What the analysis says
This is the way-of-knowing meta-question: can observational nutritional-cohort associations carry causal weight for egg-health claims at all. It conditions how much every cohort-based cluster counts, so it is pivotal. Three members: H-31 - Nutritional-cohort diet-disease associations reflect cumulative bias and cannot support causal dietary claims (reform pole, cannot), H-41 - The standard critiques of nutritional-cohort epidemiology are manageable, so well-conducted cohorts can support policy-relevant causal dietary inference (defence pole, can, high-confidence policy-grade), and H-59 - Some intermediate epistemic standing (weak/qualitative support for large effects only).
The prior sat at reform/residual co-plurality [0.351, 0.298, 0.351], defence lowest. It rested on the field-level outside-view track record marked used_for_prior: O-69 - Updated cohort meta-analyses find almost all food groups statistically significantly associated with all-cause mortality, O-70 - Taken as lifelong causal, meta-analyzed cohort estimates imply implausible life-expectancy effects including minus 6 years for one egg per day, and the canonical cohort-to-RCT reversal O-71 - Serum beta-carotene shows a strong protective mortality association in cohorts (RR 0.69) that randomized trials exclude. The three no-observation arguments A-36 (replication bounds confounding), A-37 (triangulation, corrected down from “sufficient” to “best-available”), A-38 (reverse causation removable) lifted H-41 modestly but in their hedged forms underwrote H-59 more.
The posterior landed [0.386, 0.183, 0.431]: the intermediate H-59 now leads, reform second, defence sank by ~0.11. Step-8 likelihoods split the evidence. Pulling off defence: E-96 (egg HRs ~1.06-1.14 have small E-values ~1.5-1.7, so a modest healthy-user confounder nulls them; H31=1, H41=0.5, H59=0.85, trust 0.83) and CG-34 (OPEN validation: FFQ under-reports energy ~35%, error BMI-correlated/differential; H31=1, H41=0.45, H59=0.9, trust 0.85). Pulling toward defence: E-103 (non-differential error attenuates toward null; H41=1, H31=0.55) and E-104 (FFQ validity coefficients ~0.4-0.7; H41=1, H31=0.65), both at trust 0.62. The reform-favouring edges carry higher trust, so defence net-sinks. A way-of-knowing verdict, not an object-level one.
What the model may not capture
The residual’s plurality (0.431) is the fragile result. H-59 is a broad grab-bag (“weak/qualitative support for large effects only”) that will attract mass whenever both poles are written as absolutes — its lead may be a framing artifact, not a positive finding. Relatedly, H-41 is stated at its maximal bar (“high degree of certainty for policy”), which A-37’s own correction declined to grant; the actual defence position (Satija) probably sits nearer H-59 than H-41. So H-41’s sinking is partly that it was carved as a strawman — the model is testing the defence against its most extreme form and finding it least supported. A finer carve of the middle (reliable-for-large-effects vs reliable-with-triangulation) would likely redistribute the 0.431.
Both poles trace to named partisans — Ioannidis (S-95), Satija (S-100) — and the used_for_prior observations (O-69/70/71) are the reform side’s exhibits. The counterargument least represented is the field’s genuine hits (LDL/CVD, trans-fat, folate/NTD): named in prose, scored nowhere, exactly what reform-motivated sourcing omits; scored they would lift H-41/H-59.
Is the true answer on the list? A fourth standing is exposure-class-conditional rather than effect-size-conditional: cohorts can carry causal weight for large, biomarker-anchored, well-measured exposures but not for eggs specifically. That nominally collapses into H-59 but is a different axis, and because it conditions every downstream cluster it is at least as consequential as the listed members.
What would help
- A quantitative base rate of cohort-to-RCT concordance across nutritional exposures, stratified by effect size (O-71 is one anecdote) — exists, unread.
- Egg-specific E-values computed against the actual healthy-user confounder with a plausible strength estimate — exists, unread / partly does not exist.
- Whether the uniform healthy-user confounder that survives A-36’s replication carve-out is bounded for eggs — does not exist (A-36 concedes “no unmeasured confounding” is not empirically verifiable).
Confusions and contradictions
The measurement-error evidence genuinely points both ways and the model prices the halves as opposing edges. CG-34’s BMI-correlated differential error (O-93) directly breaks the non-differential premise E-103/O-100 relies on; the same literature yields “attenuates toward null, so observed associations are underestimates” (defence) and “differential and person-specific, so calibration cannot fully fix it” (reform). Which dominates for eggs is unresolved, shipped unfixed. The beta-carotene reversal (O-71) is the second irreducible knot: a biomarker-measured RR 0.69 that RCTs exclude — reform reads it as proof that confounding, not just self-report error, manufactures associations; defence confines it to single-nutrient supplement trials unrepresentative of whole-diet inference. The sources here do not resolve it.
External consensus (post-hoc, entered no number)
Mainstream practicing nutritional epidemiologists sit near hedged-defence/triangulation (Hu, Satija); the evidence-based-medicine / GRADE camp defaults observational nutrition evidence to “low/very low” certainty. The posterior — residual-plurality with reform second — is somewhat more skeptical than working nutritional epidemiology but roughly aligned with the GRADE view. Landing on the intermediate rather than either pole is where the field’s more careful voices also sit.