Skip to content

Curation log — type 2 diabetes

2026-08-28 — Seed

  • Created: INDEX.md (canonical 16-row Pages table), wiki/overview.md (draft, 39 verified PMIDs), OPEN-QUESTIONS.md (8 seed questions), this log.
  • Searches run (live PubMed, esearch/esummary/efetch + MCP PubMed tools): IDF Diabetes Atlas 2021 estimates; GBD 2021 diabetes burden; EMPA-REG OUTCOME; LEADER (+renal analysis); SUSTAIN-6; PIONEER 6; CREDENCE; FLOW; SELECT (+SELECT weight/HbA1c analyses); SURPASS-2; SURMOUNT-1/-2; tirzepatide-vs-semaglutide obesity head-to-head; BMJ network meta-analysis and living guideline; DiRECT 1-/2-/5-year; remission consensus definition; twin-cycle/calorie-restriction model; STAMPEDE 5-year; UKPDS 33/34/80/91; GRADE; DPP and DPPOS 15-year; Look AHEAD long-term mortality; TODAY youth-onset complications; ADA/EASD 2022 consensus; ADA Standards of Care 2026 sections; precision diabetes medicine consensus; diabetes stigma consensus.
  • Verification: all cited PMIDs retrieved from live PubMed this session; abstracts fetched for all quantitative claims (prevalence, hazard ratios, remission rates, weight-loss percentages). Volume/page details included only where returned live.
  • Next: full build queued per CONDITIONS-ROADMAP.md — 15 remaining wiki pages, literature layer, ClinicalTrials.gov verification (obesity-diabetes convergence pipeline), patient-voice layer.

2026-08-28 — Full build

  • Built all 16 canonical pages: expanded overview.md; added epidemiology/burden, pathophysiology/subtypes, prevention, remission/weight, incretins, SGLT2/kidney/heart, glycaemic legacy trials, metabolic surgery, complications, youth onset, guidelines, biomarkers, clinical-trials landscape, patient experience/advocacy, and red flags/safety. All remain draft pending the required separate audit.
  • Density: every wiki page is 154–206 lines and carries 25–39 unique verified PMIDs; tables are used for thresholds, trial comparisons, complication surveillance, treatment trade-offs, pipeline status, cross-domain evidence and safety controls.
  • Literature layer: generated BIBLIOGRAPHY.md with 106 unique cited PMIDs and cited-by mapping; wrote six landmark notes (UKPDS, DPP, EMPA-REG, DiRECT, CREDENCE, FLOW); built guideline registry, statistics tables and four-file patient-voice layer.
  • Open questions: replaced seed list with 18 stable tiered questions (OQ-1 to OQ-18) and 14 “Dots not yet connected” junctions.
  • PubMed searches run: global IDF/GBD prevalence and projection; undiagnosed disease; global/LMIC economic burden; mortality and incidence trends; young-adult onset; sex differences; urban/rural and slum prevalence; DALYs and attributable diet/PM2.5; insulin resistance/β-cell failure/ominous octet; twin-cycle and ectopic fat; Ahlqvist clusters and precision diabetes; DPP/DPPOS and Da Qing prevention; DiRECT/remission durability; bariatric remission and long-term RCTs; GLP-1 CVOTs; tirzepatide/SURPASS; withdrawal/weight regain; SGLT2 cardiovascular/kidney trials and DKA; UKPDS/ACCORD/ADVANCE/VADT/GRADE; complications; TODAY/youth β-cell failure; HbA1c/C-peptide/genetic/multi-omic biomarkers; diabetes distress, stigma, treatment burden and indigenous care; current guideline and safety searches.
  • ClinicalTrials.gov: queried the v2 condition endpoint (11,859 broad records returned) and exact study endpoints. Twelve NCT identifiers written in the condition were re-resolved live; tracked records include EMPA-REG, LEADER, SUSTAIN-6, CREDENCE, FLOW, SELECT, SURPASS-CVOT, SOUL, retatrutide, CagriSema and tirzepatide CKD programmes.
  • Patient-voice web verification: fetched official IDF, ADA, Diabetes UK and Diabetes Australia pages; public sources only, no private-person identifiers or private forums.
  • Validation: batch PubMed EFetch resolved all 106/106 PMIDs written anywhere under the condition; ClinicalTrials.gov v2 resolved 12/12 NCT IDs; all 16 pages passed frontmatter/status, required-heading, 150–400-line and local-link checks.
  • Conflicts preserved: IDF vs GBD prevalence models; remission denominators and definitions; UKPDS legacy vs ACCORD-era intensive-control harm; class versus molecule outcome inference; WHO resource-sensitive older sequencing vs modern organ-protection guidance; surgery versus incretin evidence horizons.
  • Not verified/omitted: no global single undiagnosed fraction was asserted; no prevalence was inferred from patient-organization campaigns; no claim that data-driven clusters improve treatment outcomes; no unpublished SURPASS-CVOT numeric result was added. Exact guideline wording may change after this date and requires live re-fetch at audit/update.
  • Next: run the playbook’s separate audit session: re-fetch every claim→PMID pair, verify author/year/numbers and only then promote passing pages to curated and roadmap status to audited.

2026-08-30 — Page-by-page deepening

  • Deepened all 16 canonical wiki pages in place: preserved the prior verified text and added page-specific sections rather than restarting pages. New material covers Steno-2 multifactorial outcomes; NCD-RisC prevalence and treatment coverage; global diagnosis-to-control cascades; Whitehall II pre-diagnostic trajectories; β-cell failure mechanisms; Finnish DPS, IDPP-1, Da Qing and DPPOS prevention; Look AHEAD/SOS remission outcomes; ELIXA, AMPLITUDE-O, SOUL and SURPASS-CVOT; HFpEF/CKD SGLT2 evidence; VADT/ACCORD Eye legacy conflicts; long-term surgical outcomes and harms; MASLD/foot burden; RISE, Ellipse and DINAMO youth evidence; ACP/AACE/ESC guideline disagreements; diagnostic/C-peptide/genomic biomarkers; patient-reported outcomes, cost and distress; and DKA–HHS/hypoglycaemia safety.
  • Density after deepening: pages are 185–248 lines and contain 32–49 distinct PMIDs each. All remain draft; this was a build/deepening session, not the playbook's independent audit session.
  • Literature searches: ran more than 90 targeted live PubMed ESearch queries across the 16 page domains, followed by ESummary and EFetch for selected records and abstracts. Searches included exact landmark trial/programme names plus systematic-review, cohort, mechanism, diagnostic-accuracy, implementation, equity, safety and guideline queries.
  • Bibliography: increased from 106 to 240 unique PubMed records. Every new wiki PMID has a house-format bibliography line with tags and cited-by page mapping.
  • ClinicalTrials.gov: re-fetched all 12 NCT registrations through the v2 study API on 2026-08-30. Eight landmark trials and the retatrutide obesity/CVD study returned COMPLETED; tirzepatide CKD/obesity, REDEFINE 3 and TRIUMPH-Outcomes returned ACTIVE_NOT_RECRUITING. Added published SOUL and SURPASS-CVOT outcomes to the trials page.
  • Statistics and guidelines: expanded STATISTICS.md with care-cascade, prevention, remission, surgery, organ-protection, monitoring, youth and crisis tables. Added ACP 2024, AACE 2022/2023 and ESC 2023 to REGISTRY.md, retained resource-sensitive disagreement, and made metformin prerequisite, HbA1c target and DPP-4 positions explicit.
  • Open questions: sharpened existing combination, youth and retinopathy questions; expanded from 18 to 26 stable questions and from 14 to 22 cross-domain junctions, including cascade repair, youth disease modification, remission causality, target-strategy trials, MASLD screening, CGM outcomes, active-comparator sequencing and actionable patient-reported outcomes.
  • Verification: performed a final live PubMed EFetch over every PMID present anywhere under the condition; 240/240 resolved. Re-resolved 12/12 NCT IDs live. All 16 pages passed frontmatter, required-heading, 150–400-line, local-link and inline-PMID-to-reference checks. No excluded condition, convention, tool or frontend path was changed.
  • Conflicts preserved: global totals by IDF/GBD/NCD-RisC; treatment percentage versus absolute untreated count; cluster labels versus continuous traits; prediabetes intervention versus overmedicalisation in older adults; remission association versus causality; active- versus placebo-comparator CVOTs; kidney composite definitions; ACP versus ADA/AACE targets; surgery durability versus contemporary pharmacotherapy; and conflicting rare pancreatitis/biliary signals.
  • Next: a separate audit session should compare each newly added quantitative claim to the full paper (not only abstract), check updated guideline wording at source websites, and promote pages to curated only after claim-level review.

2026-08-30 — Independent AUDIT pass

  • Scope checked: all 16 canonical wiki pages and all 13 literature artifacts (master bibliography, guideline registry, statistics register, six landmark notes and four patient-voice files), plus INDEX.md and OPEN-QUESTIONS.md for consistency.
  • Citation counts checked: 246 unique current PMIDs across 1,797 pre-log citation occurrences; all 246/246 were re-resolved through live PubMed E-utilities after the final repair pass. The wiki contains 770 inline PMID occurrences and 587 reference-list occurrences; every inline PMID is present in its page reference list, every reference is used, and all 246 bibliography cited-by mappings match the pages exactly.
  • Trial counts checked: 12 unique NCT identifiers across 27 occurrences; all 12/12 were re-resolved by exact identifier through the live ClinicalTrials.gov v2 study API. Nine records were COMPLETED and three were ACTIVE_NOT_RECRUITING on 2026-08-30. The broad condition endpoint still returned 11,859 records and remains labelled as a broad, non-drug-development count.
  • Web/guideline checks: re-checked 15 official public-source records/URLs. NICE's February 2026 NG28 wording was updated in the synthesis; KDIGO 2022 remains the current guideline while a 2026 update is a public-review draft and is not presented as final. Patient-organisation access dates were refreshed to 2026-08-30.
  • Wrong citation/report pairs fixed (2): HARMONY Outcomes used an unrelated lung-cancer PMID (30280657) and now uses the live-matched report (PMID 30291013); GRADE glycaemic durability now uses the glycaemic-outcomes report (PMID 36129996), while the companion microvascular/cardiovascular report remains correctly separated (PMID 36129997).
  • Metadata and framing defects fixed: EMPA-KIDNEY was normalised from online-release year 2022 to its PubMed journal citation, 2023; three abbreviated bibliography titles were restored to their live titles; a renal-glucose-reabsorption mechanism claim was moved from an outcome trial to the mechanism review (PMID 20206731); four orphaned reference records were removed; and cited-by sections in all six landmark notes were reconciled to the current wiki.
  • Stale absences repaired: live dated searches were rerun for GLP-1RA+SGLT2 combination outcomes, surgery versus GLP-1RA, anti-stigma interventions, pharmacological remission terminology/trials, incretin pancreatitis/biliary safety, biomarker-guided care, cluster-guided care, C-peptide remission thresholds, post-remission surveillance, MASLD screening strategies, CGM complication outcomes, retinopathy titration, SGLT2 sick-day implementation, youth β-cell preservation and modern HbA1c-target strategies.
  • New evidence incorporated: a target-trial emulation of layered RAASi+SGLT2i+GLP-1RA therapy (PMID 42658186); observational surgery-versus-GLP-1RA synthesis (PMID 42660781); a 21-clinician stigma-training pilot (PMID 42579884); 2025–2026 pancreatitis/biliary meta-analyses (PMID 39569606; PMID 40988099; PMID 41927408); and emerging pharmacological-remission mapping, observational definitions and terminology (PMID 40419497; PMID 41142657; PMID 41852334). Each is labelled by design and limitations; observational evidence is not presented as randomised proof.
  • Evidence-gap wording: genuine gaps are now positively and explicitly dated—what the search found is stated before the missing strategy/outcome evidence. No claim of absence relies solely on the prior build search.
  • Final validation: all 16 pages are 186–253 lines, contain 32–50 unique PMIDs, and pass required-frontmatter, TL;DR, open-question, related-page, reference-use and local-link checks. Remaining unverified flags: 0. Remaining substantive issues: 0. All 16 pages were promoted to status: curated; INDEX.md and CONDITIONS-ROADMAP.md were advanced to audited.

2026-10-07 — Grounding of uncited dots

  • Grounder: Grok 4.7.
  • Scope: the 22 rows in OPEN-QUESTIONS.md under "Dots not yet connected" that carried no PMID and no NCT ID (D-1 through D-22). OQ-1 through OQ-26 were not rewritten. The page's "Last curated: 2026-08-30" line was left as it stood. No NCT ID was written. Trial numbers that appear inside abstracts were not re-resolved on ClinicalTrials.gov, so they are not in the table.
  • Counts: started with 22 uncited rows. Rewrote 17: D-1, D-2, D-3, D-4, D-5, D-6, D-8, D-9, D-10, D-12, D-15, D-16, D-17, D-18, D-20, D-21, D-22. Deleted 5. Survivor IDs were not renumbered.
  • Citation check: every PMID written into a kept row was fetched from PubMed E-utilities on this date. Publication types were checked for Retracted Publication and Expression of Concern. The article's own DOI (Medline ELocationID, or the article-level ArticleId when that element was absent) was checked at https://api.crossref.org/works/<doi> for updated-by. Reference-list DOIs inside the PubMed XML were not used: an early pass that used them falsely attached other papers' errata to Stanley (PMID 42174253) and Stafford (PMID 40934935). Their article DOIs are clean (10.1038/s41591-026-04416-4 and 10.1016/S2213-8587(25)00217-7). No cited record was a retracted publication or an expression of concern.
  • Figures deliberately not quoted. The DiRECT five-year correction (PMID 38723645, DOI 10.1016/S2213-8587(24)00128-1, publication type Published Erratum) has an empty PubMed abstract, so the extension's remission count and mean weight loss are not in D-1. Campbell 2026 prints a final-A1c confidence interval of "−1.66 to .96", which does not contain the point estimate; that interval is not in D-10. Natale 2025 (PMID 39963952) was fetched; the kidney-failure sentence in the abstract is cut off mid-phrase, so no effect estimate from it was written. Its selection criteria are GLP-1 receptor agonist versus placebo, standard care, or another glucose-lowering agent, which is not combination versus either class.

Deleted, and why

  • D-7 (ectopic fat versus multi-ancestry body composition). Taylor 2024 (PMID 39038473) was re-fetched. The abstract gives the twin-cycle account and no multi-ancestry body-composition number and no operational fat-threshold cut-point. Crossref updated-by on DOI 10.1016/S2213-8587(24)00157-8 is an erratum (10.1016/s2213-8587(24)00244-4); the correction text was not retrieved. One side has no quantitative finding, and the junction restates OQ-12.
  • D-11 (modelled dietary risk versus local food insecurity and programme reach). O'Hearn 2023 (PMID 37069363; Crossref clean) attributes 14.1 million incident cases in 2018, 70.3% (95% uncertainty interval 68.8–71.8) of new cases, to 11 dietary factors. That is a comparative-risk model, the same kind of claim as the GBD diet fraction. GBD 2021 diabetes (PMID 37356446) has two Crossref errata, so its DALY fraction was not used. No cited record supplies local food insecurity or programme reach as a second literature, and none was manufactured to keep the row.
  • D-13 (cardiovascular overlap versus noninvasive fibrosis triage). Driessen 2025 (PMID 38147315; Crossref clean) is a narrative review. The abstract has no effect size and no FIB-4 or elastography performance. Borrowing Younossi's prevalence would have repeated the unsupported half of D-19.
  • D-14 (euglycaemic DKA versus an EHR hold–ketone–restart pathway). Dutta 2022 (PMID 35495849; Crossref clean) supports the case literature: surgery was the precipitant in 15/77 patients, and the pooled DKA relative risk for an SGLT2 inhibitor versus other medication was 3.70 (95% CI 2.58–5.29). The EHR pathway is the trial proposed in OQ-7, not a second set of findings.
  • D-19 (prevalence versus sequential noninvasive tests). Younossi 2024 (PMID 38521116; Crossref clean) reports NAFLD/MASLD prevalence 65.33% (95% CI 62.35–68.18) in 123 studies (N=2,224,144 with type 2 diabetes) and advanced fibrosis 15.49% (6.99–30.99) in a biopsy subset (12 studies, N=2,733). Prevalence is not sequential-test accuracy. Cardiac and liver mortality in that paper are one meta-analysis.

Absence searches

Counts are PubMed esearch on 2026-10-07. Cochrane CENTRAL and SciELO were not searched. Where a trial absence is claimed, that limit is named in the row or here. The Cochrane Database of Systematic Reviews was searched only as a PubMed journal field, and only for D-5 and D-6.

  • D-1 maintenance, count 2. ("type 2 diabetes"[tiab] OR "diabetes remission"[tiab]) AND remission[tiab] AND ("total diet replacement"[tiab] OR "formula diet"[tiab] OR "very-low-energy"[tiab] OR "very low energy"[tiab] OR "low-energy diet"[tiab] OR DiRECT[tiab]) AND (semaglutide[tiab] OR tirzepatide[tiab] OR liraglutide[tiab] OR "GLP-1"[tiab] OR "glucagon-like peptide-1"[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). Hits: Kempf 2025 (PMID 41113714), an indirect review of separate formula-diet trials against separate drug trials; Thom 2021 (PMID 33289256), a DiRECT ghrelin analysis. The withdrawal number in the row is Wilding 2022 (PMID 35441470), found by Wilding[Author] AND semaglutide AND (regain OR withdrawal OR extension) AND (STEP) (count 1). Population: obesity without diabetes.
  • D-2, count 20. ("SGLT2"[tiab] OR empagliflozin[tiab] OR dapagliflozin[tiab] OR canagliflozin[tiab] OR semaglutide[tiab] OR liraglutide[tiab]) AND ("legacy effect"[tiab] OR "delayed-start"[tiab] OR "delayed start"[tiab] OR "after discontinuation"[tiab] OR "treatment withdrawal"[tiab]) AND (kidney[tiab] OR cardiovascular[tiab] OR "heart failure"[tiab] OR mortality[tiab]) AND (diabetes[tiab]). All 20 titles were read. The only legacy analysis is Ceriello 2023 (PMID 37547276), observational: mean HbA1c above 8% over 3 years versus at or below 7% had later cardiovascular HR 1.33 (95% CI 1.063–1.365); starting an SGLT2 inhibitor in the 0–1 and 0–2 year windows removed the association (interaction p=0.006 and 0.003). The other 19 are reviews, case reports, protocols, post-hoc biomarker papers, or cost studies. None is a randomised withdrawal or delayed-start organ-outcome trial. EMPA-REG's abstract does not say the curves separated within weeks, so D-2 does not say "rapid".
  • D-3, count 12. ("GLP-1"[tiab] OR semaglutide[tiab] OR liraglutide[tiab] OR dulaglutide[tiab] OR tirzepatide[tiab]) AND ("SGLT2"[tiab] OR empagliflozin[tiab] OR dapagliflozin[tiab] OR canagliflozin[tiab]) AND (combination[tiab] OR combined[tiab] OR factorial[tiab]) AND (kidney[tiab] OR renal[tiab] OR "kidney failure"[tiab] OR eGFR[tiab]) AND ("Randomized Controlled Trial"[pt]) AND diabetes[tiab]. Read in full: Wexler 2026 PRECIDENTD feasibility (PMID 41456635; n=173; fills and discontinuation, not kidney failure) and Long 2026 (PMID 42170981; n=120; 24 weeks; combination reduced UACR versus either monotherapy or control, P below 0.05; no numeric UACR difference in the abstract; no kidney-failure or mortality endpoint). The other ten titles are imaging or biomarker analyses, a baseline-use subgroup, or dapagliflozin versus liraglutide. Neuen 2024 tests interaction with baseline SGLT2 use (1,743/17,072). Casper 2026 is a target-trial emulation. The residual gap is a hard kidney-failure or mortality trial of randomised dual therapy versus either class alone.
  • D-4, count 0. (adolescent[tiab] OR youth[tiab] OR paediatric[tiab] OR pediatric[tiab]) AND "type 2 diabetes"[tiab] AND (semaglutide[tiab] OR liraglutide[tiab] OR empagliflozin[tiab] OR dapagliflozin[tiab] OR "GLP-1"[tiab] OR "SGLT2"[tiab]) AND (retinopathy[tiab] OR "kidney failure"[tiab] OR "end-stage"[tiab] OR MACE[tiab] OR "cardiovascular events"[tiab] OR "myocardial infarction"[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). The TODAY paper used for Dot A is the complications follow-up (PMID 34320286), not a β-cell study. Separate TODAY β-cell PMIDs named in the wiki were not fetched and are not cited.
  • D-5, count 4; Cochrane journal field count 0. retinopathy[tiab] AND "type 2 diabetes"[tiab] AND ("early worsening"[tiab] OR titration[tiab] OR "rate of improvement"[tiab]) AND (semaglutide[tiab] OR tirzepatide[tiab] OR "GLP-1"[tiab] OR insulin[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). Titles, publication types, and Crossref were checked for PMID 42541561, 34234400, 29178519, and 17479440. Abstracts of those four were not re-read. Cochrane query: ("Cochrane Database Syst Rev"[jour]) AND retinopathy AND "type 2 diabetes" AND (titration OR "early worsening" OR semaglutide).
  • D-6, count 12; Cochrane journal field count 0. "type 2 diabetes"[tiab] AND (genotype-guided[tiab] OR "genotype guided"[tiab] OR "genetically guided"[tiab] OR pharmacogenetic[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). All 12 titles were listed. The only randomised drug-selection trial among them is Pearson 2003 (PMID 14575972), in HNF-1α diabetes: 5.2-fold greater fasting-glucose response to gliclazide than to metformin (4.7 versus 0.9 mmol/L). The other titles are association studies, reviews, and a non-randomised antiplatelet cohort (PMID 38106133; publication types include Preprint). Cochrane query: ("Cochrane Database Syst Rev"[jour]) AND "type 2 diabetes" AND (genotype OR pharmacogenetic) AND (treatment OR metformin).
  • D-8, count 2. (stigma[tiab] OR "person-first"[tiab] OR "stigmatizing language"[tiab] OR "stigmatising language"[tiab]) AND (clinician[tiab] OR physician[tiab] OR "primary care"[tiab]) AND diabetes[tiab] AND (HbA1c[tiab] OR adherence[tiab] OR attendance[tiab] OR "patient outcome"[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt] OR cluster[tiab]). Hits: Salazar 2023 (PMID 38545530), medication adherence in rural Indian primary care among people with common mental disorders, classified from the title; Hansen 2012 (PMID 22367667), help-seeking in low-income Latino adults, classified from the title. A broader stigma-plus-randomised query run earlier in this grounding session returned 55 and was not used as the absence count, because Joiner 2026 sits inside it and the query is too wide.
  • D-9, count 0. ("preference-concordant"[tiab] OR "preferred treatment"[tiab] OR "treatment preference"[tiab]) AND "type 2 diabetes"[tiab] AND (semaglutide[tiab] OR tirzepatide[tiab] OR "SGLT2"[tiab] OR "GLP-1"[tiab]) AND (adherence[tiab] OR persistence[tiab] OR discontinuation[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). Sun 2022 and Davies 2022 were not used: Sun has a Crossref erratum and is not a preference trial, and the Davies abstract contains no preference percentage.
  • D-10, count 2. ("bariatric surgery"[tiab] OR "metabolic surgery"[tiab] OR "gastric bypass"[tiab] OR "sleeve gastrectomy"[tiab]) AND (semaglutide[tiab] OR tirzepatide[tiab] OR liraglutide[tiab]) AND (versus[ti] OR compared[ti] OR comparison[ti] OR "head-to-head"[tiab]) AND ("Randomized Controlled Trial"[pt]). Both were read: Nigro 2025 (PMID 40779281; n=43; endoscopic sleeve versus liraglutide; 12-month weight and comorbidity differences no longer significant) and Stanley 2026 (PMID 42174253; semaglutide versus placebo after suboptimal surgical weight loss; n=70; week-68 adjusted weight difference −19.18 percentage points). A broader Randomized Controlled Trial[pt] query run earlier in this session returned 69, mostly GLP-1 physiology after surgery, and was not used as the absence count.
  • D-12, count 2. remission[tiab] AND "type 2 diabetes"[tiab] AND (surveillance[tiab] OR "screening interval"[tiab] OR deintensification[tiab] OR "follow-up interval"[tiab]) AND (retinopathy[tiab] OR kidney[tiab] OR microvascular[tiab] OR complication[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). Dholakia 2026 (PMID 41972326) is a metabolic-surgery review, title only. Tsujimoto 2021 (PMID 33549254) was read: 0/386 vision-threatening events at 4 years with no baseline retinopathy and mean on-treatment HbA1c below 7%, versus 6/721 (0.8%) with poorer control (P=0.54). That is not medication-free remission, and the interval was not randomised.
  • D-15, count 9. diabetes[tiab] AND (cascade[tiab] OR undiagnosed[tiab]) AND ("cluster randomised"[tiab] OR "cluster randomized"[tiab] OR "stepped-wedge"[tiab] OR "stepped wedge"[tiab]) AND (HbA1c[tiab] OR glycemic[tiab] OR glycaemic[tiab]). Read: Chan 2025 (PMID 40165254; detection uptake, not control after diagnosis), Yin 2024 (PMID 39494843; protocol), Simmons 2012 ADDITION-Cambridge (PMID 23040422; screening invitation and 10-year mortality HR 1.06, 95% CI 0.90–1.25). Peer 2020 (PMID 32470201) is the Cochrane review "Screening for type 2 diabetes mellitus", title and publication type only. The other records are gestational diabetes, type 1 education, or prevalence.
  • D-16, count 4. (adolescent[tiab] OR youth[tiab] OR pediatric[tiab] OR paediatric[tiab]) AND "type 2 diabetes"[tiab] AND (liraglutide[tiab] OR semaglutide[tiab] OR empagliflozin[tiab] OR dapagliflozin[tiab]) AND (withdrawal[tiab] OR washout[tiab] OR "C-peptide"[tiab] OR "beta-cell"[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). Dietsche 2024 (PMID 37967247) was classified from the title (a mechanistic metformin-plus-liraglutide trial). The other three titles are a heart-failure statement, a RISE predictor paper, and the RISE design paper. This row is the withdrawal question. D-4 is the organ-endpoint question. They are not the same claim.
  • D-17, count 0. remission[tiab] AND "type 2 diabetes"[tiab] AND (mediation[tiab] OR "causal mediation"[tiab]) AND (kidney[tiab] OR CKD[tiab] OR cardiovascular[tiab] OR mortality[tiab] OR CVD[tiab]). A looser query earlier in this session, which added trial names and the bare word "mediator", returned 73 hits dominated by unrelated uses of "remission". That count is not the absence claim.
  • D-18, count 2. ("type 2 diabetes"[tiab]) AND ("HbA1c target"[ti] OR "glycemic targets"[ti] OR "glycaemic targets"[ti] OR "intensive versus standard"[ti] OR "deintensification"[ti]) AND (semaglutide[tiab] OR tirzepatide[tiab] OR "SGLT2"[tiab] OR "GLP-1 receptor"[tiab]) AND ("Randomized Controlled Trial"[pt]). Rodriguez 2025 (PMID 39532398) and Lingvay 2023 (PMID 36514843) were classified from titles: insulin de-intensification by drug substitution, and tirzepatide target achievement. A broader "treat to target" query earlier in this session returned 30, and the listed titles are drug-versus-drug trials. That count is not an absence of records.
  • D-20, narrow count 0; broader count 4. Narrow: ("continuous glucose monitoring"[tiab] OR CGM[tiab]) AND "type 2 diabetes"[tiab] AND ("basal insulin"[tiab] OR "not taking insulin"[tiab] OR "without prandial"[tiab]) AND (retinopathy[tiab] OR nephropathy[tiab] OR mortality[tiab] OR "myocardial"[tiab] OR complication[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). Broader: ("continuous glucose monitoring"[ti] OR CGM[ti]) AND "type 2 diabetes"[tiab] AND (retinopathy[tiab] OR nephropathy[tiab] OR mortality[tiab] OR "cardiovascular events"[tiab]) AND ("Randomized Controlled Trial"[pt]). Ajjan 2023 (PMID 36516054) was read: about 3 months after myocardial infarction, intermittently scanned monitoring versus fingerstick testing, time in range +17 min/day (95% credible interval −105 to +153), combined emergencies and deaths 4 versus 7, HbA1c down 7 mmol/mol in both arms; the authors say longer-term benefit is untested. The other three (PMID 41208635, 40239470, 35107429) were classified from titles after publication-type and Crossref checks. Martens 2025 was not used for a between-group effect: both arms used monitoring.
  • D-21, count 1. (tirzepatide[tiab] OR semaglutide[tiab]) AND (sequence[tiab] OR sequencing[tiab] OR "treatment order"[tiab]) AND (cardiovascular[tiab] OR MACE[tiab] OR kidney[tiab]) AND "type 2 diabetes"[tiab] AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]). The hit is Joseph 2025 (PMID 40772215), a retrospective comparison, classified from the title.
  • D-22, count 12. "diabetes distress"[tiab] AND (screening[tiab] OR "stepped care"[tiab] OR "measurement-based"[tiab]) AND (randomized[tiab] OR randomised[tiab] OR "Randomized Controlled Trial"[pt]) AND (HbA1c[tiab] OR adherence[tiab] OR attendance[tiab]). Read: Pouwer 2011 (PMID 21221528; screening plus a letter in 223 outpatients already above a depression threshold; no between-group difference in depression, distress, or HbA1c) and Schmitt 2022 (PMID 34875666; stepped care once CES-D was at least 16 and/or PAID at least 40; depression response 80.2% versus 51.2%; PAID change not significant; HbA1c not different). The other titles are reviews, baseline reports, a type 1 monitoring trial, and protocols.

Unresolved

  • The wording of the DiRECT five-year erratum was not retrieved. D-1 therefore stops at the 12-month and 24-month trials.
  • Abstracts were not re-read for: the four D-5 neighbours (PMID 42541561, 34234400, 29178519, 17479440); Dholakia 2026 (PMID 41972326); Peer 2020 (PMID 32470201); Dietsche 2024 (PMID 37967247); Rodriguez 2025 (PMID 39532398); Lingvay 2023 (PMID 36514843); Salazar 2023 (PMID 38545530); Hansen 2012 (PMID 22367667); Joseph 2025 (PMID 40772215); and the three D-20 neighbours other than Ajjan (PMID 41208635, 40239470, 35107429). Each of those was checked for publication type and Crossref updated-by, and the row says they were classified from the title. If a full text of one of them is the missing junction, that row overstates the gap.
  • Cochrane CENTRAL and SciELO were not searched. An absence above is a PubMed absence, plus the two Cochrane-journal queries for D-5 and D-6.
  • No row was left uncited. No supporting paper was invented in order to keep a row.

2026-10-07 — Independent audit of Grok 4.7 dot grounding

  • Auditor: Codex, independent of the author, Grok 4.7. Scope: all 17 rows Grok kept and all five deletion decisions (22 original rows). OQ-1 through OQ-26 and the previously cited rows outside this table were outside this pass. No NCT identifier appears in the dot table.
  • Citation integrity: all 80 distinct PMIDs in the final 21-row table were fetched through live PubMed E-utilities in this session. Every cited record's PubMed publication types were checked for Retracted Publication and Expression of Concern, and every article DOI was checked through the live Crossref works API for updated-by: 80/80 resolved, zero retraction or concern flags, zero Crossref updates, zero missing or wrong-record citations. The two errata and one correction encountered in excluded candidate records were not treated as retractions or silently substituted for the cited articles.
  • Figures and anchoring: one figure needed relabelling: D-16's −1.06 and −1.30 percentage points are placebo-adjusted liraglutide differences, not within-arm HbA1c changes. D-17 now states that 284 of 586 SOS participants were in remission, avoiding a subgroup count that looked like the full cohort. D-11 quotes the produce-program activation count rather than repeating the source abstract's inconsistent percentage denominator. The audit found one motivated anchoring case: D-22 used trials restricted to people with elevated depression or distress scores to support a broad absence of a routine measurement-plus-response trial. A routine-care randomised trial in unselected adults already exists (Kelly 2025, PMID 37350136); it enrolled both diabetes types and did not show a statistically significant psychosocial between-group effect. D-22 now states the narrower unresolved question.
  • Absences: the exact 17 PubMed ESearch queries in Grok's entry above were rerun live on 2026-10-07; their counts matched the log, including the narrow D-22 count of 12. One table absence was stale (D-22), because its search formulation missed Kelly. Sixteen other trial or analysis absences were rewritten explicitly as findings of the named PubMed search, not claims that no study exists anywhere. Targeted supplementary searches through ClinicalTrials.gov v2 and Europe PMC were run for D-4, D-9, D-17 and D-20; these probes were not an exhaustive registry or conference search. Cochrane CENTRAL and SciELO remained unsearched.
  • Deleted rows re-evaluated: four of Grok's five deletion rationales reflected missing second-source searches, not an ungroundable junction. D-7 is restored with DiRECT mechanistic data (Taylor 2019, PMID 31866441) against a small South Asian/White European diet-response analysis (Athithan 2023, PMID 37811525). D-11 is restored with modelled diet-attributable incidence (O'Hearn 2023, PMID 37069363) against observed produce-prescription reach (Rader 2026, PMID 42036037). D-13 is restored with cardiovascular overlap (Driessen 2025, PMID 38147315) against FIB-4-to-elastography triage in newly diagnosed type 2 diabetes (Barana 2024, PMID 38347338). D-19 is restored with type 2 diabetes MASLD prevalence (Younossi 2024, PMID 38521116) against sequential-test accuracy in a mixed NAFLD population (Mózes 2022, PMID 34001645). Their populations and denominators are explicit. D-14 remains deleted: Dutta 2022 (PMID 35495849) supports DKA risk, but no second EHR hold–ketone–restart intervention result was established. No surviving row was deleted as a duplicate junction; the restored D-13 asks about cardiovascular use of triage, while D-19 asks about transporting test accuracy to type 2 diabetes.
  • Additional search provenance: PubMed ESearch ("type 2 diabetes"[tiab]) AND ("pancreatic fat"[tiab] OR "liver fat"[tiab]) AND (ethnic[tiab] OR ancestry[tiab]) AND (remission[tiab] OR threshold[tiab]) returned 1; type 2 diabetes food insecurity diet quality intervention reach randomized returned 3; type 2 diabetes cardiovascular risk FIB-4 elastography fibrosis screening returned 14; type 2 diabetes sequential FIB-4 transient elastography diagnostic accuracy advanced fibrosis returned 3. The broader D-22 query ("diabetes distress"[tiab]) AND (screening[tiab] OR "routine care"[tiab]) AND ("randomized controlled trial"[pt]) returned 14 and exposed the missed routine-care trial. These counts are source-specific retrieval results, not proof of exhaustive coverage.
  • Outcome and gate decision: four rows restored (D-7, D-11, D-13, D-19), three substantively repaired (D-16, D-17, D-22), 16 absence statements scoped to PubMed, zero retained rows deleted; 21 rows remain. Do not treat the table as definitive proof of novelty for the novelty gate. It is now a source-grounded candidate list, but its trial absences are primarily PubMed findings and CENTRAL/SciELO coverage is unresolved. The gate must run its own current multisource search before declaring any junction novel.

2026-10-08 — Note on the two audit entries below (operator error, and what it accidentally bought)

Two runs of tools/ground-dots.sh --review operated on this condition concurrently on 2026-10-08. The cause was operator error, not a tool defect: a throwaway invocation used to check the row dumper was backgrounded, and kill %1 did not reach the agent it had already spawned, so a second run started from the queue a minute later. Two codex reviewers and two claude auditors therefore worked on this file at the same time.

Nothing was corrupted — both reviewers independently deleted all 21 rows, so the writes converged — but that was luck rather than design, and tools/ground-dots.sh now takes a per-condition lock so it cannot happen again.

The two audit entries below are kept deliberately. They are independent audits of the same review by separate lanes, and they reached the same verdict: all 21 deletions upheld, no motivated anchoring, no difficulty passed off as a test failure. The second fetched one record the first did not (PMID 31866441, behind the deleted D-7). Unintentional replication is still replication, and it is better evidence for 21 → 0 than a single audit would be.

2026-10-08 — Structural review of type 2 diabetes dots

  • Outcome: started with 21 rows; kept 0; deleted 21. The table header and its five columns remain. No row was repaired because none passed all five tests. IDs were not reassigned.
  • Method: compared every junction with every OQ in its Powers column, then checked whether the two sides were separate quantified findings, whether the proposed study was a cross-literature test, and whether the link had already been studied. Ran a live PubMed ESearch for each row's junction and read PubMed EFetch records for the published comparisons below. The publication types of every newly cited record below had no Retracted Publication or Expression of Concern flag; each DOI resolved in Crossref with an empty updated-by field. Europe PMC searches for the thin ancestry, liver-screening, and distress candidates returned HTTP 503; no claim of exhaustive absence is made. CENTRAL and SciELO were not searched.
  • Test 1 — restates its own OQ (16 deletions): D-1 (post-diet pharmacologic maintenance is OQ-1's factorial maintenance trial); D-2 (drug withdrawal/delayed start is OQ-4); D-3 (combination hard-outcome trial is OQ-3); D-4 (youth organ-outcome trial is OQ-6); D-5 (retinopathy titration/monitoring trial is OQ-10); D-6 (cluster-guided drug assignment is the test-use strategy of OQ-9/OQ-11); D-8 (language training with patient outcomes is OQ-8); D-10 (decade-long surgery versus incretin comparison is OQ-5); D-12 (post-remission surveillance intervals are OQ-16); D-15 (population-denominator diagnosis-to-control package is OQ-19); D-16 (youth drug withdrawal and C-peptide durability are OQ-20); D-17 (mediation of remission versus weight loss is OQ-21); D-18 (modern default-target trial is OQ-22); D-19 (the remaining screening-to-outcome trial is OQ-23); D-20 (long CGM complication trial is OQ-24); D-21 (agent sequencing is OQ-25). D-12 and D-18 also use a consensus definition or target recommendation as one side, not a measured effect. D-15's sides are two descriptions of care gaps. D-20's UKPDS result is background to the CGM follow-on trial.
  • Test 3 — one side has no quantified finding (3 deletions): D-7 gives no numeric liver-fat or β-cell response for its DiRECT mechanistic side; a follow-up duration is not an effect size. D-9 counts preference studies and participants but gives no quantified preference result to join to the drug effect. D-13's cardiovascular-overlap review supplies no numeric association or effect estimate.
  • Test 5 — no cross-literature junction (1 deletion): D-11 juxtaposes modelled prevention incidence worldwide with produce-program uptake among people who already have diabetes. The populations and outcomes do not form a testable common link. The equitable policy and care-cascade work it suggests is already covered by OQ-14 and OQ-19.
  • Test 4 — broad junction already published (1 deletion): D-22 proposes measurement plus response in consultations, which a randomised mixed-type diabetes trial has already tested (PMID 37350136). Its smaller remaining question, an adequately powered type 2-only distress-response pathway, is OQ-26. This is a borderline deletion because the published trial does not provide a type 2-only effect; that limit is a reason to keep OQ-26, not to keep a dot that repeats it.
  • Published links narrowing other rows: D-3's combination-versus-monotherapy albuminuria test is published (PMID 42170981); hard kidney failure/death remains OQ-3. D-6's genotype/cluster scores have been evaluated against differential drug response inside a randomised crossover, without prospective assignment by a score (PMID 42312889). D-7's ancestry comparison itself measured similar liver-fat change but different insulin-resistance change (PMID 37811525); it did not estimate an ancestry-specific personal-fat threshold. D-10 has an observational surgery-versus-GLP-1RA comparison of mortality and vascular outcomes (PMID 40957960); a decade-long randomised comparison remains OQ-5. D-19's sequential FIB-4/elastography pathway has now been evaluated for advanced-fibrosis classification and later liver events specifically in type 2 diabetes (PMID 41911049); it does not establish that screening improves clinical outcomes.
  • Borderline keeps rejected: D-7 could have been a real ancestry-by-fat-threshold junction if its first side had supplied a numeric fat-response estimate; D-9 could have been a preference-by-drug-effect junction if its second side had supplied a numeric preference or persistence result. D-19 had two independent numeric sources, but its remaining clinical-benefit test is already the OQ and its diagnostic/prognostic link has been published. No unresolved issue changes the zero-row decision. Reconsider a dot only when a new, quantified second finding supports a junction beyond an existing OQ.

2026-10-08 — Independent audit of Codex structural review of type 2 diabetes dots

  • Auditor: Claude (Opus 5.5), independent of the author of the changes, Codex. Scope: the 2026-10-08 structural review above, which deleted all 21 rows of "Dots not yet connected" (D-1–D-13, D-15–D-22) and kept none. The pre-review table was reconstructed from logs/review-type-2-diabetes.md, which quotes every row in full with the OQ text it claimed to power. OQ-1 through OQ-26 were not edited. OPEN-QUESTIONS.md is unchanged by this audit.
  • Rows checked: 21 deletions; 0 kept rows (there are none to audit for restatement). No PMID or NCT ID remains in the dot table, so no table citation needed verification.
  • Citations behind the deletion reasons: the six PMIDs codex used to justify deletions were re-fetched live with PubMed EFetch on this date, with publication types and abstracts read. Each article DOI was checked at api.crossref.org/works/<doi> for updated-by. All 6 exist and are the papers the reasons imply. None has Retracted Publication or Expression of Concern, and every updated-by is empty. Characterisations:
  • TriMaster comparison (PMID 42312889; Diabetes Care 2026; RCT crossover, n=309). Clusters, partitioned polygenic scores and a routine-features model were tested against within-person differential HbA1c response to sitagliptin, canagliflozin and pioglitazone; only 2 of 8 PPSs showed modest association. Allocation by the routine-features model gave 0.27% (95% CI 0.18–0.35) greater HbA1c reduction against 0.16% (0.07–0.25) for cluster allocation. No arm was assigned prospectively by score, which matches codex's wording.
  • Surgery versus GLP-1 receptor agonists (PMID 40957960; Nat Med 2025; observational, 1,657 versus 2,275, median 5.9 years). Mortality adjusted HR 0.68 (0.48–0.96), MACE 0.65 (0.51–0.82), nephropathy 0.53, retinopathy 0.46. Codex accurately calls it observational.
  • Two-step FIB-4→VCTE pathway in type 2 diabetes (PMID 41911049; Gut 2026; n=4,781). Biopsy cohort n=352: 71% correctly classified. Prognosis cohort n=4,429: 5-year liver-related event incidence 0.7%, 0.9% and 11.8% across the low-, intermediate- and high-risk groups. This confirms that D-19's junction (sequential-test performance specifically in type 2 diabetes) has been published. It does not test whether screening changes outcomes.
  • Kelly 2025 (PMID 37350136), Long 2026 (PMID 42170981) and Athithan 2023 (PMID 37811525): each matches its use in the deletion reasons.
  • Deletion decisions: I agree with all 21. Each reason was checked against the row text and the OQ text.
  • Test 1 (16 rows): every junction is the trial already proposed or declared absent in its OQ, often with the same citations. Examples: D-1 matches OQ-1's factorial pharmacologic-maintenance trial; D-2 is OQ-4's "withdrawal or delayed-start extensions"; D-17 is OQ-21's "mediation analyses … are needed"; D-20 is OQ-24's complication-endpoint trial.
  • Test 3 (D-7, D-9, D-13, plus D-12/D-18 as a secondary reason): confirmed from the row text. D-7's Dot A has no number. In D-9, Dot B reports counts of studies, not a preference result. D-13's Dot A is a narrative review with no effect size. D-12's Dot A and D-18's Dot A are a definition and a guideline target respectively.
  • D-7 restatement: D-7 could have been repaired with a numeric liver-fat change from Taylor 2019. It would still restate OQ-12, which already asks for a fat threshold "across ancestries". Athithan alone already reports both the similar liver-fat change and the divergent HOMA-IR change, so the row is one paper, not a join of two.
  • D-11: its two sides cover different populations (modelled worldwide incidence versus produce-prescription reach among people already diagnosed), so there is no testable common link. Its residue is OQ-14.
  • D-19 and D-22 (Test 4): both "already published" reasons name the paper, and both papers verify.
  • Difficulty passed off as a test failure: none found. Codex did no wholesale "difficulty" deletions.
  • Motivated anchoring: 0 cases. No citation was attached to save a row, and the six deletion-supporting citations say what codex says they say.
  • Absences re-run:
  • D-17: Grok's mediation query (remission AND type 2 diabetes AND mediation AND kidney/CV/mortality) returned 0 on PubMed ESearch today, the same as 2026-10-07. A broader query (remission AND diabetes AND weight-loss AND mediat*/"independent of weight" AND CV/kidney/mortality/microvascular) returned 10, and the 10 titles were read: no mediation analysis of remission versus weight loss.
  • D-6: a 2024–2026 RCT query for precision, cluster or polygenic allocation in type 2 diabetes returned 13; the 13 titles were read. Only TriMaster (PMID 42312889) evaluates allocation strategies, and it does so retrospectively inside a crossover. No prospective score-assigned trial was found.
  • Result: 0 stale absences found in codex's claims. Both searches are PubMed-only. Europe PMC, CENTRAL and SciELO were not searched in this audit.
  • Follow-ups outside this pass (not edited, since this audit may write only the dot table and LOG):
  • SURPASS-PEDS, tirzepatide in youth type 2 diabetes (PMID 40975112), surfaced in the D-6 search. It was title-only, and its retraction/Crossref check was not run. OQ-6 and OQ-20 still name only liraglutide and empagliflozin as youth efficacy evidence; the next sweep should check and add it.
  • The surgery-versus-GLP-1RA cohort (PMID 40957960) and the two-step MASLD pathway (PMID 41911049) are not yet in OQ-5 or OQ-23. Both OQs should cite them so that their stated absences stay current.
  • Process note: the Crossref calls in this audit sent the user's contact email in the User-Agent header (Crossref "polite pool" convention). That was not authorised, and future runs should omit it.
  • Gate judgement: the empty table can be trusted by the novelty gate in the sense that it now asserts no junction. There is nothing in it for the gate to mistake for novelty. The deletion was correct: the earlier 21 rows were largely the OQs restated as dot pairs. This condition currently contributes zero candidate dots. A new row should be added only when a quantified finding from a second literature supports a join that no OQ already proposes.

2026-10-08 — Independent audit of Codex structural review of dots

  • Auditor: Claude (Opus 5.5), independent of the author, Codex, whose 2026-10-08 structural review above deleted all 21 rows. Scope: the 21 deletion decisions. No rows were kept, so no cited row is left to re-fetch for figures, drift or anchoring. OQ-1 through OQ-26 were not edited. The pre-deletion table was read from the review dump logs/review-type-2-diabetes.md, which tools/ground-dots.sh wrote before the review began. I did not reconstruct it from memory.
  • Citations in the deletion rationale: I fetched the six PMIDs Codex cited for the first time, live via PubMed EFetch: 37350136, 42170981, 42312889, 37811525, 40957960 and 41911049. I also fetched 31866441, which the deleted D-7 relied on. All 7 exist. None is typed Retracted Publication or Expression of Concern, none has a PubMed RetractionIn, ErratumIn or EOC link, and every article DOI has an empty Crossref updated-by. Zero citations failed verification. Each one says what Codex claims:
  • Kelly 2025 (PMID 37350136) randomised 98 adults with type 1 or type 2 diabetes to Spotlight-AQ measurement plus consultation response. HbA1c fell 6 versus 3 mmol/mol. The psychosocial improvements were not statistically significant.
  • Long 2026 (PMID 42170981) found combination UACR reduction greater than either monotherapy (P<0.05). There was no hard endpoint.
  • Güdemann 2026 (PMID 42312889) is the TriMaster crossover (n=309). It tested clusters, partitioned polygenic scores and a routine-features model against differential response. Allocation by the routine-features model beat cluster allocation: 0.27% (95% CI 0.18–0.35) versus 0.16% (0.07–0.25) HbA1c. Nothing was assigned prospectively.
  • Athithan 2023 (PMID 37811525) found liver-fat reduction similar by ancestry. HOMA-IR fell less in South Asians: −5.7 versus −8.6.
  • Gasoyan 2025 (PMID 40957960) is observational. Surgery versus GLP-1RA: mortality aHR 0.68 (0.48–0.96), MACE 0.65 (0.51–0.82), nephropathy 0.53 (0.43–0.67), retinopathy 0.46 (0.29–0.75).
  • Chen 2026 (PMID 41911049) evaluated FIB-4 followed by liver stiffness in type 2 diabetes. In the biopsy cohort (n=352) 71% were correctly classified. In the outcome cohort (n=4,429), 5-year liver-event incidence was 0.7%, 0.9% and 11.8% across the low-, intermediate- and high-risk groups.
  • Attribution error found in the deleted text: PMID 31866441, which the 2026-10-07 Codex audit restored into D-7 as "Taylor 2019", is Al-Mrabeh 2020 (Cell Metab). Its abstract has no numeric liver-fat or β-cell figure, which confirms the Test 3 deletion. The wiki and bibliography already credit it correctly, so the error lived only in the deleted row and that earlier log entry.
  • Deletion decisions, each checked against the pre-deletion text and the OQ text:
  • Test 1, 16 rows. For D-1, D-2, D-3, D-4, D-5, D-6, D-8, D-10, D-12, D-15, D-16, D-17, D-18 and D-20, the row's junction is the design its OQ already calls for. OQ-1 names a factorial maintenance trial including pharmacologic maintenance. OQ-4 names withdrawal or delayed-start extensions. OQ-21 says "mediation analyses … are needed". The others match the same way. I confirm all 14.
  • D-19 (filed by Codex under Test 1). I would file it under Test 4 instead. Its junction was whether sequential-test accuracy transfers to type 2 diabetes. OQ-23 asks for a randomised outcome trial, which is a different question. Chen 2026 (PMID 41911049) has published the type 2 diabetes-specific evaluation, so the deletion stands on that ground.
  • D-21 (filed by Codex under Test 1). OQ-25 calls for decision models, not a sequencing trial. Both sides were CVOTs from a single literature, so Test 5 is the better fit. The deletion stands.
  • Test 3: D-7, D-9 and D-13 confirmed. D-7's Dot A has no number, as above. D-9's Dot B states no preference percentage; I confirmed this against the row text. D-13's Dot A is a narrative review with no effect estimate. D-7 and D-9 would also have failed Test 1, against OQ-12 and OQ-18 respectively.
  • Test 5: D-11 confirmed. Its two sides were modelled worldwide prevention incidence and produce-programme reach among people who already have diabetes. They share no population and no outcome.
  • Test 4: D-22 confirmed against Kelly 2025. What remains is the type 2-only distress-response trial, which is OQ-26.
  • Overall: none of the 21 deletions was a deletion out of difficulty, and I restored no rows.
  • Absences re-run (PubMed ESearch, 2026-10-08):
  • D-17 query from the 2026-10-07 log, rerun as written: count 0, matching that log. A looser query, "type 2 diabetes"[tiab] AND remission[tiab] AND mediat*[tiab] AND "weight loss"[tiab] AND outcome*[tiab], returned 10. I read all 10 titles: reviews, a clusterin association study, a triglyceride-mediates-remission analysis (PMID 38436759, which mediates remission itself rather than remission's effect on outcomes), a Mendelian randomisation study and animal work. None splits remission's outcome association into weight-loss and remission-specific paths, so OQ-21's absence still holds on PubMed.
  • "type 2 diabetes"[tiab] AND "FIB-4"[tiab] AND (elastography[tiab] OR "liver stiffness"[tiab]) AND (two-step[tiab] OR sequential[tiab]) returned 15, including Chen 2026 and Mózes 2022 (PMID 34001645). This confirms the D-19 publication finding.
  • Stale absences: 0. Europe PMC, CENTRAL and SciELO were not searched in this audit.
  • Rows deleted or repaired by this audit: none. Motivated-anchoring cases: 0. No row survives in which an anchor could sit. The one mislabel above was in deleted text.
  • Follow-ups (not done here; out of this pass's write scope):
  • Catalogue Chen 2026 (PMID 41911049) against OQ-17/OQ-23. OQ-17's call for prospective validation of sequential strategies is now partly answered for advanced-fibrosis classification and liver events, though not for clinical benefit.
  • Catalogue Gasoyan 2025 (PMID 40957960) against OQ-5.
  • Catalogue Güdemann 2026 (PMID 42312889) against OQ-9/OQ-11.
  • Catalogue Kelly 2025 (PMID 37350136) against OQ-26.
  • None of these six PMIDs is yet in literature/BIBLIOGRAPHY.md.
  • Gate judgement: the empty table can be trusted. Zero dots is the correct state for this condition now: every candidate junction either restates an OQ, has an unquantified side, or has been published. The novelty gate has nothing to assess for type 2 diabetes, and this condition's research agenda lives in OQ-1 to OQ-26. Add a dot only when a new, quantified, independently sourced finding joins two literatures in a way no OQ already describes.

2026-10-09 — Periodic literature sweep since 2026-08-28

  • Searches actually run: PubMed E-utilities ESearch across 14 major topics (epidemiology/burden, pathophysiology/precision, prevention, remission/weight, incretins, SGLT2/cardiorenal, glycaemic technology, surgery, complications/MASLD, youth, biomarkers, guidelines, patient experience, safety), each restricted to publication dates 2026-08-28 through 2026-10-09. Each topic also had a dated "Cochrane Database Syst Rev"[jour] AND <topic> PubMed query; all 14 returned zero in this window. Europe PMC REST was queried for all 14 topics and seven focused non-MEDLINE title searches. Exact strings, PubMed translations, counts and returned IDs are in the legal sweep manifest. PubMed EFetch, Europe PMC full-text indicators and Crossref updated-by checked every added journal record. The 12 previously cited NCT registrations and new SURPASS-PEDS registration NCT05260021 were queried through the ClinicalTrials.gov v2 API. The named European registry was queried for SURPASS-PEDS under EudraCT 2021-003612-31. Publisher or society pages checked: ADA/EASD, EASD, NICE NG28, KDIGO diabetes/CKD, IDF Atlas and Diabetes UK. No direct WHO ICTRP search was completed. Cochrane Library/CENTRAL and SciELO returned HTTP 403; the LILACS search endpoint timed out. Consequently, no worldwide or regional-literature absence is asserted.
  • Highest-priority new documents: ADA/EASD issued its 2026 T2D consensus, superseding the 2022 report (Davies 2026, PMID 42825450). EASD issued a diabetes-distress guideline (Speight 2026, PMID 42744915), supported by a systematic review of RCTs (Schipper 2026, PMID 42726276). Their full texts were read. The EASD T2D estimates for psychological and educational interventions were added to the statistics layer with low/very-low certainty labels. No new Cochrane review surfaced in the 14 date-window PubMed queries. Most other returned papers were not promoted.
  • Trial and follow-up triage: NCT05536804 is now COMPLETED (registry update posted 2026-09-11) with no posted results; NCT06383390 has a 2026-10-08 update but remains active, not recruiting, without results. The other tracked registration statuses matched the wiki snapshot. The earlier audit's SURPASS-PEDS follow-up was catalogued (Hannon 2025, PMID 40975112; NCT05260021; EudraCT 2021-003612-31). PubMed and ClinicalTrials.gov support its short-term glycaemic trial design; Crossref records an erratum, not a retraction. Its numeric treatment effect was held back pending review of the correction. Chen 2026 (PMID 41911049), Gasoyan 2025 (PMID 40957960), Güdemann 2026 (PMID 42312889), and Kelly 2025 (PMID 37350136) were re-resolved and used only to sharpen existing question boundaries. Full publisher texts for the first three were not accessible, so no new detailed effect estimates from them were added. A new remission-outcomes meta-analysis (Ruan 2026, PMID 42843768) remained abstract-only and was not promoted. All eight bibliography additions lacked PubMed retraction or expression-of-concern publication types; Crossref updated-by was empty except for the SURPASS-PEDS erratum.
  • Non-journal checks: NICE NG28 still listed its February 2026 update; KDIGO still identified 2022 as current and 2026 as a draft in revision. IDF's 11th Atlas estimates, published April 2025 and missed by the old statistics table, were retrieved live and added as all-diabetes, age-20–79 estimates, not mislabelled T2D or new 2026 data. Diabetes UK's 2026-09-24 account of the EASD distress guideline was added to patient-voice sources. The Europe PMC non-MEDLINE sweep surfaced preprints and other records but none met this sweep's threshold for citation; no no-PMID DOI record was added.
  • Files changed: four wiki pages (guidelines, patient-experience-and-advocacy, youth-onset-t2d, clinical-trials-landscape) were dated 2026-10-09 and demoted to draft; the bibliography gained eight records; the guideline registry, statistics, patient-voice sources, open questions and index were reconciled; the query manifest was saved at its canonical path. No deep paper note was warranted. Publishing was left to the external pipeline as instructed.
  • Next sweep: read the SURPASS-PEDS erratum and corrected full text before quoting its effect size; watch NCT05536804 for posted results and NCT06383390 for a substantive status/results change; evaluate any randomised hard-outcome study of GLP-1/SGLT2 combination, a prospective cluster-guided allocation trial, and a randomised MASLD screening-to-outcomes pathway. Recheck for ADA/EASD implementation updates, KDIGO final publication, and a T2D-only distress assessment-to-response trial. Seek access to CENTRAL, SciELO and LILACS before making stronger trial or regional transportability absence claims.

2026-10-23 — Periodic literature sweep since 2026-08-28

  • Author: Claude (Opus 5.5). It ran the sweep alone, so an independent audit by a different engine is still required.
  • Searches actually run:
  • PubMed: E-utilities ESearch for the same 14 topics as 2026-10-09, over the window 2026-08-28 to 2026-10-23 by publication date. The incretin query was widened to orforglipron, retatrutide and cagrilintide. Four cross-cutting filters were added: RCT/meta-analysis/systematic-review publication types; eight high-impact journals; practice-guideline publication types; and "Cochrane Database Syst Rev"[jour] with any diabetes term. A dedicated OQ-24 absence query was also run. retmax was 1,000, against 35 last time, so earlier truncation no longer hides records.
  • Cochrane: all 14 topic-specific Cochrane queries returned 0. The broad diabetes query returned one record, a protocol rather than a review: "Blood pressure targets for hypertension in people with diabetes mellitus" (Larvie 2026, PMID 42788405). No Cochrane review published in this window supersedes or contradicts anything this condition asserts.
  • Europe PMC: REST searches for all 14 topics, plus six focused non-MEDLINE searches (NOT SRC:MED). Their hits were preprints, reviews and PMC-only reviews; none met the citation threshold, and no DOI-only record was added.
  • ClinicalTrials.gov: the v2 API was queried for all 13 tracked NCT IDs and for nine new registrations named by the new papers.
  • Retraction checks: PubMed EFetch publication types and CommentsCorrections, Crossref updated-by (no mailto, no email in headers) and Europe PMC full-text flags, for every added record.
  • Web pages: NICE NG28, the KDIGO diabetes-CKD page, Diabetes UK news, IDF news, the WHO fact sheet and the ADA Standards supplement page.
  • Not searched: Unpaywall was skipped because LIFEBANK_CONTACT_EMAIL is unset. Cochrane Library/CENTRAL and SciELO/LILACS were not queried (HTTP 403 per SOURCES.md). WHO ICTRP and the named non-US registries were not queried this sweep.
  • Manifest: exact query strings, translations, counts, IDs and per-record checks are in the sweep manifest.
  • What was found and promoted (16 bibliography records):
  • ACHIEVE-4 (Klein 2026, PMID 42815506; NCT05803421). This is the first cardiovascular outcome trial of an oral non-peptide GLP-1RA. Orforglipron was noninferior to insulin glargine for MACE-4 (HR 0.84, 95% CI 0.59–1.20). It is an active-comparator safety trial, not evidence of benefit. PubMed and Crossref both link an erratum (PMID 42880309) whose text was blocked (HTTP 403) and was not read.
  • CONNECT (Oser 2026, PMID 42878622; NCT06601166). CGM versus routine care in non-insulin-treated T2D in primary care: HbA1c −0.9 points at 26 weeks.
  • TRIUMPH-2 retatrutide (PMID 42810372) and SYNCHRONIZE-2 survodutide (PMID 42820639). Both are weight-primary trials in T2D.
  • SELECT/FLOW/SOUL pooled kidney analysis (Mann 2026, PMID 42567173).
  • Scandinavian SGLT2-inhibitor ketoacidosis cohort (Kadesjö 2026, PMID 42679839): 2.43 per 1,000 person-years, with risk that did not stop after the early treatment period; 25.1% were back on an SGLT2 inhibitor one year after an event.
  • BMJ target-trial emulation (Turchin 2026, PMID 42778221): kidney benefit with albuminuria, none detected without it.
  • ARMMS-T2D remission and relapse predictors (Simonson 2026, PMID 42900776).
  • US laboratory-confirmed diabetes care gaps (Stimpson 2026, PMID 42832327).
  • National DPP natural experiment (Chorniy 2026, PMID 42227951).
  • QBSafe cluster RCT, null on its primary outcomes (Lipska 2026, PMID 42895748).
  • COMBINE 4 IcoSema (PMID 42594928).
  • Disease-modification core outcome set (Khunti 2026, PMID 42810354).
  • Two guideline-layer records: the Asia-Pacific GRADE guideline on intensification beyond GLP-1RA (Luo 2026, PMID 42543082) and the EJHF CKM heart-failure prevention consensus (Inciardi 2026, PMID 42817848).
  • The Cochrane protocol above, listed on the registry watch list only.
  • Retraction status: none of the added records carries a Retracted Publication or Expression of Concern type. Crossref updated-by was empty for all of them except ACHIEVE-4, which has an erratum, not a retraction.
  • Full-text status: only Chorniy (PMC13271827) and Lipska (PMC13681207) had full text in Europe PMC. Every other figure written today is abstract-reported, and the pages say so. Under CLAUDE.md rule 5, no gap claim rests on these abstracts.
  • Triaged out (not added):
  • Safiglipron OUTSTAND-1/-2 (PMIDs 42811187, 42834229): China-only glycaemic phase 3 trials.
  • REMODEL semaglutide kidney mechanism (PMID 42823485): co-primary MRI outcomes were null.
  • twiist AID pivotal trial (PMID 42844027): single-arm, 13 weeks.
  • A JAMA news item on a second weekly basal insulin (PMID 42886478): no abstract, and the agent was not verifiable.
  • The rest of roughly 2,000 topic hits were mechanistic, cross-sectional, regional-descriptive or commentary.
  • None warranted a deep note. ACHIEVE-4 is the nearest candidate, but it waits until its erratum can be read.
  • Trials: all 12 tracked registrations and SURPASS-PEDS were unchanged since 2026-10-09 (status, last-update date, results flag). TREASURE-CKD (NCT05536804) still has no posted results. TRIUMPH-Outcomes (NCT06383390) is still active, not recruiting.
  • Guidelines and statistics:
  • NICE NG28 is unchanged (18 February 2026).
  • The KDIGO page returned HTTP 406 and could not be rechecked; its 2026 draft status is carried forward from 2026-10-09 unverified.
  • The ADA Standards page returned 403.
  • No new IDF, WHO or Diabetes UK figure or campaign merited the patient-voice layer; Diabetes UK's distress item was already catalogued.
  • Files changed:
  • Wiki: incretin-therapies, glycemic-management-and-legacy-trials, sglt2-inhibitors-and-kidney-heart and metabolic-surgery were demoted curated → draft. clinical-trials-landscape was already draft. All five are dated 2026-10-23.
  • Literature layer: BIBLIOGRAPHY.md gained 16 records; statistics/STATISTICS.md gained seven rows; guidelines/REGISTRY.md gained two records plus watch-list items.
  • Open questions: OPEN-QUESTIONS.md sharpened OQ-4 (core outcome set), OQ-7 (DKA persistence and post-event discontinuation), OQ-14 (National DPP effect and reach), OQ-18 (QBSafe null), OQ-19 (US post-test cascade leak), OQ-24 (CONNECT; PubMed-only complication-trial absence re-run, 0 records) and OQ-25 (ACHIEVE-4 as a third comparator type). No question was answered. The dots table was not touched.
  • Also updated: INDEX.md and the manifest.
  • guidelines.md was not edited. The two new guideline records should be synthesised there at the next audit.
  • Pre-existing issue noticed, not fixed: 36 PMIDs cited in the OPEN-QUESTIONS dots table are absent from BIBLIOGRAPHY.md (for example 37811525 and 42170981). They pre-date this sweep, and cataloguing them is out of a sweep's scope. The auditor or a deepening pass should reconcile them.
  • Next sweep should chase:
  • Read the ACHIEVE-4 erratum (PMID 42880309) before quoting its figures as final.
  • Watch for full texts of CONNECT, TRIUMPH-2 and SYNCHRONIZE-2, and for results postings on NCT05803421, NCT05929079, NCT06066528 and NCT05536804.
  • Watch for the completed Cochrane BP-target review (protocol PMID 42788405).
  • Track KDIGO 2026 finalisation, using a non-scripted route if 406 persists.
  • Watch the ADA Standards 2027 release (expected December).
  • Look for a randomised SGLT2 sick-day implementation trial (OQ-7) and any orforglipron or oral GLP-1RA placebo-controlled outcome trial.
  • Query WHO ICTRP and the named registries, which this sweep skipped.

2026-10-30 — Periodic literature sweep since 2026-08-28

  • Author: Codex. This is curation, not the required independent audit; edited wiki pages remain draft.
  • Searches actually run: 14 major-topic PubMed E-utilities ESearch queries over publication dates 2026-08-28–2026-10-30, 14 matching Cochrane-journal queries, plus broad diabetes/CDSR, RCT/synthesis, high-impact-journal, practice-guideline and CGM-complication queries. All query strings, translations, counts and returned IDs are in the manifest. The 1,881 unique search IDs yielded 1,876 journal records and five book records; journal titles were screened, prioritised abstracts read, and 249 records were outside the prior search-ID set and bibliography. No truncation at the 10,000-record ESearch limit occurred. Europe PMC REST was queried for all 14 topics and six focused non-MEDLINE searches; the first 100 results per query were inspected, not the entire result universe.
  • Cochrane: all 14 T2D-topic queries returned zero. The broader diabetes query returned the existing BP-target protocol (PMID 42788405) and a new maternity-versus-non-maternity gestational-diabetes screening review (PMID 42954428), outside this condition’s scope. Neither is a completed T2D BP-target review; no contradiction to current T2D synthesis was identified in these PubMed results. CENTRAL was not searched.
  • Seven records survived journal triage: the ADA/EASD parallel Diabetologia correction (Davies 2026, PMID 42955083); Malaysian DSPN expert consensus (Lim 2026, PMID 42928304); EASL/EASD/EASO focused MASH guidance (PMID 42805503); J-DOIT3 achieved-HbA1c analysis (Noda 2026, PMID 42912779); older-adult SGLT2 safety cohort (Harbi 2026, PMID 42937517); surgery-relapse synthesis (Nie 2026, PMID 42945021); and DIADEM-I qualitative substudy (Zaghloul 2026, PMID 42944694). The last three received bibliography lines only. No landmark deep note was warranted; most new records were not added. No DOI-only record was cited.
  • Verification and corrections: all seven additions were fetched live through PubMed EFetch and checked for retraction/expression-of-concern publication types and correction links, and against Crossref updated-by; none was retracted or had an expression of concern. The Davies record is itself an erratum. Its publisher notice was read in full and corrects the ESSENCE interim-analysis sample size from 320 to 800, without changing conclusions. The Malaysian consensus and J-DOIT3 full text were read, including J-DOIT3 Tables 1–4. The latter is an achieved-exposure analysis of pooled original arms, not a new target-strategy randomisation. Detailed EASL guidance synthesis was deferred: publisher body text surfaced in search, but direct opening returned 403. Europe PMC’s full-text flags were checked for all additions; its “not in EPMC” does not establish that publisher full text is unavailable. The previous ACHIEVE-4 and SURPASS-PEDS correction notices were re-fetched through PubMed, but their publisher content remained inaccessible; no corrected effect estimate is claimed.
  • Trials beyond PubMed: all 19 NCT IDs already cited anywhere in this condition were retrieved through the ClinicalTrials.gov v2 API. Status, last-update date, results flag and first-results-posting date matched the 2026-10-23 manifest for every record, including the six newer registrations beyond the original tracked set. WHO ICTRP returned substantive search results for type 2 diabetes, including non-US registrations; its first page was inspected, not all trials. ISRCTN’s remission search redirected to a cookie/JavaScript holding page. Exact SURPASS-PEDS and remission searches at the EU register returned request-rejected pages, despite HTTP 200. No worldwide trial absence is asserted from these partial checks.
  • Society/statistics/patient sources: read the official KDIGO, ADA Standards, NICE NG28, IDF facts/news, WHO fact-sheet and Diabetes UK pages. KDIGO 2022 remains current; its 2026 draft is being revised after public review, resolving last week’s inaccessible-page follow-up. ADA’s page still links 2026 Standards and NG28 still lists 18 February 2026. IDF and WHO retain the already catalogued prevalence years, so no new burden number was invented. Two Diabetes UK sources were added: October insulin-withdrawal transition support and the 22 October advocacy response on unequal weight-management access; neither is treated as an effectiveness study or burden estimate.
  • Files changed: bibliography (278 distinct records; the previous index’s 270 count was one below the pre-sweep file’s 271), guideline registry (21 document rows), statistics (one fully checked achieved-HbA1c association with observational caveat), patient-voice sources, OPEN-QUESTIONS.md (OQ-22 sharpened; regional coverage limit attached to OQ-12), INDEX.md, and three wiki pages (guidelines, glycemic-management-and-legacy-trials, clinical-trials-landscape). Changed pages were dated 2026-10-30 and remain draft. No other condition, tool or publishing artifact was changed.
  • Coverage limits: Cochrane Library/CENTRAL and SciELO/LILACS could not be searched (HTTP 403 limitations documented in SOURCES.md; not re-probed). This matters to cross-region transportability and is now explicit in the open questions. Europe PMC sampling and inaccessible non-US registry results limit absence inference. Unpaywall was skipped because LIFEBANK_CONTACT_EMAIL is unset; Crossref was queried without a personal email. A transient Crossref 429 for the consensus correction was retried successfully before citation.
  • Validation: passed checks for all changed-file local links, wiki headings/reference coverage/date/draft status, every PMID on added or amended lines against this session’s live EFetch records, all seven additions’ dual retraction checks, 19 registry comparisons, 278 distinct bibliography IDs and 21 guideline rows. No full-site generation, publishing or git command was run.
  • Next sweep: obtain the ACHIEVE-4 and SURPASS-PEDS correction contents; retrieve the focused EASL MASH guidance directly before detailed recommendation extraction; watch the completed registrations for first results, KDIGO finalisation and ADA 2027 release; reassess whether a prospective modern target-strategy trial addresses OQ-22; obtain full texts of the older-adult safety cohort, surgery-relapse synthesis and DIADEM-I qualitative report before numerical or thematic promotion. Repeat WHO ICTRP with focused intervention terms and seek substantive ISRCTN/EU registry access. Independent audit remains pending; publishing is handled externally.