or maintains → reinforcing / stabilizing factor. This matches glycemic index effects on late-postprandial hunger [32] and on weight regain after weight loss [33], and hyperinsulinemia-driven adiposity in mice [15].

The classification concerns the strong common-obesity claim. It does not deny insulin's physiological role, or that glycemic load can matter for some people and outcomes.

8. Boundary checks, caps and falsification

Boundary check (Step 9). The field-valid perturbation is taken from rubric 06: "Genetic or pharmacologic intervention producing reproducible lifespan/healthspan extension". It is applied here with adiposity and weight regulation in place of lifespan and healthspan (see the rubric-routing note above). Clause 5 of Step 9 is quoted verbatim: "Is that perturbation's result positive? A study of exactly the right design that returns a null or contradicts the hypothesis does not cross Δ." - SC1 passes: the perturbation is positive and reproduced, and the thresholds are tested upward to Η (see §1). - SC2 fails at question 5 at human scope. The human isocaloric fat-balance result is negative [1,2,21], and the positive mouse result binds only within its own system [15]. - SC3 answers unclear at question 5 for the mechanism as stated. Glycemic index effects on hunger and intake are positive in defined contexts [7,8,32,33], but mediation failed [7] and the largest insulin contrast went the opposite way [3]. - SC4 answers contested at question 5: positive or null depending on analysis [9,10,24,25], with no fat-loss correlate [1,2,21]. - SC5 fails at question 2, because the proposed boundary has not been independently reproduced [11,34,35]. - SC6 fails at questions 2 and 5.

Caps (verbatim). - Universal cap 4, "Biomarker change without functional/phenotypic change caps at Γ", binds SC2 and SC4. Insulin secretion changes substantially without the predicted adiposity outcome. - Rubric 06 Γ hard rules bind two subclaims. "Biomarker-only improvement without organismal benefit" binds SC2 and SC4. "Single-species or single-strain evidence" binds SC2 and SC6, whose only in vivo causal demonstration of insulin-driven adiposity is in one engineered mouse model [15]. - Cap precedence. The mouse gene-dosage result crosses Δ within the mouse system. The cap governs the inference from mouse to human diet-driven common obesity. This is recorded, not resolved silently.

Falsification patterns (rubric 06, adapted). - "Pathway manipulation changes biomarkers but not lifespan/healthspan" is triggered for SC2 and SC4. Large reductions in insulin secretion did not produce the predicted fat-balance change [1,2,21]. - "Necessity fails when the proposed pathway is removed" is triggered for SC6: lowering glycemic load does not prevent or reverse obesity beyond a small margin [11–13,36]. - For SC3, the mediator prediction failed [7].

These patterns are the same observations that made the Step 9 answers negative or unclear, and they already set the Γ ceilings. No further tier deduction is applied, because that would count one observation twice. Applied literally, one extra tier per pattern would put SC2–SC6, and the composite, at Β. Β's descriptor ("reproducible correlation … without demonstrated causal dependency") does not fit a body of evidence that includes a causal endocrine link at Η and positive human perturbations in defined contexts.

Upper-ladder equivalents were considered. - Mendelian randomization is available [16]. It instruments stimulated insulin secretion rather than diet, explains little variance, and its heterogeneity traces to a GIPR-region variant. - Multiple human dietary perturbations exist, but for the full chain they are negative, mixed or contested [1–3,7,9–11,21–24]. - Withdrawal and rechallenge analogues (crossovers) are short and have no washout [3–6]. - Dose–response across populations is confounded by food form, energy density and protein. - No therapeutic perturbation of the pathway has a hard long-term obesity endpoint with mediation. There is no positive convergent set being re-capped.

9. Independence and correction audit

Reports were grouped by study, not counted as separate papers: - NCT03878108 comprises the 2021 trial, the Soto-Mota reanalysis and reply, the two Sciarrillo analyses, the Hengist appetite-hormone analysis and the 2025 correspondence [3–6,26–29]. - NCT02068885 comprises the Framingham trial, its reanalysis, the investigators' response and the energy-requirement analysis [9,10,24,25]. - NCT01826591 comprises DIETFITS and its exploratory reanalysis [11,35]. - NCT05804942 comprises the 2025 meal trial and its correspondence [7,30,31]. These debates improve interpretation but add no participants or independent causal lines.

Corrections are attached to their parent records; none has its own row. - Soto-Mota 2024 [4]: corrigendum; Crossref lists it as updating three related J Nutr items. - Liu 2025 [7]: erratum correcting Figure 2F. - Ebbeling 2018 [9]: correction of a table heading. - DIETFITS [11]: two corrections, units of measure and the funding section. - Ultra-processed trial [14]: two errata. - Bray 2012 [18]: "Percentage Error". None affects a number carried here.

For SC1, independence is recorded in the scorecard (independence_basis). The dietary-perturbation line and the monogenic β-cell glucose-sensing line fail in different ways. A food-matrix, protein co-ingestion or immunoassay error could not produce a genetic dose–response shift. A developmental or compensatory effect of a lifelong mutation could not produce within-person responses to dietary switching.

Access. Full text was read for refs 1, 2, 4–6, 9–14, 16, 18–21, 25 and 27. Ref 3 (Nature Medicine) and ref 7 (Cell Metabolism) could not be retrieved in full in this pass. Only their abstract numbers, plus open full-text corroboration from the same study (refs 6 and 27) and the trial registry, are carried. The remaining records were read at abstract depth, and no number beyond the abstract is used. Correspondence records (refs 26, 28–31) carry only their existence and stated position.

10. Steelman and final adjudication

The case for Δ, one rung higher. - The endocrine link is at Η [1,2,42,43]. - Glycemic index manipulation changes late-postprandial glucose and hunger in a macronutrient- and palatability-matched blinded crossover [32], raises next-meal intake [8], and at group level in the 2025 trial reduced intake on low-GI meals [7]. In a large independent trial it reduced weight regain [33]. - In the six longer controlled-feeding studies, lower-carbohydrate diets raised expenditure [22]. That is what the Framingham trial's published analysis and energy-requirement analysis found [9,24], and a graded crossover points the same way [23]. - Insulin-secretion effect modification appeared in two trials and one exploratory reanalysis [9,34,35]. - Mouse gene dosage and human MR support insulin-driven adiposity [15,16]. This is a serious case. It establishes that the pathway can contribute in defined contexts, and it is the reason Model 3 ranks above the nominal.

Why it does not cross Δ for the composite. Under the preregistered weakest-link rule, every link must cross. SC6 has no positive necessity, initiation or rate-limitation evidence. SC5's one prespecified independent replication was null [11]. SC3's mediation test failed [7], and the largest insulin contrast produced lower intake [3]. SC4's key trial is null on its registered analysis [10], its signal has no fat-loss correlate [1,2,21], and the positive longer-duration pool is small (six trials, three measured by doubly labelled water) [22]. Even if SC3 and SC4 crossed, the composite would remain Γ. The blocker is the missing positive human test of the full chain: a long-duration trial in which lower glycemic load, at controlled energy, protein and food architecture, raises expenditure and reduces fat accumulation or ad libitum intake, with insulin shown to mediate.

The case for Β, one rung lower. The direct human tests of the chain are negative or contested. Five falsification patterns are triggered. Much of the positive evidence is small, comes from one group, or is post hoc.

Why Β is rejected. Β describes associational evidence without demonstrated causal dependency. The endocrine link is causal at Η. Insulin-driven adiposity is causal in mice. Several glycemic index perturbations in humans are positive in defined contexts. The falsification evidence already sets the Γ ceiling, and deducting again would double-count it.

The final assessment is:

Γ — Mechanistically plausible, partial system support · High confidence

Why High. Confidence reflects how dense and reproducible the evidence is, not a belief that the model is false. The composite is bound by SC6. Its placement rests on: - two opposing syntheses of 29–32 controlled-feeding studies [21,22]; - several large independent long-term trials [11–13,33] and a 53-trial meta-analysis [36]; - inpatient feeding trials from more than one laboratory [1–3,14], plus an independent free-living replication [44]; - the proponents' own position that the comparative validity of the models still requires "rigorous research" [19]. SC2–SC5 are individually at Moderate confidence. High does not mean carbohydrate quality never matters, that lower-glycemic-load diets are not useful for some people, or that everyone responds alike. It means current evidence does not support the submitted primary-cause chain across common human obesity.

11. Momentum, stability and divergence

Momentum: Mechanistic only. Recent work is active but acute, reanalytic or correspondence-based: the 2025 glycemic-index meal trial [7], the reanalyses and replies of 2023–2025 [4–6,26–31,35], and appetite-hormone analyses [27]. The one purpose-built longer residential test has published only its design [37]. It addresses SC4 alone. No trial addresses the composite's binding subclaim, SC6.

Stability: Stable. The named trigger is the primary publication of the residential isocaloric trial [37]. It bears on SC4, in either direction, and at 65% of planned completion it has limited power. Under weakest-link it cannot move the composite while SC3, SC5 and SC6 remain at Γ.

Divergence: Contested. The field is genuinely split. Proponents present the CIM as better reflecting "knowledge on the biology of body weight control" [19]. Opponents present the energy balance model as "a more robust theory" [20]. A neutral perspective concludes that "neither has yet provided a validated mechanistic account" [39], and a joint workshop including authors from both camps calls for more testing [40]. For transparency: the submitted wording "primary cause", read literally, implies Η, which is +4 rungs above the assessed Γ (severe overstatement). The proponents' own wording ("central importance"; "rigorous research is needed") implies a lower claim than the submitted wording. M, S and D never modify the level or confidence.

PROPOSED DECISIVE EXPERIMENTS

1. Factorial residential feeding trial

Randomize adults with overweight or obesity to a factorial design. One factor is high versus low glycemic load. The other is minimally processed versus ultra-processed food architecture. Run it for at least 8–12 weeks, so that adaptation is past, then add a prespecified ad libitum phase. Clamp protein and offered energy density, and harmonize palatability targets, fibre type, sodium, meal schedule and food mass.

Measure: - insulin, C-peptide and glucagon; - glucose, free fatty acids and ketones; - substrate oxidation; - chamber and doubly-labelled-water expenditure, with food quotient recorded; - hunger and daily intake; - body fat and metabolic flux.

If insulin separates widely but appetite, intake, expenditure and fat mass do not follow the predicted order, the generalized chain is falsified. If a glycemic-load effect appears reproducibly in both processing conditions and is quantitatively mediated by insulin, the chain is supported.

2. Insulin mediation and rescue

Within a controlled diet platform, dissociate carbohydrate exposure from insulin using two mechanistically distinct and clinically acceptable perturbations. Then add a pathway-specific rescue that restores insulin signalling without restoring the carbohydrate exposure. Prespecify the temporal chain from insulin to circulating fuels to hunger and intake.

If hunger and intake track carbohydrate or food properties when insulin is dissociated, or persist despite selective insulin reduction, the proposed mediator fails. If selective reduction abolishes the phenotype and rescue restores it, the mechanism is materially narrowed.

3. Prospective responder replication

Fix one baseline insulin-secretion definition, such as INS-30 with a registered cut point, before enrollment. Use it in at least two independent randomized trials with harmonized low- and high-glycemic-load diets. Maintain meaningful diet separation and match protein and diet quality.

No interaction, inconsistent directions, or dependence on a post-outcome cut point would falsify the responder claim. A replicated, clinically meaningful interaction in the same prospectively defined phenotype would raise the context-specific model (Model 3) without establishing universal primacy.

LIMITATIONS OF THIS REVIEW

• No rubric in the HSEF packet covers body-weight or energy regulation. Rubric 06's thresholds were adapted from lifespan and healthspan to adiposity, and this is declared, not hidden.
• Two central records (refs 3 and 7) could not be read in full in this pass. Three original records (refs 8, 15, 17) and sixteen added records (refs 22–24, 32–44) were read at abstract depth, and no number beyond the abstract is carried from them. Correspondence records (refs 26, 28–31) carry only their stated positions.
• Diet trials differ in carbohydrate dose, glycemic load, ketogenic status, energy prescription, protein, food form, adherence, duration and participant phenotype. The review treats these as mapped boundaries and does not assume all "low-carbohydrate" interventions are interchangeable.
• Doubly labelled water under changing macronutrient composition, and provided-food studies with imperfect adherence, both create measurement uncertainty. The Framingham dispute is therefore preserved rather than resolved by counting votes.
• Short ward studies give strong internal control but generalize poorly to the long term. Long free-living trials give duration at the cost of diet separation.
• The HSEF validator checks structure, provenance links, arithmetic and required fields. It does not reproduce the studies or endorse the judgement.

CALIBRATION STATUS

Calibration status: not a calibration entry. This is a standalone evaluation. No independently authored expected band was sealed before grading, so it is not a blinded calibration run.

VALIDATION NOTE

The scorecard was checked with HSEF validator 2.2 during the 2026-09-30 red-team pass, and re-checked with validator 2.4 (the archived HSEF v5.4 validator) for the consensus-consistency supplement and again on 2026-10-14 after adversarial review; the current output is validation-report.json in this folder. A clean pass means the schema, canonical descriptors, aggregation arithmetic, caps, independence basis, claim-to-source links, source counts and required prose blocks are internally consistent. It is not independent scientific validation.

RED-TEAM SYNTHESIS NOTE

This edition is the cross-model red-team synthesis of the original HSEF v5.4 review (authoring model: OpenAI Codex). The red team was Claude (Anthropic). The overall grade is unchanged: Γ, High confidence; role reinforcing / stabilizing factor.

The main changes: 1. SC1 raised from Ε to Η. Thresholds were tested upward, and two independent causal lines were recorded. Θ is blocked by the modest predictive power of carbohydrate content for insulin demand. 2. The claim was restated to match the proponents' current model: glycemic load, the insulin-to-glucagon ratio, fat storage driving positive balance, and a weeks-long adaptation. The strawman "independently of energy balance" was removed from SC2. 3. Balance restored. Positive evidence the original omitted is now included: both controlled-feeding meta-analyses, the 2012 expenditure crossover, the energy-requirement analysis and the investigators' response, the 2013 hunger crossover, DIOGENES, the 2007 effect-modification trial, the DIETFITS reanalysis and the long-term meta-analytic advantage. The 2025 meal trial is described correctly as mixed. 4. The ultra-processed-food trial is no longer treated as a falsifier of primacy, because glycemic load was not matched. 5. Scoring details corrected. Every rubric quotation is now verbatim. The rubric's poor fit is declared. Confidence for SC2–SC5 is lowered to Moderate. Divergence is set to Contested. The pending residential trial is named.

All 20 original identifiers and their 8 corrections were verified; none was confabulated. Twenty-four references were added, and one unverifiable number was removed. Full audit trail: redteam/REDTEAM_REPORT.md (25 numbered findings, citation table) and redteam/SYNTHESIS_CHANGE_LOG.md (before→after table and itemised changes).

ADVERSARIAL REVIEW AND APPROVAL

Authoring model: Sol 6.1 (original authoring recorded as OpenAI Codex with HSEF v5.4; cross-model red-team synthesis by Claude, 2026-09-30). Adversarial reviewer: Claude Opus 5.5 (Anthropic). Date: 2026-10-14.

Scope of checks. PMIDs verified 44/44 against PubMed get_article_metadata: first author, title, journal, year and DOI matched for every reference; 0 unresolvable, 0 wrong-record substitutions. The seven linked correction notices cited in the reference list were separately confirmed to exist (PMID 38609206, 40306286, 33144344, 29614160, 29710144, 31269427, 33027677). Retraction / expression-of-concern screen on all 44 records: none found. 12 load-bearing claims spot-checked against primary abstracts or full text, including every figure that carries the adjudication: Ebbeling 2020 energy requirement 181 (8–353) to 246 (64–427) kcal/d; Ebbeling 2012 TEE falls of −423 / −297 / −97 kcal/d with P for trend by glycemic load < .001; Ludwig 2021 meta-analysis of 29 trials, 617 participants, 23 shorter trials −50.0 kcal/d and 6 longer trials +135.4 kcal/d; Hall & Guo 32 controlled-feeding studies with both expenditure (26 kcal/d) and fat loss (16 g/d) greater on lower-fat diets; Horton 1995 75–85% versus 90–95% of surplus stored; Bray 2012 body fat rising similarly across 5/15/25% protein; Hengist 2024 greater intake on the low-carbohydrate arm; DIETFITS null INS-30 interaction; the Wong 2024 residential-trial figures — 77 randomized, 70 completers, "65% of planned completion" — which appear verbatim in that abstract. A fresh PubMed search on 2026-10-14 confirms the residential trial's primary results are still unpublished, so the review's statement about them is correct both at its stated cutoff and today. Scoring logic re-derived against hsef-hypothesis-review-v5/SKILL.md, references/hsef-framework.md and references/evidence-rubrics/06-metabolism-aging.md.

Scoring logic — checked and upheld. Hard caps are applied before thresholds, quoted verbatim, and correctly attributed: universal cap 4 and rubric 06's "Biomarker-only improvement without organismal benefit" bind SC2 and SC4, and "Single-species or single-strain evidence" binds SC2 and SC6. The cap-precedence rule is applied exactly as the framework writes it — the mouse gene-dosage result crosses Δ within the mouse system, and the cap governs the inference from that system to human diet-driven obesity — and the resolution is recorded rather than made silently. The declared weakest-link rule is applied as written (SC1 Η, SC2–SC6 Γ → Γ) and both alternatives are reported and correct: primary-subclaim (SC6) Γ, and scope-matched Γ, since SC6 is the only subclaim whose scope matches the full claim. The review also declines to deduct a further tier for each triggered falsification pattern, correctly, because those are the same observations that already set the Γ ceilings — double-counting them would have driven the composite to Β against a body of evidence containing an Η-level causal endocrine link. Role of mechanism is derived by walking the five-step procedure in order and stopping at step 5; the tournament is explicitly not used as the input, which matters here because the nominal model ranks last. Momentum "Mechanistic only" is correct under lead policy and its artifact names specific records. Stability "Stable" names a real trigger with its direction left open. Divergence "Contested" is supported with the implied-level arithmetic stated (+4 rungs for the submitted wording). The rubric's poor fit is declared rather than hidden, as SKILL.md Step 5 requires.

Rejected findings (checked, no change). 1. Confidence High beside a Γ letter. Checked against hsef-framework.md § Confidence band and the lead's two-step test, and upheld. The rationale contains no reference to reviewer access depth; it rests on evidence density and reproducibility — two opposing syntheses of 29 and 32 controlled-feeding studies, four large independent long-term randomized trials, a 53-trial meta-analysis, inpatient trials from more than one laboratory and an independent free-living replication. The framework explicitly contemplates this pair: "A Γ-level hypothesis can have high confidence (we are confident the evidence does not yet justify more than Γ)." Nothing to re-derive. 2. One reference whose print year differs by one from the PubMed e-pub year (Soto-Mota, J Nutr 2024 vs 2023). Standing convention. 3. "Best-supported rival" in AT A GLANCE naming the integrated energy-homeostasis model. Score card confirms rank_in_tournament: 1 with is_nominal: false; the review's own nominal (M1) ranks 4. Defect absent. 4. Collective author. Reference 33 carries the DIOGENES consortium and renders correctly as "Larsen TM, Dalskov SM, van Baak M, et al.; Diogenes Project" — the consortium is present, with no blank author or stray leading comma. This is the one review in this group with a collective author, and the builder handled it. 5. Rendering defects. No glued numerals, no empty DOI: field, no "PMIDnot supplied" string. Checked explicitly. 6. REVIEWER NOTES vs the spliced Consensus section. This review has no statement that Consensus work was not performed; § Consensus provenance reconciliation and the addendum are consistent with each other. 7. Momentum. Already "Mechanistic only", with a trial that has published only its design treated as not satisfying "Decisive work underway". Correct under lead policy.

Residual limitations. Unchanged from the review's own statement: refs 3 and 7 could not be read in full, and refs 8, 15, 17 and 22–44 were read at abstract depth, so figures attributed to them rest on their abstracts. No rubric in the packet covers body-weight or energy regulation, and rubric 06's thresholds remain an adaptation. This pass did not re-derive any number from raw trial data.

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

REFERENCES

1. Hall KD, Bemis T, Brychta R, et al. Calorie for Calorie, Dietary Fat Restriction Results in More Body Fat Loss than Carbohydrate Restriction in People with Obesity. Cell Metabolism. 2015;22(3):427-436. PMID 26278052 · PMC full text · DOI
2. Hall KD, Chen KY, Guo J, et al. Energy expenditure and body composition changes after an isocaloric ketogenic diet in overweight and obese men. American Journal of Clinical Nutrition. 2016;104(2):324-333. PMID 27385608 · PMC full text · DOI
3. Hall KD, Guo J, Courville AB, et al. Effect of a plant-based, low-fat diet versus an animal-based, ketogenic diet on ad libitum energy intake. Nature Medicine. 2021;27(2):344-353. PMID 33479499 · DOI
4. Soto-Mota A, Jansen LT, Norwitz NG, Pereira MA, Ebbeling CB, Ludwig DS. Physiologic Adaptation to Macronutrient Change Distorts Findings from Short Dietary Trials: Reanalysis of a Metabolic Ward Study. Journal of Nutrition. 2024;154(4):1080-1086. PMID 38128881 · PMC full text · DOI. Corrigendum: J Nutr 2024;154(4):1486, PMID 38609206 · DOI (Crossref lists it as updating this article and refs 5 and 26).
5. Sciarrillo CM, Guo J, Hengist A, Darcey VL, Hall KD. Flawed reanalysis fails to support the carbohydrate-insulin model of obesity. Journal of Nutrition. 2024;154(3):1058-1060. PMID 38286396 · PMC full text · DOI
6. Sciarrillo CM, Guo J, Hengist A, Darcey VL, Hall KD. Diet order significantly affects energy balance for diets varying in macronutrients but not ultraprocessing in crossover studies without a washout period. American Journal of Clinical Nutrition. 2024;120(4):953-963. PMID 39163976 · PMC full text · DOI
7. Liu Y, Mei H, Xue L, et al. Testing the carbohydrate-insulin model: Short-term metabolic responses to consumption of meals with varying glycemic index in healthy adults. Cell Metabolism. 2025;37(3):606-615.e3. PMID 40043690 · DOI. Erratum (Figure 2F): Cell Metab 2025;37(6):1453, PMID 40306286 · DOI. Registry: ClinicalTrials.gov NCT05804942.
8. Ludwig DS, Majzoub JA, Al-Zahrani A, Dallal GE, Blanco I, Roberts SB. High glycemic index foods, overeating, and obesity. Pediatrics. 1999;103(3):E26. PMID 10049982 · DOI
9. Ebbeling CB, Feldman HA, Klein GL, et al. Effects of a low carbohydrate diet on energy expenditure during weight loss maintenance: randomized trial. BMJ. 2018;363:k4583. PMID 30429127 · PMC full text · DOI. Correction (table heading): BMJ 2020;371:m4264, PMID 33144344 · DOI.
10. Hall KD, Guo J, Speakman JR. Do low-carbohydrate diets increase energy expenditure? International Journal of Obesity. 2019;43(12):2350-2354. PMID 31548574 · PMC full text · DOI
11. Gardner CD, Trepanowski JF, Del Gobbo LC, et al. Effect of Low-Fat vs Low-Carbohydrate Diet on 12-Month Weight Loss in Overweight Adults and the Association With Genotype Pattern or Insulin Secretion: The DIETFITS Randomized Clinical Trial. JAMA. 2018;319(7):667-679. PMID 29466592 · PMC full text · DOI. Corrections: PMID 29614160 (units of measure) and PMID 29710144 (funding/support section).
12. Sacks FM, Bray GA, Carey VJ, et al. Comparison of weight-loss diets with different compositions of fat, protein, and carbohydrates. New England Journal of Medicine. 2009;360(9):859-873. PMID 19246357 · PMC full text · DOI
13. Foster GD, Wyatt HR, Hill JO, et al. Weight and metabolic outcomes after 2 years on a low-carbohydrate versus low-fat diet: a randomized trial. Annals of Internal Medicine. 2010;153(3):147-157. PMID 20679559 · PMC full text · DOI
14. Hall KD, Ayuketah A, Brychta R, et al. Ultra-Processed Diets Cause Excess Calorie Intake and Weight Gain: An Inpatient Randomized Controlled Trial of Ad Libitum Food Intake. Cell Metabolism. 2019;30(1):67-77.e3. PMID 31105044 · PMC full text · DOI. Errata: PMID 31269427 and PMID 33027677.
15. Mehran AE, Templeman NM, Brigidi GS, et al. Hyperinsulinemia drives diet-induced obesity independently of brain insulin production. Cell Metabolism. 2012;16(6):723-737. PMID 23217255 · DOI
16. Astley CM, Todd JN, Salem RM, et al. Genetic Evidence That Carbohydrate-Stimulated Insulin Secretion Leads to Obesity. Clinical Chemistry. 2018;64(1):192-200. PMID 29295838 · PMC full text · DOI
17. Horton TJ, Drougas H, Brachey A, Reed GW, Peters JC, Hill JO. Fat and carbohydrate overfeeding in humans: different effects on energy storage. American Journal of Clinical Nutrition. 1995;62(1):19-29. PMID 7598063 · DOI
18. Bray GA, Smith SR, de Jonge L, et al. Effect of dietary protein content on weight gain, energy expenditure, and body composition during overeating: a randomized controlled trial. JAMA. 2012;307(1):47-55. PMID 22215165 · PMC full text · DOI. Correction ("Percentage Error"): JAMA 2012;307(10):1028, DOI.
19. Ludwig DS, Aronne LJ, Astrup A, et al. The carbohydrate-insulin model: a physiological perspective on the obesity pandemic. American Journal of Clinical Nutrition. 2021;114(6):1873-1885. PMID 34515299 · PMC full text · DOI
20. Hall KD, Farooqi IS, Friedman JM, et al. The energy balance model of obesity: beyond calories in, calories out. American Journal of Clinical Nutrition. 2022;115(5):1243-1254. PMID 35134825 · PMC full text · DOI
21. Hall KD, Guo J. Obesity Energetics: Body Weight Regulation and the Effects of Diet Composition. Gastroenterology. 2017;152(7):1718-1727.e3. PMID 28193517 · PMC full text · DOI
22. Ludwig DS, Dickinson SL, Henschel B, Ebbeling CB, Allison DB. Do Lower-Carbohydrate Diets Increase Total Energy Expenditure? An Updated and Reanalyzed Meta-Analysis of 29 Controlled-Feeding Studies. Journal of Nutrition. 2021;151(3):482-490. PMID 33274750 · PMC · DOI
23. Ebbeling CB, Swain JF, Feldman HA, et al. Effects of dietary composition on energy expenditure during weight-loss maintenance. JAMA. 2012;307(24):2627-2634. PMID 22735432 · PMC · DOI
24. Ebbeling CB, Bielak L, Lakin PR, et al. Energy Requirement Is Higher During Weight-Loss Maintenance in Adults Consuming a Low- Compared with High-Carbohydrate Diet. Journal of Nutrition. 2020;150(8):2009-2015. PMID 32470981 · PMC · DOI
25. Ludwig DS, Lakin PR, Wong WW, Ebbeling CB. Scientific discourse in the era of open science: a response to Hall et al. regarding the Carbohydrate-Insulin Model. International Journal of Obesity. 2019;43(12):2355-2360. PMID 31586125 · PMC full text · DOI
26. Soto-Mota A, Jansen LT, Norwitz NG, Pereira MA, Ebbeling CB, Ludwig DS. Reply to C M Sciarrillo et al. Journal of Nutrition. 2024;154(3):1061-1063. PMID 38316213 · DOI
27. Hengist A, Sciarrillo CM, Guo J, Walter M, Hall KD. Gut-derived appetite hormones do not explain energy intake differences in humans following low-carbohydrate versus low-fat diets. Obesity (Silver Spring). 2024;32(9):1689-1698. PMID 39113385 · PMC full text · DOI
28. Ludwig DS, Willett WC, Putt ME. Concern for the validity of short-term dietary crossover trials. Nature Medicine. 2025;31(9):2913-2914. PMID 40846772 · DOI
29. Hall KD. Reply to: Concern for the validity of short-term dietary crossover trials. Nature Medicine. 2025;31(9):2915-2916. PMID 40846773 · DOI
30. Ludwig DS, Brand-Miller JC, Ebbeling CB, Friedman MI, Norwitz NG, Soto-Mota A. Testing the carbohydrate-insulin model: The data are supportive! Cell Metabolism. 2025;37(6):1239-1240. PMID 40466618 · DOI
31. Speakman JR, Liu Y. Testing the carbohydrate-insulin model: Some aspects are consistent, but overall the data do not support the model. Cell Metabolism. 2025;37(6):1241-1242. PMID 40466619 · DOI
32. Lennerz BS, Alsop DC, Holsen LM, et al. Effects of dietary glycemic index on brain regions related to reward and craving in men. American Journal of Clinical Nutrition. 2013;98(3):641-647. PMID 23803881 · PMC · DOI
33. Larsen TM, Dalskov SM, van Baak M, et al.; Diogenes Project. Diets with high or low protein content and glycemic index for weight-loss maintenance. New England Journal of Medicine. 2010;363(22):2102-2113. PMID 21105792 · PMC · DOI
34. Ebbeling CB, Leidig MM, Feldman HA, Lovesky MM, Ludwig DS. Effects of a low-glycemic load vs low-fat diet in obese young adults: a randomized trial. JAMA. 2007;297(19):2092-2102. PMID 17507345 · DOI
35. Soto-Mota A, Pereira MA, Ebbeling CB, Aronica L, Ludwig DS. Evidence for the carbohydrate-insulin model in a reanalysis of the Diet Intervention Examining The Factors Interacting with Treatment Success (DIETFITS) trial. American Journal of Clinical Nutrition. 2023;117(3):599-606. PMID 36811468 · DOI
36. Tobias DK, Chen M, Manson JE, Ludwig DS, Willett W, Hu FB. Effect of low-fat diet interventions versus other diet interventions on long-term weight change in adults: a systematic review and meta-analysis. Lancet Diabetes & Endocrinology. 2015;3(12):968-979. PMID 26527511 · PMC · DOI
37. Wong JMW, Ludwig DS, Allison DB, et al. Design and conduct of a randomized controlled feeding trial in a residential setting with mitigation for COVID-19. Contemporary Clinical Trials. 2024;140:107490. PMID 38458559 · DOI
38. Ludwig DS. Carbohydrate-insulin model: does the conventional view of obesity reverse cause and effect? Philosophical Transactions of the Royal Society B. 2023;378(1888):20220211. PMID 37661740 · PMC · DOI
39. Flier JS. Moderating "the great debate": The carbohydrate-insulin vs. the energy balance models of obesity. Cell Metabolism. 2023;35(5):737-741. PMID 37086719 · DOI
40. Magkos F, Sørensen TIA, Raubenheimer D, et al. On the pathogenesis of obesity: causal models and missing pieces of the puzzle. Nature Metabolism. 2024;6(10):1856-1865. PMID 39164418 · DOI
41. Holt SH, Miller JC, Petocz P. An insulin index of foods: the insulin demand generated by 1000-kJ portions of common foods. American Journal of Clinical Nutrition. 1997;66(5):1264-1276. PMID 9356547 · DOI
42. Byrne MM, Sturis J, Clément K, et al. Insulin secretory abnormalities in subjects with hyperglycemia due to glucokinase mutations. Journal of Clinical Investigation. 1994;93(3):1120-1130. PMID 8132752 · PMC · DOI
43. Pearson ER, Flechtner I, Njølstad PR, et al. Switching from insulin to oral sulfonylureas in patients with diabetes due to Kir6.2 mutations. New England Journal of Medicine. 2006;355(5):467-477. PMID 16885550 · DOI
44. Dicken SJ, Jassil FC, Brown A, et al. Ultraprocessed or minimally processed diets following healthy dietary guidelines on weight and cardiometabolic health: a randomized, crossover trial. Nature Medicine. 2025;31(10):3297-3308. PMID 40760353 · PMC · DOI

Documentary record

Every grade on this site is reproducible from its record. The files below are the posted edition exactly as assessed.

• ADVERSARIAL_REVIEW.md
• SHA256.txt
• adversarial-findings.json
• carbohydrate-insulin-obesity.review.md
• carbohydrate-insulin-obesity.scorecard.json
• codex-correction-provenance.json
• consensus-addendum.md
• consensus-validation.json
• edition-approval-audit.json
• research-update-2026-10-05.bib
• research-update-2026-10-05.html
• research-update-2026-10-05.json
• research-update-2026-10-05.md
• research-update-2026-10-05.ris
• validation-report.json
• validation-round2-2026-10-14.json
• validation.txt
• video-insert-2026-10-05.md

Sources: 44 · load-bearing claims: 31 · verified: 31 · unverified: 0 · contradicted: 0 · schema 5.4