What the analysis says
This cluster answers the “for whom (diabetics)” crux: in people who already have diabetes, prediabetes, or metabolic syndrome, is habitual egg intake net-safe/beneficial or net-harmful cardiometabolically? It carves onto a single net-verdict axis — a safe/beneficial pole ([[H-4 - High egg intake does not harm blood lipids or glycemic control in people with prediabetes-T2D]] surrogate neutrality, [[H-19 - Substituting eggs for high-carb breakfast foods improves or does not worsen glycemic-insulin outcomes in adults at risk for T2D]] glycemic benefit via substitution) versus a harm pole ([[H-24 - Egg consumption raises CVD risk specifically in people with type 2 diabetes]] endpoint harm, [[H-33 - Impaired insulin sensitivity in T2DM deranges cholesterol transport, making egg-derived dietary cholesterol more atherogenic in diabetics]] its atherogenic-cholesterol mechanism), plus residual [[H-52 - Some other net egg effect in diabetics]]. The prior leaned safe (safe_vs_harm = 1.8): [0.347, 0.212, 0.233, 0.078, 0.130]. The posterior [0.312, 0.183, 0.388, 0.032, 0.086] flips the leader to the harm-endpoint member H-24 and nearly empties the mechanism member H-33.
Two moves drive that. First, three observational diabetic-specific harm signals all anchor on H-24: the KoGES egg-CVD HR of 2.81 in diabetics only ([[CG-11 - HC-4 joint over O-73+O-74]]), the Harvard egg×T2D interaction at P<0.001 with individually null strata ([[E-77 - O-45 x HC-4 — significant egg-T2D interaction on CVD, strata null]]), and the diabetic-restricted meta RR 1.25-1.40 ([[E-78 - O-49 x HC-4 — diabetic-subgroup meta egg-CVD RR 1.25-1.40]]). The RCT surrogate nulls (DIABEGG at ~12 eggs/week, [[CG-9 - HC-4 joint over O-4+O-5+O-6]]; Mutungi; Maki) discriminate only weakly against endpoint harm — the model gives H-24 likelihoods of 0.4-0.55 against them, since a 3-month lipid/glycemic null need not carry a hard-CVD footprint. Second, those same nulls refute H-33: Mutungi held LDL and the LDL:HDL ratio even in hyper-absorbers under 640 mg/day cholesterol (likelihood 0.18), and the KoGES cohort showing HR 2.81 shows no egg-lipid association (0.45), so the mechanism collapses to 0.032 while the endpoint claim it existed to support rises.
What the model may not capture
The exclusivity is imposed, not found. H-4 (“does not harm blood lipids or glycemic control”) and H-24 (“raises CVD risk”) are not logically contradictory — eggs can be surrogate-safe and still raise hard endpoints via a pathway no surrogate captures. The “single net verdict” axis forces a winner between two claims that could both be true, and the KoGES pattern (harm with a null lipid stratum) is precisely that co-holding. The arithmetic reads it as H-24 beating H-4; the reality may be “both,” which the block form cannot express.
The harm pole rests on confounding-prone designs. Diabetics eating more eggs plausibly differ in diet quality, adiposity, and illness severity; the model docks trust (t=0.5 on the 79-case KoGES subgroup, CI lower bound 1.25 and winner’s-curse-inflated per the approved caveat; t=0.75 on the 1-of-10-tests Harvard interaction; t=0.7 on the heterogeneous meta), but docking trust is not modelling that all three signals may share one confounding source and so are not three independent witnesses. The population is also lumped — DIABEGG in prediabetes/T2D, Mutungi in metabolic-syndrome men, the harm cohorts in established T2D — so treating a MetS-men HDL rise and a T2D CVD hazard as one estimand assumes an unargued homogeneity.
Is the true answer on the list? An unlisted candidate is more consequential than any listed one: endpoint harm in diabetics via a non-lipid channel (glycation, TMAO, an insulin-secretion effect) rather than H-33’s cholesterol story — that would produce exactly the observed pattern (CVD excess, no lipid trace) yet is not H-33, and residual H-52’s 0.086 is not obviously carrying it. A second unlisted answer — no single net verdict, because the effect is contingent on glycemic control or background diet — is handled at whole-question level in HC-10 but sits uneasily inside this cluster’s demand for one verdict.
What would help
A long-duration hard-CVD-endpoint RCT of eggs in diabetics — does not exist; it would be decisive and its absence is why surrogate-vs-endpoint cannot be resolved here. Egg-stratified lipoprotein-subclass / ApoB data in diabetics to test H-33 directly — does not exist (the note flags H-33 as untested). Mendelian-randomization or negative-control analyses isolating the egg effect from the Western-pattern confounder in diabetic strata — exists, unread (such MR work is plausibly available and would move the harm pole most). The DIABEGG 12-month extension beyond the 3-month data priced here — exists, unread.
Confusions and contradictions
The cluster’s defining conflict is shipped unresolved. RCT surrogate safety (DIABEGG, Mutungi, Maki all null-to-favorable on lipids and glycemia) and observational diabetic CVD harm (KoGES HR 2.81, Harvard interaction, meta RR 1.25-1.40) point in opposite directions, and the model does not reconcile them — it keeps both poles alive, landing H-24 at 0.388 and H-4 at 0.312 with neither dominant. What makes it genuinely irreducible rather than fixable: the one cohort that reports large diabetic harm (KoGES) reports no lipid difference in the same people, so if the harm is real it travels through a mechanism none of these trials measured, and if it is not real it is confounding — and the assembled evidence cannot presently say which.
External consensus (labelled — entered no number)
Major bodies (ADA, 2020 US Dietary Guidelines) treat eggs as acceptable in moderation for diabetics while several diabetic-subgroup meta-analyses flag an egg-CVD signal, i.e. the field itself is unresolved on exactly this “for whom” question. The posterior’s near-even H-24/H-4 split, tilted to harm, sits within that live disagreement rather than outside it.