Curation note (step 2b) — egg-health case
- initial_count: 101 (
sourcenodes S-1…S-101, contiguous) - live_count: 98 (three TRUE duplicates flagged
duplicate_of: S-56→S-25, S-71→S-48, S-79→S-53) - curated_target_N: 25 (soft; coverage + quality first)
- curated_count: 28
- generated: 2026-07-23
trust_baseline
0.70 (set by the orchestrator; not lowered further here).
Rationale: the cut normally uses baseline 0.8, and the spec says to lower it as a corpus-fit knob when little clears it. Here the reduction is kept deliberately minimal — only 0.10 below default, not the 0.5–0.6 corpus-fit drop the spec would otherwise license — so that “curated” still means genuinely reliable data (the trust_score each source carries is what downstream Bayesian steps rest on). At 0.70, the strong tier (feeding/metabolic-ward + methodology sources up to ~0.85, plus the best nutrition-epi cohorts ~0.72–0.82) ranks cleanly and 24 sources clear the baseline, yielding ~25 by rank before coverage adjustments. combined_score = usefulness · (trust_score − 0.70) is written on all 101 sources.
Comparability: the ten 2a scorers’ slices were checked against the consolidated overview and read as mutually consistent (nutrition-epi ~0.5–0.72; feeding/metabolic-ward + methodology up to ~0.85). No slice was systematically high or low, so no trust rescore was applied — scores are kept as scored. Trust was never deflated to fit the cut.
scoring_rubric
trust_score ∈ [0,1]: probability the key finding survives a clean well-powered replication with a similar effect — judged from design/statistics/corroboration, never from whether the conclusion is believable or surprising. Carried downstream as each source’s data-reliability.usefulness ~1–5(log scale): how much the data would move the main question if true.- Cut key:
combined_score = usefulness · (trust_score − 0.70);≤ 0 ⇒ below baseline. Ranking is arithmetic; departures are coverage/skew decisions logged below.
D-node reconciliation (job 1, done before the cut)
Parallel 2a scorers minted duplicate data-basis nodes for the same dataset. Merged by dataset identity (survivor = lowest id; every source’s data_basis repointed to the survivor; merged nodes moved to data-bases/merged/, none deleted):
| merged | → survivor | dataset | repointed source |
|---|---|---|---|
| D-5 | D-2 | Physicians’ Health Study | S-1 |
| D-19 | D-7 | Harvard NHS/NHSII/HPFS | S-19 |
| D-23 | D-9 | Cohort of Swedish Men | S-45 |
| D-17 | D-12 | China Health & Nutrition Survey | S-13 |
| D-18 | D-13 | EPIC | S-16 |
| D-15 | D-11 | Million Veteran Program | S-66 |
Result: 18 live D nodes (down from 24), each with ≥1 inbound data_basis, no two live D nodes representing the same dataset, and no lingering references to a merged id. All other D nodes are singletons and were left as-is (D-1 LRPP, D-3 DIABEGG, D-4 Cleveland GeneBank, D-6 WHS, D-8 Jackson Heart, D-10 JPHC, D-14 China Kadoorie, D-16 KoGES, D-20 PREDIMED, D-21 SUN, D-22 Ginsberg, D-24 Swedish Mammography).
Fernandez / UConn feeding-lab cluster (correlated, NOT merged): S-30, S-32, S-36, S-72, S-55 (responder subject-pool series) and same-lab distinct-cohort trials S-48, S-53, S-43, S-64, S-59, S-67 are distinct trials sharing lab/PI/protocol and mostly Egg Nutrition Center / American Egg Board funding. They keep self-links ([[S-N]]), not a shared D — but step 5 should treat them as correlated on lab practice + COI (already in motivatedness). Only one member (S-67 Mutungi) is curated, so the concentration does not enter the cut; the rest are non-curated.
angle normalization (before the skew check)
Scorer-introduced variants were collapsed to one consistent 15-tag set (curation_select.py groups on angle): egg-diabetic-subgroup → diabetic-interaction (S-14, S-16, S-17); allergy-rct/allergy-epi → allergy (S-15, S-18); global-cohort → asian-cohort (S-91 PURE, folded into the non-Western/multinational cohort bucket); dietary-pattern → asian-cohort for the multinational MI cohort (S-93 INTERHEART) and → methodology-critique for the confounding-demonstration (S-101 MESA dietary-pattern). Final vocabulary: western-cohort, asian-cohort, egg-T2D-cohort, diabetic-interaction, feeding-rct-lipids, responder-genotype, egg-rct-clinical, tmao-mechanism, tmao-null, nutrition-benefit, foodborne-harm, allergy, methodology-critique, methodology-defence, substitution.
The cut
Coverage-first, because the question is contested and multi-sided. Core = the 24 sources with combined_score > 0 (trust > 0.70). Of these, 16 were kept and 8 were dropped (feeding-rct-lipids and methodology-critique each capped at ~3–4 curated; plus two same-position redundancies). 12 below-baseline sources were added as logged coverage-includes so that every contested position carries ≥1 curated source. Final skew: no angle exceeds 3 curated members and no data_basis exceeds 2 — well under the ~4 ceiling; the script reports no strong skew on the pool. Curated stay in sources/; the 73 non-curated (3 duplicates + 8 skew/redundancy trims + 62 below-baseline) moved to sources/non-curated/. Nothing deleted, no id renumbered.
Final curated set (28), by position
- Western CVD harm — S-21 (coverage)
- Western CVD null — S-19 (core); global null S-91 (core)
- Asian/multinational protective — S-80 (coverage); multinational null S-91 (core)
- Asian within-region heterogeneity (U-shape, “at what level”) — S-81 (coverage)
- egg-T2D US-elevated — S-4 (core)
- egg-T2D null/inverse — S-8 European (core), S-9 Asian/Japan (coverage)
- Diabetic-subgroup CVD interaction — S-14 (coverage)
- Lipid feeding LDL-up mechanism — S-49, S-42 (core)
- Lipid feeding HDL-up (pro-egg) — S-67 (coverage)
- Responder hypo/hyper — S-25 (core)
- ApoE genotype (“for whom”) — S-62 (coverage)
- Egg RCT in diabetics (safe, interventional) — S-26 (coverage)
- Egg RCT adverse counterweight — S-57 (coverage)
- TMAO harm — S-28, S-47 (core)
- TMAO null-from-eggs — S-60 (core)
- Beyond-CVD benefit (choline essentiality) — S-7 (core)
- Allergy — S-15 prevention RCT, S-18 prevalence (core)
- Foodborne — S-20 (core)
- Methodology critique — S-99 (measurement), S-94 (confounding), S-95 (reform viewpoint) (S-99/S-94 core, S-95 coverage)
- Methodology defence — S-100 (coverage)
- Substitution — S-89 (coverage), plus S-57 (substitution-angle)
Angle distribution (all ≤3): egg-T2D-cohort 3, feeding-rct-lipids 3, asian-cohort 3, methodology-critique 3, allergy 2, western-cohort 2, responder-genotype 2, tmao-mechanism 2, substitution 2; diabetic-interaction / methodology-defence / foodborne-harm / egg-rct-clinical / tmao-null / nutrition-benefit 1 each. Shared data_basis in the cut (correlated pairs step 5 must not double-count): D-7 = S-4 + S-19 (Harvard cohorts); D-1 = S-21 + S-89 (Zhong 6-cohort pool). Everything else is a unique D or a self-link.
coverage_includes (12; each combined_score ≤ 0, trust kept as scored)
This is a deliberate, documented extension of the cut’s single-exception default: covering a contested multi-sided question needs several. Each keeps its low trust_score so downstream steps still discount it.
| id | trust | combined | position / reason |
|---|---|---|---|
| S-21 | 0.57 | −0.52 | Western CVD harm anchor (Zhong 6-cohort dose-response) — no harm-direction Western source is core |
| S-80 | 0.60 | −0.38 | Asian protective anchor (China Kadoorie) — protective verdict-direction otherwise unrepresented |
| S-81 | 0.68 | −0.08 | Asian within-region U-shape (China-PAR) — only nonlinearity source, the “at what level” clause |
| S-9 | 0.65 | −0.13 | egg-T2D Asian null (JPHC) — Asian pole of the first-class US-vs-Asia axis |
| S-14 | 0.55 | −0.48 | diabetic-subgroup CVD interaction (Jang, HR 2.81) — “for whom (diabetics)“ |
| S-26 | 0.70 | 0.00 | DIABEGG RCT — sole interventional “high-egg safe in T2D” anchor |
| S-57 | 0.62 | −0.23 | Maki adverse-LDL substitution — the one counterweight to the null-to-beneficial egg-RCT slice |
| S-62 | 0.60 | −0.32 | Weggemans ApoE null (largest pool) — the “for whom (genotype)” disagreement |
| S-67 | 0.55 | −0.41 | Mutungi whole-egg-raises-HDL — pro-egg lipid side otherwise absent (Fernandez/ENC-AEB COI) |
| S-89 | 0.62 | −0.30 | Zhong substitution analysis — substitution position otherwise unrepresented (shares D-1 with S-21) |
| S-95 | 0.60 | −0.30 | Ioannidis “nutritional-epi needs reform” — the meta-level critique viewpoint explicitly required |
| S-100 | 0.62 | −0.22 | Satija defence of nutritional-epi — balances the critique cluster |
rank_departures
Coverage-includes added out of rank (below-baseline): S-21, S-80, S-81, S-9, S-14, S-26, S-57, S-62, S-67, S-89, S-95, S-100 (see table above).
Core (combined > 0) dropped from the rank-cut:
- S-1 (0.06) — redundant with S-4 on the US-elevated egg→T2D position (same direction); dropped for coverage room.
- S-45 (0.07) — redundant with S-19 on the Western-null position; dropped for coverage room.
- S-35 (0.25) — feeding-rct-lipids skew cap; Sacks LDL-rise redundant with kept Keys/Ginsberg.
- S-44 (0.20) — feeding-rct-lipids skew cap; shares D-22 (Ginsberg) with kept S-42, dropped for independence.
- S-52 (0.41) — feeding-rct-lipids skew cap; Mattson metabolic-ward dose-response redundant with kept Keys equation.
- S-90 (0.45) — methodology-critique skew cap; substitution-caution overlaps kept substitution (S-89) and critique set.
- S-97 (0.34) — methodology-critique skew cap; cookbook-test data-dredging critique overlaps kept OPEN/E-value/Ioannidis.
- S-98 (0.05) — methodology-critique skew cap; GRADE red-meat template carries no egg data.
Skew consciously accepted: none outstanding. The two capped angles (feeding-rct-lipids, methodology-critique) are trimmed to 3 curated each; no angle > 3 and no data_basis > 2 in the final cut, so curation_select.py reports no strong skew. The field dimension is dominated by nutrition/epidemiology sub-labels — inherent to the question, not a correctable skew. The only correlated pairs (D-7, D-1) are intentional and handed to step 5 via shared D-node identity.
step1_corrections
Factual errors the 2a scorers found in step-1 summaries (so step 3 reads the paper, not the wrong summary). For the three that ended up curated, a one-line **Correction (2a read):** was appended to the node body (additive):
- S-9 Kurotani 2014 (JPHC) — NULL egg→T2D in both sexes. The “40% lower risk in men” was a mis-summary; the inverse OR was for dietary cholesterol in postmenopausal women, not eggs. (appended — curated)
- S-42 Ginsberg 1994 — a 20-man, 8-week crossover feeding 0/1/2/4 eggs/day (128–858 mg cholesterol/d), not a “P:S-ratio 0/750/1500 mg” design. (appended — curated)
- S-47 Miller 2014 — an n=6 within-subject dose-response (0/1/2/4/6 yolks), not a larger-n trial. (appended — curated)
- S-16 Trichopoulou 2006 — the egg-specific HR is secondary-sourced (via the DMSO diabetic-CVD review), not the paper’s headline endpoint. (recorded only — non-curated)
- S-92 Djoussé 2021 — a 9-US-cohort pooling (not Harvard-dominant); self-linked, and shares MESA with S-101. (recorded only — non-curated)
curation_select.py output (baseline 0.70, target-n 25) — pasted
baseline 0.70 | 101 scored | 24 clear the baseline | cut = top 25
RANKING (* = in the prospective cut)
id comb trust use title
* S-19 0.50 0.82 4.2 Drouin-Chartier 2020 BMJ - egg consumption and CVD i
* S-99 0.50 0.85 3.3 The OPEN study- how much self-reported diet data act
* S-49 0.48 0.85 3.2 Keys, Anderson & Grande 1965 Metabolism — foundation
* S-20 0.46 0.83 3.5 St Louis 1988 - Emergence of grade A eggs as a major
* S-90 0.45 0.85 3.0 Why substitution analysis in nutritional epidemiolog
* S-52 0.41 0.85 2.7 Mattson, Erickson & Kligman 1972 AJCN — Procter & Ga
* S-15 0.40 0.80 4.0 Natsume 2017 - PETIT trial- two-step egg introductio
S-56 0.36 0.82 3.0 McNamara et al. 1987 JCI — heterogeneity of choleste [duplicate]
* S-42 0.34 0.82 2.8 Ginsberg 1994 Arterioscler Thromb — dose-response of
* S-97 0.34 0.82 2.8 The cookbook test- almost every ingredient is 'linke
* S-94 0.33 0.83 2.5 The E-value- how much unmeasured confounding would i
* S-4 0.32 0.78 4.0 Drouin-Chartier 2020 AJCN - egg consumption and T2D
* S-91 0.32 0.78 4.0 PURE study- egg intake, blood lipids, CVD and mortal
* S-35 0.25 0.80 2.5 Sacks 1984 Lancet — egg feeding raises LDL and ApoB
* S-18 0.20 0.78 2.5 Eggesbo 2001 - Population-based prevalence of egg al
* S-44 0.20 0.78 2.5 Ginsberg 1995 ATVB — dietary cholesterol-egg dose-re
* S-8 0.18 0.75 3.5 Wallin 2016 Diabetologia - egg consumption and T2D i
* S-28 0.09 0.72 4.3 Gut flora metabolism of phosphatidylcholine promotes
* S-47 0.08 0.73 2.8 Effect of egg ingestion on trimethylamine-N-oxide pr
* S-25 0.07 0.72 3.3 McNamara et al 1987 - Heterogeneity of cholesterol h
* S-45 0.07 0.72 3.3 Larsson, Akesson and Wolk 2015 AJCN - egg consumptio
* S-60 0.06 0.72 3.2 Dietary choline supplements, but not eggs, raise fas
* S-1 0.06 0.72 3.0 Djoussé 2009 Diabetes Care - egg consumption and T2D
* S-7 0.05 0.72 2.5 Zeisel 1991 - Choline, an essential nutrient for hum
* S-98 0.05 0.72 2.3 GRADE-ing red-meat RCT evidence as a template for re
S-26 0.00 0.70 3.5 DIABEGG 3-month RCT — high-egg vs low-egg diet in pr [below baseline]
S-22 0.00 0.70 3.3 Katan and Beynen 1986 - Existence of consistent hypo [below baseline]
S-31 0.00 0.70 2.8 DIABEGG weight-loss and 12-month follow-up phase in [below baseline]
S-68 0.00 0.70 3.4 TMAO response to animal source foods varies among he [below baseline]
S-87 0.00 0.70 3.3 Japan Public Health Center cohort- egg intake not li [below baseline]
S-92 0.00 0.70 3.5 Egg-diabetes-CHD risk after adjusting for overall di [below baseline]
S-63 -0.06 0.68 2.8 Quintao, Grundy & Ahrens 1971 JLR — dietary choleste [below baseline]
S-39 -0.06 0.68 3.0 Intestinal microbiota metabolism of L-carnitine, a n [below baseline]
S-83 -0.06 0.68 3.0 Guangzhou Biobank Cohort Study egg consumption and C [below baseline]
S-40 -0.06 0.68 3.2 Gut microbial metabolite TMAO enhances platelet hype [below baseline]
S-48 -0.06 0.68 3.2 Intake of up to 3 eggs-day increases HDL cholesterol [below baseline]
S-81 -0.08 0.68 4.0 China-PAR project- U-shaped egg intake vs incident C [below baseline]
S-29 -0.08 0.68 4.0 Intestinal microbial metabolism of phosphatidylcholi [below baseline]
S-78 -0.09 0.65 1.8 Crowe et al 2013 AJCN - ischemic heart disease risk [below baseline]
S-9 -0.13 0.65 2.6 Kurotani 2014 BJN - cholesterol and egg intakes and [below baseline]
S-17 -0.15 0.65 3.0 Díez-Espino 2017 Clinical Nutrition - egg consumptio [below baseline]
S-33 -0.19 0.65 3.8 Key et al 2019 Circulation - meat, fish, dairy, and [below baseline]
S-101 -0.21 0.62 2.6 Eggs load onto the same 'Western' dietary pattern as [below baseline]
S-100 -0.22 0.62 2.7 In defence of nutritional epidemiology's methods and [below baseline]
S-93 -0.22 0.60 2.2 INTERHEART- eggs within the 'Western' dietary patter [below baseline]
S-32 -0.23 0.60 2.3 Herron et al 2002 J Am Coll Nutr - Pre-menopausal wo [below baseline]
S-41 -0.23 0.60 2.3 Dawber et al 1982 AJCN - eggs, serum cholesterol, an [below baseline]
S-57 -0.23 0.62 2.9 Substituting eggs for high-carbohydrate breakfast fo [below baseline]
S-50 -0.24 0.62 3.0 Abdollahi et al 2019 AJCN - egg and cholesterol inta [below baseline]
S-11 -0.25 0.60 2.5 Lee & Kim 2018 Nutrition Research and Practice - egg [below baseline]
S-12 -0.25 0.60 2.5 van Vliet 2017 - Whole eggs versus egg whites and po [below baseline]
S-53 -0.26 0.60 2.6 Greene et al 2006 Nutr Metab - Plasma LDL and HDL ch [below baseline]
S-10 -0.27 0.60 2.7 Vishwanathan 2009 - Egg yolks increase macular pigme [below baseline]
S-54 -0.27 0.60 2.7 Nettleton et al 2008 J Am Diet Assoc - egg, whole-gr [below baseline]
S-69 -0.27 0.62 3.4 TMAO is associated with mortality- impact of modestl [below baseline]
S-58 -0.28 0.60 2.8 Qureshi et al 2007 Med Sci Monit - regular egg consu [below baseline]
S-2 -0.30 0.55 2.0 Vander Wal 2005 - Short-term effect of eggs on satie [below baseline]
S-95 -0.30 0.60 3.0 The case that nutritional-epidemiology cohort method [below baseline]
S-89 -0.30 0.62 3.8 Substitution analysis- replacing eggs with other pro [below baseline]
S-62 -0.32 0.60 3.2 Weggemans, Zock, Ordovas, Pedro-Botet and Katan 2001 [below baseline]
S-51 -0.33 0.55 2.2 Selenium-enriched vs zeaxanthin-enriched eggs in T2D [below baseline]
S-36 -0.34 0.55 2.3 Herron et al 2004 Metabolism - High cholesterol inta [below baseline]
S-59 -0.34 0.55 2.3 Effects of egg consumption and choline supplementati [below baseline]
S-24 -0.35 0.60 3.5 USDA-FSIS Salmonella Enteritidis risk assessment for [below baseline]
S-82 -0.36 0.58 3.0 Katz et al. 2005 Int J Cardiol — egg vs. oatmeal cro [below baseline]
S-61 -0.37 0.55 2.5 Eggs in plant-based diets, cardiometabolic risk in a [below baseline]
S-80 -0.38 0.60 3.8 China Kadoorie Biobank egg consumption and CVD (Qin [below baseline]
S-66 -0.38 0.58 3.2 Djousse et al 2020 Clinical Nutrition - egg consumpt [below baseline]
S-6 -0.39 0.55 2.6 Djoussé 2016 Clinical Nutrition - egg consumption an [below baseline]
S-34 -0.40 0.50 2.0 One egg-day vs oatmeal breakfast crossover trial in [below baseline]
S-38 -0.40 0.50 2.0 Zazpe et al 2011 Eur J Clin Nutr - egg consumption a [below baseline]
S-55 -0.40 0.50 2.0 Missimer, DiMarco and Fernandez 2019 Curr Dev Nutr - [below baseline]
S-67 -0.40 0.55 2.7 Mutungi et al. 2008 J Nutr — egg-derived cholesterol [below baseline]
S-30 -0.42 0.55 2.8 Herron et al 2003 J Nutr - Men classified as hypo- o [below baseline]
S-74 -0.42 0.55 2.8 Kern 1991 NEJM - Normal plasma cholesterol in an 88- [below baseline]
S-70 -0.43 0.55 2.9 Lopez-Miranda et al 1994 J Lipid Res - Effect of apo [below baseline]
S-73 -0.43 0.55 2.9 Njike et al. 2010 Nutrition Journal — daily egg cons [below baseline]
S-13 -0.45 0.55 3.0 Wang, Li and Shi 2020 BJN - higher egg consumption a [below baseline]
S-43 -0.45 0.55 3.0 Whole egg vs egg substitute in carbohydrate-restrict [below baseline]
S-88 -0.45 0.55 3.0 NIPPON DATA80- egg consumption, cholesterol, and cau [below baseline]
S-75 -0.45 0.52 2.5 Kim & Campbell 2018 Nutrients — dietary cholesterol [below baseline]
S-14 -0.48 0.55 3.2 Jang 2018 Nutrition and Diabetes - egg consumption a [below baseline]
S-84 -0.48 0.55 3.2 Rural China (Anhui) 14-year cohort- high egg intake [below baseline]
S-64 -0.48 0.50 2.4 Eggs plus spinach in a plant-based diet, metabolic s [below baseline]
S-27 -0.49 0.55 3.3 Djousse and Gaziano 2008 AJCN - egg consumption, CVD [below baseline]
S-76 -0.49 0.55 3.3 Al-Ramady et al 2022 Clin Nutr ESPEN - egg consumpti [below baseline]
S-86 -0.50 0.52 2.8 China Family Panel Studies nationwide cohort- egg in [below baseline]
S-37 -0.52 0.50 2.6 Energy-restricted high-protein diet with eggs vs lea [below baseline]
S-21 -0.52 0.57 4.0 Zhong 2019 JAMA - dietary cholesterol-egg consumptio [below baseline]
S-46 -0.54 0.50 2.7 One egg-day vs egg substitute in pre--type-2-diabete [below baseline]
S-77 -0.54 0.50 2.7 Vorster et al. 1992 AJCN — egg intake does not chang [below baseline]
S-16 -0.56 0.50 2.8 Trichopoulou 2006 J Intern Med - diet, physical acti [below baseline]
S-79 -0.56 0.50 2.8 Greene et al. 2006 Nutrition & Metabolism — egg feed [below baseline]
S-85 -0.58 0.52 3.2 China Health and Nutrition Survey- habitual egg inta [below baseline]
S-71 -0.59 0.48 2.7 DiMarco et al. 2017 Lipids — dose-escalation egg RCT [below baseline]
S-65 -0.66 0.48 3.0 Sarkkinen et al 1998 AJCN - Effect of apolipoprotein [below baseline]
S-3 -0.67 0.45 2.7 Vander Wal 2008 - Egg breakfast enhances weight loss [below baseline]
S-23 -0.70 0.42 2.5 Djousse and Gaziano 2008 Circulation - egg consumpti [below baseline]
S-96 -0.87 0.40 2.9 The 'fatal flaws' case against food-frequency-questi [below baseline]
S-5 -0.92 0.42 3.3 Caudill 2018 - Maternal choline supplementation impr [below baseline]
S-72 -0.95 0.36 2.8 Herron et al 2006 J Nutr - The ABCG5 polymorphism co [below baseline]
REPRESENTATION OF THE CUT
angle:
methodology-critique 5/24 of cut, 5 in pool
feeding-rct-lipids 5/24 of cut, 5 in pool
egg-T2D-cohort 3/24 of cut, 3 in pool
western-cohort 2/24 of cut, 2 in pool
allergy 2/24 of cut, 2 in pool
tmao-mechanism 2/24 of cut, 2 in pool
foodborne-harm 1/24 of cut, 1 in pool
asian-cohort 1/24 of cut, 1 in pool
responder-genotype 1/24 of cut, 1 in pool
tmao-null 1/24 of cut, 1 in pool
nutrition-benefit 1/24 of cut, 1 in pool
data_basis:
unknown 3/24 of cut, 3 in pool
D-22 2/24 of cut, 2 in pool
(remaining values 1/24 each)
No strong skew detected in the prospective cut.
Note: the pasted ranking is the script’s pure-rank prospective cut (top-24 that clear baseline 0.70). The applied cut departs from it per the coverage-first mandate above — 8 of those 24 were trimmed (skew caps + two redundancies) and 12 below-baseline coverage-includes were added — giving the final 28. The script’s angle skew flags (feeding-rct-lipids 5, methodology-critique 5) are resolved in the applied cut, where each is trimmed to 3.