What the analysis says
The cluster asks which exposure carries the egg-CVD association: H-3 - Dietary fat quality is a larger determinant of plasma cholesterol than dietary cholesterol quantity accompanying dietary fat quality (saturation), H-26 - Dietary cholesterol content, not egg per se, is the operative exposure behind the egg-CVD association egg-derived dietary cholesterol, or residual H-54 - Some other operative exposure (a non-cholesterol egg component, or no real exposure — pure dietary-pattern confounding). The prior [H-3, H-26, H-54] = [0.451, 0.324, 0.225] moved to posterior [0.475, 0.182, 0.342]: fat quality stays ahead, the pure-cholesterol thesis roughly halves, and the residual/artefact pole climbs most. Three witnesses drove it. CG-13 - HC-5 joint over O-1+O-3 (McNamara 1987, trust 0.72) anchors H-3 — 69% of metabolic-ward periods buffered added cholesterol and plasma cholesterol tracked fat quality over cholesterol dose (H-26 scored 0.35x). CG-14 - HC-5 joint over O-41+O-58+O-60 (Zhong pooled cohorts, trust 0.57) anchors H-26 — a monotonic dietary-cholesterol dose-response plus O-60 - Egg-CVD and egg-mortality associations become non-significant after adjusting for dietary cholesterol (6 US cohorts), the egg signal going non-significant specifically on cholesterol adjustment. CG-15 - HC-5 joint over O-43+O-44 (Harvard cohorts, trust 0.82) anchors H-54 — the crude egg-CVD association (HR 1.10) reversing to null (0.93) on lifestyle adjustment, the confounding signature. The higher-trust confounding and buffering witnesses outweigh the lower-trust cholesterol one, which is why H-26 collapses.
What the model may not capture
The three members are not cleanly exclusive: in every cohort cholesterol, saturated fat, and an unhealthy-eater pattern co-vary, so “the operative exposure” presumes a single lever where reality may be a small real cholesterol effect swamped by fat and pattern. H-54 also pools two very different answers — “artefact, no exposure” and “a non-cholesterol component (choline to TMAO, fed by CG-16/17/18)” — into one 0.342 mass, so a reader cannot tell whether the residual means “it’s confounding” or “it’s TMAO.” Is the answer on the list? A “genuinely multifactorial, no single operative exposure” answer is closest to but not the same as the artefact pole (HC-6 says a small real cholesterol effect does exist), and it would matter more than any listed member because it dissolves the “act on eggs vs act on the cholesterol in eggs” framing the cluster exists to serve.
What would help
Within-cohort mediation separating egg cholesterol from co-consumed fat and lifestyle — exists, unread (the pooled cohorts hold the covariates) to does not exist (a clean instrument). Whether the non-cholesterol (TMAO) branch is real — handled in HC-8/HC-9, exists elsewhere.
Confusions and contradictions
The core tension: O-60 - Egg-CVD and egg-mortality associations become non-significant after adjusting for dietary cholesterol (6 US cohorts) (egg signal vanishes on cholesterol adjustment, trust 0.57) reads as cholesterol mediation and anchors H-26, while O-44 (egg signal vanishes on lifestyle adjustment, not cholesterol, trust 0.82) reads as confounding and anchors H-54 — the same statistical move, opposite causal meaning, in two different cohorts. Cohorts cannot tell mediation from confounding here, so the cholesterol-adjustment attenuation ends up cutting against, not for, the pure-cholesterol thesis. Genuinely irreducible from the observational corpus.