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.

Curiosity research: status and what survived review

Four Opus deep-research reports, each given a Fable adversarial review (primary-source verification, fallacy hunt, hypothesis ranking with prior art) and then corrected.

Report	Reviewed	Corrected	Trust level
mechanisms-beyond-density	yes	yes	reviewed; verdicts inline
theories-and-anomalies	yes	yes	reviewed; verdicts inline
counterintuitive-levers	yes	yes	reviewed; verdicts inline
frontier-2024-2026	yes	yes	reviewed; verdicts inline

All four were written 2026-09-06 and reviewed 2026-09-06. Every withdrawn claim is left in place inside a bracketed review note so the reasoning trail survives.

What survived review (mechanisms and theories reports)

Held up, and changes how to think:

The 2019 and 2026 ward trials reconcile in principle. Hall 2019 matched total energy density only when beverages were counted; non-beverage density was 1.96 vs 1.06 kcal/g. The 2026 trial manipulated exactly that. The remaining puzzle (2019 stacked all three factors and got +508; the 2026 density-only contrast got +662) awaits publication of diet compositions.
Energy density is the dominant proximal mediator, not proof that processing is a proxy. Density is something processing produces, so conditioning on it and declaring processing null is conditioning on the mediator. The intervention implication is the same either way: lower non-beverage calorie density.
Texture and eating rate are a real candidate lever, but not yet a clean one. Lasschuijt 2023 (n=18, density equalised) moved eating rate 85% and intake 33% with texture alone. The longer Forde 2026 trial (n=41, 369 kcal/day) was not matched for fat (22 vs 33% of energy) or protein (21 vs 16%), so texture and composition are confounded there.
Atwater calorie labels overstate intact, fibrous foods by roughly 4 to 11% and almonds by about 20%. Confirmed by Baer 1997 and Zou 2007. Not novel, but under-known.
Bread and cereal associations are mixed by outcome, not simply protective: inverse for diabetes (EPIC) and cardiovascular disease (US cohorts), slightly positive for mortality. Savoury snacks flip sign across cohorts. Some EPIC inverse hazard ratios (0.46, 0.65) are implausibly large, which argues for artifact in both directions.
A negative-control check exists and was positive: ultra-processed intake predicted accidental death in EPIC. That is a measured confounding floor and should temper every cohort hazard ratio in this field, including the ones for sugary drinks and processed meat.
The additive cohort literature does not obviously reduce to one latent “processed food” factor. The 2026 preservative paper reports persistence after adjusting for percent ultra-processed and low inter-additive correlation. The single-factor hypothesis stays plausible (about 50%), not probable.
The rubric’s headline 4 is correct. Removing sugary drinks and processed meat takes total cardiovascular disease from 1.11 to 1.00 in the US cohorts. Coronary disease alone goes to 1.06, not fully to null.

Falsified or downgraded:

Protein leverage asymmetry (“a floor, flat above 15%”): falsified before it was proposed, by Martens 2013 and Weigle 2005 (leverage above 15%, none below).
Eating rate in grams per minute as a “defended trait invariant to diet”: the registry nulls have confidence intervals of about ±7%, texture was not manipulated in that trial, and the same lab moved eating rate 85% with texture.
“Hyperpalatable foods are not more palatable”: rests on remembered online ratings and one wide-interval null; a 2026 tasted study found large liking differences.
Gut hormones “not where the effect lives”: Hall 2019 measured fasting hormones only, and those moved in the satiety direction on the unprocessed arm. The trial is silent on postprandial signalling.
“Solved” and “direct demonstration” language throughout: replaced with “reconciled in principle” and “not detected”.

What survived review (levers and frontier reports)

Held up:

Glucose-lowering tricks move glucose, not weight, as far as anyone has measured. Food order, vinegar, soup and water preloads all flatten the post-meal curve. No well-controlled, non-industry trial shows a weight effect, but none was powered to detect one kilogram either, so this is “not detected”, not “never”.
Water immediately before a meal cuts that meal by about a fifth in young adults; water 30 minutes before does nothing. Timing, not age, looks like the moderator. No weight endpoint exists.
The 2-hour glucose metric cannot see a late cost. Adding fat flattens the first two hours; in type 1 diabetes on fixed insulin it doubles late hyperglycemia. Whether that late cost exists in people with working insulin is unmeasured.
Almond calories are overstated by about a quarter on the label, and grinding removes the gap. Lipid encapsulation is the mechanism; it does not generalize to starchy foods (coarse versus fine rye differ by 0.4 points).
Distracted eating costs more at the next meal than at the current one. Plate size is folklore once portion is fixed. “Morning calories burn hotter” is a circadian-rhythm artifact.
Texture-driven eating rate is a real within-ultra-processed lever, clean in one short trial (n=18, 33% less intake at matched density), confounded with fat and protein in the only long one.
The erythritol cohort signal is probably largely endogenous production (serum banked before commercial use still predicts death). That does not answer the platelet question, where the only human intervention (n=10) is positive, so the rubric’s caution stays until the running trial reads out.
Titanium dioxide now has a human crossover trial, but the signal is a transcriptome enrichment with null inflammation markers and no DNA-damage assay. Wording changes, weight does not.
Nitrosyl-heme is null for colorectal cancer in EPIC, but that is one route of the nitrite hypothesis measured crudely; it does not remove nitrite.
GLP-1 drugs dampen wanting; selectivity for ultra-processed food was not detected, with confidence intervals too wide to call it “global”.

Falsified or downgraded:

“Every design-matched trial is null on weight”: positive trials exist (protein-fibre preload n=206, vinegar plus restriction, meal-sequence cluster), all with funding or blinding defects.
Preston 2025 as a “eucaloric” counter-example: weight rose on the ultra-processed arm in both conditions, the registry primary outcome was sperm methylation, and three published letters contest it.
“Erythritol should move down” and “titanium dioxide should move up”: both answer a different question from the one the rubric weight encodes.
Protein dilution as an unmeasured fourth factor: Hall 2019 matched protein and bounded leverage at half the effect.

Hardest open questions carried forward

In the 2026 ward trial, energy density was manipulated by swapping foods, not by hydrating identical foods. What else differed between arms?
What observation would distinguish “ultra-processed is a proxy for density” from “ultra-processed causes overeating via density”?
Why is a bread hazard ratio of 0.65 treated as artifact while a sugary-drink ratio of 1.25 from the same share model is treated as real?
Fazzino, Hall et al. 2023 regressed 2,733 ward meals on density, eating rate and hyperpalatability and found each independently predictive. No single-variable theory has reconciled itself with that result.
For a tray-meal scoring function with no beverages, which of these hypotheses would change a single weight?