Skip to content

Curation log — chronic kidney disease

Newest entries first. Every content-changing session appends an entry: date, what changed, what was searched, follow-ups for next time. See CLAUDE.md.

2026-09-02 — Full build (Codex)

Built all 24 canonical wiki pages at status: draft (3,994 lines total), with 25–60 live-resolved PubMed records per page reference set. Built the literature layer: master bibliography (125 unique records), five landmark notes, 12-document guideline registry, quantitative statistics tables, and four-file patient-voice layer. Replaced the seed open questions with 20 stable tiered questions and a 12-row “Dots not yet connected” table. Rewrote the index with counts, reading paths and page statuses.

Live searches run. PubMed E-utilities searches and batch record fetches covered: KDIGO 2024; CKD definition/staging; eGFR and albuminuria prognosis; global burden and awareness; race-free eGFR; hyperfiltration/fibrosis/AKI transition; KFRE; RENAAL/IDNT/ONTARGET/VA NEPHRON-D; DAPA-CKD/EMPA-KIDNEY/CREDENCE; FIDELIO/FIGARO/FIDELITY; FLOW; CHOIR/CREATE/TREAT/ASCEND-D; EVOLVE; SHARP/4D/AURORA; MDRD and nutrition reviews; AMBER and potassium binders; VALOR-CKD; IDEAL; conservative care; symptom instruments and SONG; ADPKD TEMPO/REPRISE; IgA NefIgArd/PROTECT; FSGS DUPLEX; genomic diagnosis; contrast safety; and current guideline documents. Anchor PMIDs were re-fetched rather than inherited.

ClinicalTrials.gov v2 searches. Queried active CKD studies and exact trial terms/records. Catalogued 17 live-resolved NCT identifiers and dated statuses. Plausible identifiers that resolved to unrelated cancer, dementia or other trials were explicitly discarded and do not appear in content.

Web sources checked. Official KDIGO, NICE, UK Kidney Association and Kidney Health Australia guideline pages; National Kidney Foundation, Kidney Care UK, National Kidney Federation, Kidney Foundation of Canada, Kidney Health Australia, National Kidney Foundation of South Africa, Kidney Failure Awareness & Advocacy SA, Kidney Federation of India, Hong Kong Kidney Foundation and World Kidney Day public pages.

Scope decisions applied. Dialysis and transplantation are restricted to timing, choice and outcomes; no technique, access engineering or immunosuppression protocol content. Diabetic CKD owns kidney endpoints only. Hypertension owns BP measurement and target synthesis. AKI appears only as a progression border.

Follow-up for independent audit. Re-fetch every cited PMID and NCT identifier; verify each quantitative claim against abstract/full text; check ePub-versus-print citation details; review national-guideline completeness outside English; re-run asserted evidence gaps; and promote only pages that pass. No page was self-promoted.

2026-09-02 — Seeded (Claude)

Created the scaffold: INDEX.md with a 24-page plan (22 topical + 2 standing house pages) and a 15-record anchor table, a seed OPEN-QUESTIONS.md, this log, and the empty wiki/ and literature/ layers. No wiki pages written. Excluded from the public site until it has pages.

Searches run (live PubMed E-utilities, 2026-09-02). Scoping: chronic kidney disease 244,964; chronic kidney disease[MeSH Major Topic] 117,221; dialysis 240,604; kidney transplantation 167,381; diabetic kidney disease OR diabetic nephropathy 58,971. Per-page topic counts recorded in INDEX.md, including cardiovascular outcomes 19,834, AKI 16,931, anaemia 13,648, glomerulonephritis 11,475, CKD-MBD 6,280, epidemiology 4,379, progression/risk prediction 3,244, polycystic 3,261, SGLT2 2,325, finerenone/MRA 1,855, eGFR/albuminuria classification 980, symptom burden 558, conservative care 473, race and eGFR 329, diet/potassium 343.

Query-hygiene note. Three initial topic queries were malformed — unparenthesised OR binds loosely in E-utilities, so chronic kidney disease AND finerenone OR mineralocorticoid antagonist returned 15,304 (i.e. every MRA paper) rather than 1,855. All affected counts were re-run parenthesised and are marked † in INDEX.md. Any future scoping pass must parenthesise Boolean groups; an unchecked count would have justified a page the literature does not support.

Anchor records resolved live: 15, listed in INDEX.md.

Scoping decisions. Third-largest literature in the repository, so scoped by border like hypertension and melanoma. Three exclusions: the dialysis-technique and transplant-immunology corpora each get one decision-and-outcomes page rather than a section of the condition (each of those literatures individually exceeds this condition's MeSH-major count); glycaemic management stays with the diabetes conditions; blood pressure as an exposure stays with hypertension. AKI is treated as a border covered inside the progression page and recorded as a candidate future condition rather than absorbed.

Flagged for the build pass. Anchors were resolved in a scoping session and must be re-verified. No anchor exists yet for RAS blockade landmarks, anaemia trials, CKD-MBD, statins in CKD, dietary protein restriction, dialysis-initiation timing, conservative care, ADPKD or IgA nephropathy — all live-search tasks.

Follow-ups. Build with tools/build-condition.sh chronic-kidney-disease; auditor must differ from the writer.

2026-09-02 — Independent audit (auditor: Claude; author: codex)

Full citation-to-claim audit of all 24 wiki pages and every literature artifact. A different engine from the build, per CLAUDE.md.

What was checked

Object Count Method
Unique PubMed records 132 (125 at build + 7 added at audit) Re-fetched live via E-utilities esummary and efetch; author, journal, year, volume, pages and title compared with every bibliography and reference line
Records with retrievable abstracts 130 of 132 efetch XML; the two exceptions are the KDIGO 2024 CKD guideline (38490803) and the KDIGO 2017 CKD-MBD update (30675420), which carry no PubMed abstract
Claim→citation pairs 442 inline citations outside evidence ledgers and reference lists Every page read in full; each claim compared with its source abstract
Numeric claim lines 59 Every effect size, confidence interval, sample size, percentage and follow-up duration checked digit-by-digit against the abstract
ClinicalTrials.gov records 20 (17 catalogued + 3 added) v2 API; title, status, phase, enrolment and dates re-verified
Non-journal URLs 15 (11 patient organizations + 4 guideline pages) Re-requested; all returned HTTP 200; NICE, UKKA and Kidney Health Australia page text re-read
Asserted absences re-searched 7 Fresh PubMed searches on finerenone outside type 2 diabetes, combination/factorial CKD trials, semaglutide in CKD without diabetes, conservative-management RCTs

Errors found and fixed

1. Stale absences — the most consequential finding. Seven records published before the build date were missed, and six pages asserted gaps that had already closed:

  • FIND-CKD has reported. mineralocorticoid-receptor-antagonists.md said "Evidence for non-diabetic CKD … remains incomplete" and clinical-trials-landscape.md listed the trial as merely COMPLETED with publication "to be tracked". Heerspink 2026 (PMID 42246672, N Engl J Med 2026;395:533-545) randomized 1,584 adults with CKD and no diabetes: total eGFR slope difference 0.7 mL/min/1.73 m²/year (95% CI 0.3–1.1), composite kidney-or-cardiovascular HR 0.77 (0.60–0.99), hyperkalaemia 17.0% vs 13.3%.
  • The INFINITY pooled analysis exists (Neuen 2026, PMID 42248158, Lancet): 14,574 participants across FIDELIO-DKD, FIGARO-DKD and FIND-CKD; kidney composite HR 0.76 (0.68–0.86), kidney failure alone 0.85 (0.74–0.99), all-cause death 0.88 (0.79–0.99).
  • A randomized combination trial exists. diabetic-kidney-disease.md said "randomized evidence for the full combination and sequence is absent" and clinical-trials-landscape.md said "the major missing design is factorial or strategy randomization". CONFIDENCE (Agarwal 2025, PMID 40470996) randomized 779 people 1:1:1 to finerenone, empagliflozin or both; combination lowered uACR 29% more than finerenone alone (ratio 0.71, 0.61–0.82) and 32% more than empagliflozin alone (0.68, 0.59–0.79) at day 180. Its prespecified secondary analysis (PMID 41493296) found combination did not mitigate hyperkalaemia versus finerenone alone. Claims rewritten to say precisely what remains missing: an outcome-powered trial across all four classes.
  • Finerenone has type 1 diabetes evidence (FINE-ONE, PMID 41780000): 242 participants, uACR 25% lower than placebo at six months (ratio 0.75, 0.65–0.87). The table row reading "T2D only" was corrected.
  • A glomerular-disease analysis exists (Neuen 2026, PMID 42246414, JAMA): prespecified exploratory FIND-CKD subgroup, 903 participants, eGFR slope difference 0.73 (0.22–1.24), kidney failure or ≥40% eGFR decline HR 0.74 (0.57–0.97).
  • Semaglutide's kidney signal now extends beyond diabetic CKD (Mann 2026, PMID 42567173): SELECT + FLOW + SOUL pooled, 30,787 participants, kidney composite HR 0.84 (0.77–0.91).
  • Three new NCT records verified live and added: NCT05254002 (CONFIDENCE), NCT05901831 (FINE-ONE), plus the FIND-CKD publication link on NCT05047263.

2. Systematic study-design mislabelling — 194 rows across all 24 pages. The per-page "Evidence ledger" assigned each record a design descriptor; 41 of 125 records carried the wrong one, in the direction that most misleads about evidence strength. CREDENCE (PMID 30990260, on 15 pages), AMBER (9 pages), ONTARGET (7), VA NEPHRON-D (7), NefIgArd, BEACON, HALT-PKD, CHOIR, AURORA, ASCEND-D, NEUTRALIZE and the FIDELITY stage-4 analysis were all labelled "Observational or conceptual evidence; association is not treatment effect" — every one is a randomized trial. Five meta-analyses (including Tangri 2016 and the Inker 2019 IPD meta-analysis) were labelled observational; two editorials and a KDIGO controversies conference were labelled as primary evidence; and four modelling studies plus four consensus/Delphi exercises were labelled "Intervention study", including Neuen 2024, which is an actuarial projection, and Geroldinger 2023, which is a retrospective cohort. All corrected against PubMed publication types with hand adjudication, and two new categories added for modelled projections and consensus exercises.

3. Pseudo-citations in per-page open questions — 120 instances. All 24 pages carried an identical five-question boilerplate block, each question tagged with a citation the source does not support: a question about minimum outcome sets was attributed to Hostetter 1981, a 1981 rat micropuncture study; a surrogate-validation question to Wen 2014, a paper on race-specific relative risks; a competing-death question to Tangri 2016, a KFRE validation. Tangri 2016 was cited this way on eight pages without appearing in those pages' reference lists at all. All 24 blocks were replaced with page-specific questions grounded in that page's verified sources, cross-linked to the OQ IDs in OPEN-QUESTIONS.md.

4. Wrong number. kidney-replacement-therapy-decisions.md stated IDEAL's "median start was 9.0 versus 7.2". The trial reported median time to initiation of 1.80 versus 7.40 months; 9.0/7.2 is a different quantity and appeared without units. Replaced with the abstract-verified figures, plus mortality 37.6% vs 36.6% over median 3.59 years. The condition's own landmark note (cooper-2010-ideal.md) already had it right.

5. Direction of effect misrepresented. diet-and-nutrition.md said later syntheses find "small or uncertain effects" of protein restriction, citing Yan 2018. That meta-analysis of 19 trials (2,492 participants) actually found kidney failure OR 0.59 (0.41–0.85) and ESRD OR 0.64 (0.43–0.96) — a significant positive result. Rewritten to present MDRD's null primary endpoint and the meta-analysis's positive one as the live contradiction they are, rather than collapsing both into "small".

6. Overstatements corrected against source. - causes-and-aetiology.md and inherited-and-glomerular-disease.md used Benson 2020 to support genomic testing resolving uncertain diagnoses. That study found a molecular diagnosis in 2 of 50 (4%) unselected biopsy referrals and concluded testing is most valuable when a heritable form is already suspected. The yield and the authors' own qualification were added. - mineralocorticoid-receptor-antagonists.md said the FIDELITY stage-4 subgroup "retains benefit". The paper reports that the kidney composite violated the proportional-hazards assumption, with protection only to about two years and inconsistent direction thereafter; the cardiovascular HR was 0.78 (0.57–1.07), crossing 1.0; hyperkalaemia was 26% vs 13%. Stated accordingly. - ras-blockade-and-blood-pressure.md attributed increased acute kidney injury and hyperkalaemia jointly to ONTARGET and VA NEPHRON-D. ONTARGET's renal paper reports neither; it reports a worse composite renal outcome (HR 1.09, 1.01–1.18) and steeper eGFR decline. The AKI and hyperkalaemia rates come from VA NEPHRON-D alone. Split and quantified. - mineral-and-bone-disorder.md said Xu 2022 found "limited evidence for patient-important benefit". Its outcome was radiologic calcification, not patient-important endpoints; it found data insufficient or conflicting for calcification itself, with magnesium and sodium thiosulfate the only consistent signals across 77 trials of median size 50 and median duration 12 months. - symptom-burden-and-patient-experience.md said a standardized life-participation measure "was developed"; the SONG workshops endorsed SONG-LP by consensus and stated explicitly that piloting and validation in non-KRT CKD remain outstanding. - pathophysiology-and-progression.md said the remnant-nephron model "established" that hyperfiltration becomes injurious; Hostetter 1981 is a rat micropuncture study whose own conclusion is that hyperfiltration "may" be maladaptive. Softened, with the species and study design made explicit. - conservative-and-supportive-care.md attributed Wongrakpanich's HR 0.53 to all 11,515 participants; it comes from a subset (age ≥65, eGFR <15, models adjusted for age and comorbidity). Qualified.

7. Wrong PMID on a landmark note. literature/notes/empa-kidney-2023.md attributed the acute-dip and chronic-slope figures to the 2023 primary paper (PMID 36331190). They are from the 2024 prespecified secondary analysis (PMID 38061371). Corrected, with a note recording the change.

8. Unsupported non-journal claims. The guidelines registry stated the Kidney Health Australia handbook was published in 2024 and "endorsed as an accepted clinical resource by Australian professional bodies listed on the official page". The live page states neither; its only endorsement text is Kidney Health Australia's Deductible Gift Recipient charity status, which is unrelated. Both claims removed from the registry and from wiki/guidelines.md. By contrast, the NICE NG203 dates (published 25 Aug 2021, updated 24 Nov 2021, reviewed 19 Aug 2025), its supersession chain, and the UKKA eCKD Guide's February 2024 update were all confirmed word-for-word from the live pages and are now recorded with that precision.

9. Design mischaracterization in the patient-voice layer. Schade van Westrum 2025 was described as qualitative; it is a nationwide descriptive survey of 414 patients. Replaced with its actual findings (94.7% often fatigued; 86.3% ranked it top-three; 32.1% never or rarely discuss it with a physician; 67.8% untreated).

10. Reference-list gaps. Nine PMIDs were cited in page bodies without appearing in those pages' reference lists. All added. Every page now passes a cited↔referenced consistency check in both directions, and all 132 bibliography "cited by" lists were regenerated from the actual reference sections.

Density improvements made while verifying

Where a claim was correct but thin, the verified numbers were added rather than left as adjectives: 4D and AURORA event rates and confidence intervals, RENAAL's 16%/25%/28% risk reductions, ASCEND-D's co-primary results, CHOIR's HR 1.34 and CREATE's HR 0.78, BEACON's HR 1.83 for heart failure, Tangri 2016's C-statistic of 0.90 across 31 cohorts and 721,357 people, Grams 2023's finding that added slope and comorbidity variables did not improve the KFRE, MDRD's full two-study design, and the DAPA-CKD diabetes/no-diabetes slope interaction (2.26 vs 1.29, p-interaction 0.0049) that the page had flattened into "relative effects were consistent". STATISTICS.md gained 14 rows for the newly found trials and four new caveats; its "Change in CKD deaths +41.5%" row was relabelled, since that figure is the all-age mortality rate change, and the age-standardised rate change was not significant.

Promotion result

22 of 24 pages promoted to curated. Every citation on them resolves, matches the paper claimed, and supports the claim made, after the fixes above.

2 held at draft, with the reason stated on the page and in INDEX.md:

  • guidelines.md — its substance is a synthesis of what guideline documents recommend, and its two central sources (PMIDs 38490803 and 30675420) have no PubMed abstract, so those characterizations could not be checked against retrievable source text this session. Everything about the documents as documents — dates, supersession, URLs, PMIDs — was verified live.
  • red-flags-and-safety-concerns.md — the trial-derived "surrogate traps" material is fully verified, but the opening diagnostic red-flag list and the medication-reconciliation guidance rest on the same non-retrievable KDIGO full text.

Both need a pass by an agent with access to the guideline PDFs rather than their PubMed records.

For the next sweep

  • OQ-2 is now largely answered and OQ-1 narrowed; both were rewritten rather than deleted, and the residual gaps stated. Re-read them before opening new questions in this area.
  • Watch for the FIND-CKD-era evidence entering guidelines: no document in the registry post-dates it, so guidance restricting non-steroidal MRA to diabetic CKD currently reflects its evidence cycle rather than the trial base.
  • The finerenone literature moved substantially between the build and this audit on the same day. Search windows for this condition should not assume the last sweep date is recent enough.
  • Structural note for whoever next edits these pages: the "Decision and interpretation matrix", "What can and cannot be concluded" and "Research-design checklist" blocks are byte-identical across all 24 pages and contribute roughly 40 lines per page of non-page-specific text. They are not errors and were left alone, but they inflate the line counts and should be either differentiated per page or consolidated into one shared methods page.

2026-09-02 — Literature-deepening pass (Claude)

Purpose. The condition was structurally complete but sourced narrowly: 24 pages, 167 lines and 27 distinct citations per page, but only 130 distinct PubMed records across the whole condition — roughly 5 per page — because the same small pool of records was recycled across many pages, largely through the byte-identical "Evidence ledger" blocks. Reference conditions run 15–24 distinct records per page. The instruction for this pass was to widen the evidence base, not to lengthen pages with the same references.

Result.

Measure Before After
Distinct PubMed records across the condition 130 306
Distinct records per page (total ÷ 24) 5.4 12.8
Distinct ClinicalTrials.gov records 19 41
Total wiki lines 4,030 5,377
Mean lines per page 168 224
Pages at status: draft 2 24
Records in BIBLIOGRAPHY.md 132 306
Lines in STATISTICS.md 96 292
Documents in guideline REGISTRY.md 12 17

Every one of the 306 PMIDs and all 41 NCT IDs was resolved live in this session — PMIDs through PubMed E-utilities esearch/esummary/efetch, NCT IDs through the ClinicalTrials.gov v2 API. A final verification pass re-resolved all 306 in a single esummary call (306 requested, 306 returned, no errors) and requested each NCT record individually (41 of 41 returned HTTP 200). No citation in this pass was written from memory; every claim with a number was written from the abstract or record text retrieved in this session.

What was deepened, page by page

Each page received sections built from searches run specifically for that page's topic, not from the shared pool. Representative additions:

  • definition-staging-and-measurement (25→43 records) — within-person variability arithmetic (creatinine CV 5.4% versus random-spot uACR 50.6%), creatinine–cystatin C discordance prevalence and prognosis in 861,000 people, CKD-EPI versus EKFC head-to-head P30 and bias, the age-adaptation controversy with both sides quantified, PCR/dipstick conversion performance, surrogate-endpoint performance for both eGFR slope and albuminuria, the AKD construct, and the eGFR-C monitoring study.
  • epidemiology-and-burden (26→37) — GBD 2023 replacing GBD 2017 as the current estimate (788 million adults, 14.2%, ninth leading cause of death), Hill's stage-specific pooled prevalence, lifetime risk of treated kidney failure, the sex gradient reversing between prevalence and progression, screening ICERs with sensitivity to SGLT2 inhibitor efficacy, workforce capacity, and ambient heat as a measured exposure in DAPA-CKD.
  • pathophysiology-and-progression (28→39) — human single-nephron GFR and nephron number, the human kidney cell atlas and its 28 injury states, G2–M arrest and TASCC biology, the hypoxia circuit, local complement, AKI→CKD quantified from two syntheses whose estimates differ three-fold, and the two urate trials that closed the urate hypothesis.
  • causes-and-aetiology (29→40) — APOL1 with absolute risks by genotype, exome-sequencing yield by phenotype category, aristolochic acid, heat-exposed occupational cohorts with biomarker data, PPI and lithium effect sizes, obesity-related glomerulopathy, MGRS, sickle cell trait, and biopsy-proven CKDu.
  • race-and-the-egfr-equation (25→34) — reclassification magnitudes in three different populations, the masked high-risk subgroup, oncology trial-eligibility exclusion, the OPTN waiting-time modification and its 32% uptake, APOL1 in living donors, and the paediatric CKiD analysis that isolates the creatinine mechanism.
  • Therapy pages — pooled class effects (13-trial and 8-trial SGLT2 meta-analyses), the AKI-reduction paradox, participant-level safety data, sotagliflozin, transplant physiology, cost-effectiveness; FIGARO, the finerenone mediation analysis, two negative spironolactone trials, FINEARTS-HF kidney outcomes, FIDELIO hyperkalaemia; SELECT kidney outcomes, GLP-1-on-SGLT2 additivity, endothelin antagonism, HIF-PHI safety, tirzepatide versus dulaglutide; REIN and its follow-up, benazepril, STOP-ACEi, SPRINT, chlorthalidone, and the 285,124-participant blood-pressure IPD meta-analysis.
  • Complication pages — TREAT, PIVOTAL and its infection analysis, FIND-CKD iron, daprodustat heart failure; EVOLVE, LANDMARK, PRIMO and the Cochrane binder review as four negative results with three competing explanations; SHARP versus AURORA, ISCHEMIA-CKD, CKM staging, anticoagulation, coronary calcification; UBI versus VALOR-CKD, fruit-and-vegetable alkali, DIAMOND, DIALIZE-Outcomes, RAASi discontinuation.
  • inherited-and-glomerular-disease (26→35) — TEMPO 3:4 and REPRISE, DUPLEX as a surrogate counterexample, MENTOR, and the IgA nephropathy pipeline including iptacopan's conversion of a proteinuria signal into an eGFR slope and kidney-failure benefit; plus voclosporin, belimumab and avacopan.
  • Decision and experience pages — PD-versus-HD from day 0 and day 90, FHN Nocturnal and its post-trial mortality signal, incremental initiation, haemodiafiltration, MyTEMP, transplantation versus waitlist, the two Cochrane modality reviews; the Dutch conservative-care cohort with hospital-free days; global symptom-burden prevalence, SONG-HD fatigue, difelikefalin, two exercise trials, depression treatment, sleep disorders, dialysis withdrawal.
  • guidelines (36→47) — the 2026 AHA/ACC/ADA/ASN CKM syndrome guideline (published in duplicate), ADA Standards §11 with its annual cycle, the KDIGO 2013 lipid guideline and its deliberate abandonment of LDL targets, the ERBP commentary on KDIGO 2026 anaemia, and an explicit treatment of grading-system incompatibility and cycle length as sources of disagreement.
  • clinical-trials-landscape (32→36 records, 19→41 NCTs) — the shift to combination-on-background-SGLT2 (EASi-KIDNEY at 11,000 and baxdrostat/dapagliflozin at 5,000), the first nephrology adaptive platform, and a 12-trial table of the glomerular-disease pipeline showing that proteinuria is carrying almost all of it.

Controversies added with citations on both sides

Age-adapted versus fixed CKD definition; population-specific versus race-free EKFC Q-values; whether race-free eGFR helped or harmed access; AKI as cause versus revealed low reserve; whether correcting acidosis slows CKD (open-label positive versus blinded-but-unseparated null); albuminuria surrogacy validated in IgA nephropathy and refuted in FSGS; steroidal versus non-steroidal MRAs as one class or two; why LDL lowering works before dialysis and not on it; the FINEARTS-HF acute dip counted as kidney events; hydration and urate as observationally strong and causally null; potassium restriction versus plant-forward dietary patterns; ESA versus HIF-PHI first-line positioning.

Literature layer

  • BIBLIOGRAPHY.md was rebuilt programmatically from the wiki reference sections, so every record's "cited by" list is now derived from the pages rather than maintained by hand. 306 records across the ten existing topic sections; pre-existing citation text was preserved verbatim.
  • STATISTICS.md gained nine new tables — updated global burden, measurement variability and equation performance, age-adaptation and screening economics, surrogate-endpoint performance, mechanism and progression, aetiology, pooled therapy class effects, complications and replacement, symptoms and glomerular disease, safety and negative results — and nine new entries in the closing conflicts section.
  • REGISTRY.md gained five documents, four per-guideline entries, six disagreement rows (including grading-system incompatibility and revision cadence) and five watch-list rows.
  • OPEN-QUESTIONS.md — OQ-3, OQ-5 and OQ-9 were rewritten with the new evidence and their residual gaps restated rather than being deleted; eight new questions were added (OQ-21 to OQ-28) covering the age-adapted definition, acidosis correction, aldosterone synthase inhibition, anti-inflammatory therapy, sickle cell trait counselling, the failure of observationally-derived targets, pregnancy in prognostic models, and the delivery constraint. The "dots not yet connected" table gained eight junctions (D13–D20).

Deliberately left alone

  • The five landmark notes in literature/notes/ were not touched. They are deep notes on DAPA-CKD, EMPA-KIDNEY, race-free eGFR, IDEAL and FLOW; none of the new material supersedes them, and CONVENTIONS reserves notes for papers that changed practice or opened a field. Candidates from this pass for future notes: the GBD 2023 CKD analysis (PMID 41213283), the APPLAUSE-IgAN 24-month result (PMID 41910396), and the 285,124-participant blood-pressure IPD meta-analysis (PMID 42035778).
  • The patient-voice layer was not modified. It is a non-journal layer with its own ethics rules and verification method (organizations verified by fetching their sites with access dates); re-verifying it was outside the scope of a literature-deepening pass and partially updating it would have left the access dates inconsistent.
  • The shared boilerplate blocks — "Decision and interpretation matrix", "Evidence ledger", "What can and cannot be concluded" and "Research-design checklist" — were left in place on every page. The previous audit flagged them as byte-identical across all 24 pages and recommended consolidating or differentiating them. That recommendation stands and was not acted on here because it is a structural change, not a sourcing one, and this pass was explicitly about substance rather than structure. Note that the "Evidence ledger" tables are the main reason the pre-existing per-page citation counts looked adequate while the distinct-record count was low: they list records that the page body does not use.
  • Existing verified content was extended, never replaced. No previously verified claim, citation or section was deleted from any page.

Errors found and fixed during this pass

  • My own editing helper silently dropped appended reference lines. The script computed the set of "already present" PMIDs after inserting the new body text, so every new reference was skipped as a duplicate. This was caught when the bibliography rebuild reported 130 records instead of 306. A repair script then appended the 215 missing reference lines across all 24 pages, ordered by first appearance in each page body, and a verification pass confirmed that every inline PMID now appears in its page's ## References section with no duplicate numbering.
  • Citation lines for six records were written with incomplete or guessed locator details before the exact record had been retrieved (Canaud 2019, Portilla 2025, Smeijer 2024, Song 2025, McCoy 2025, Ndumele 2026 JACC). Each was re-fetched and corrected. The abstract fetcher was patched to emit the ELocationID pii when a record has no conventional pagination, so that particular guess cannot recur. Four of the six had to be corrected twice, because the first correction was applied before the reference-line bug above was found and therefore matched nothing.
  • One pre-existing bibliography entry from the original build carried a paraphrased rather than actual title (PMID 36316605, recorded as "SGLT2 inhibitors in advanced CKD: systematic review and meta-analysis"; the real title is "Effects of sodium-glucose co-transporter-2 inhibitors on kidney, cardiovascular, and safety outcomes in patients with advanced chronic kidney disease: a systematic review and meta-analysis of randomized controlled trials"). It was corrected in the bibliography and on all nine pages that carried it.
  • Whole-corpus citation verification. After the repairs, all 1,177 citation lines across the 24 pages and the bibliography — 306 distinct records — were checked programmatically against their PubMed XML on journal abbreviation, publication year, volume, and title-word overlap. Zero mismatches remain. This checks that each line describes the record it names; it does not check that the record supports the claim it is cited for, which is the re-audit's job.
  • One trial acronym ("CAPTIVATE") was written for NCT06058585 without evidence from the registry record and was removed.

Status and what the next agent should do

All 24 pages are set to status: draft. A different engine must re-audit. The audit should prioritise:

  1. The quantitative claims added in this pass — every effect size, confidence interval, sample size and percentage — against the source abstracts. There are several hundred of them.
  2. The 215 reference lines appended by the repair script: confirm each matches the PMID it is paired with, since they were written from a mapping file rather than re-derived per page.
  3. The two pages whose pre-existing draft reason is unresolved (guidelines.md, red-flags-and-safety-concerns.md): both still rest partly on KDIGO full text with no retrievable PubMed abstract, and both would benefit from an agent with access to the guideline PDFs.
  4. Assertions of absence — "no trial has", "no guideline specifies", "no study has measured" — appear frequently in the new material because the pass deliberately surfaced gaps. Stale absences are the most common real error; each should be re-searched.

For the next sweep

  • The distinct-record density is now 12.8 per page against a reference range of 15–24. Further widening is possible but should be driven by page-specific gaps rather than by the metric; the pass stopped adding where additional citations would not have supported anything a reader needed.
  • The IgA nephropathy pipeline is moving fastest of any area here: six agents in phase 3, none comparative, almost all with proteinuria primary endpoints. Expect this section to need revision at every sweep.
  • The two aldosterone synthase inhibitor outcome trials (NCT06531824, NCT06742723) do not report before 2028 but are the largest bet in the current pipeline; the watch list tracks them.
  • GBD 2023 has superseded GBD 2017 for CKD burden. The two use different age denominators and are not a time series — a caveat now recorded in STATISTICS.md that any future sweep should preserve.

2026-09-02 — Independent audit of literature-deepening additions (auditor: Codex; author: Claude)

Audited the new material added by Claude's literature-deepening pass on all 24 pages that it returned to status: draft. The change boundary was reconstructed without repository history: the union of the pre-existing evidence ledgers is exactly 130 PMIDs, matching the logged pre-deepening corpus, while the post-deepening wiki contained 306. The pass therefore added 176 distinct PubMed records and 215 per-page reference lines.

Verification performed

Object Count Method
Deepening-pass PubMed additions 176 distinct records Re-fetched through live PubMed E-utilities; identifier, title, author, journal and year checked against the claim and reference line
Appended per-page reference lines 215 Checked against the live records; all resolved to the paper named
New claim lines using appended records 409 Read against retrieved abstracts/record text; 315 contained quantitative content and were checked for value, direction, endpoint, population and follow-up context
Current wiki PubMed corpus after corrections 310 distinct records Final live batch efetch: 310 requested, 310 returned
Current ClinicalTrials.gov corpus after corrections 42 NCT records Every identifier requested individually from the live v2 API; 42 returned
Asserted evidence gaps 19 targeted search strategies Fresh PubMed esearch/esummary plus ClinicalTrials.gov searches where the register was the relevant source

Errors and padding removed

  1. Wrong order of magnitude. epidemiology-and-burden.md and overview.md called lifetime risk of treated kidney failure “roughly two orders of magnitude” below CKD prevalence. The cited values are 3.14% in men and 1.42% in women versus 14.2% adult prevalence — about fourfold and tenfold lower, respectively, and not directly commensurable because populations and methods differ. Corrected in both places.
  2. Stale transplant absence. sglt2-inhibitors.md and clinical-trials-landscape.md said transplant evidence stopped at a 52-person physiology study and that no outcome trial was recruiting. Live searches found a 208-person 12-month randomized trial in chronic allograft dysfunction (PMID 42102257) and the recruiting 330-person DEAK chronic-slope trial (PMID 42315275; NCT05788276). The pages now state that randomized evidence exists but remains surrogate-based and is not powered for graft failure, cardiovascular events or mortality.
  3. Stale comparative-trial absence. The IgA-nephropathy material said none of the agents was comparative. A 65-person randomized crossover trial compared ambrisentan, henagliflozin and their combination (PMID 41949916). The gap is now stated precisely: no phase 3 comparative clinical-outcome trial establishes sequence.
  4. New potassium-liberalisation evidence omitted. A 16-person randomized crossover pilot tested a potassium-rich fruit/vegetable/nut diet in advanced CKD with sodium zirconium cyclosilicate rescue (PMID 42636918). Added with its two hyperkalaemia events and explicit limits: six weeks, selected patients, biochemical feasibility rather than clinical benefit.
  5. Pipeline overstatement. “Every large CKD outcome trial” was said to place SGLT2 inhibition in both arms, although the same table included SGLT2-versus-placebo and non-SGLT2 designs. Narrowed to the three largest general-CKD phase 3 programmes to which the statement applies.
  6. Historical overstatement. The IgA-nephropathy page said the field had no disease-specific licensed therapy in 2023. Removed; disease-specific approval preceded that date.
  7. Missing context. AURORA 1 quoted the voclosporin response as 41% against only “significantly fewer” controls. Added the meaningful comparator: 23% (40/178), OR 2.65 (1.64–4.27).
  8. Citation padding and cross-study rhetoric. Removed the uncited inference that G1–G3 prevalence means patients “feel well”; removed rankings of intensive glycaemic control and bariatric surgery against unrelated drug trials; removed the claim that sodium-restriction effects were “comparable to adding a drug”; changed statistical albuminuria mediation from causal language; removed causal claims from achieved haemodiafiltration dose, the FINEARTS-HF acute dip and the urate trials; and removed a transplant hazard-ratio ranking across unlike study designs. These citations supported the underlying numbers but not the rhetorical conclusion attached to them.
  9. Additional precision fixes. The heat analysis is now framed as an exposure requiring prospective testing rather than an established modifiable cause; HIF-PHI placebo-versus-ESA asymmetry is not assigned a mechanism; CKD incident-risk development-cohort incidence is not presented as population risk; BMI attenuation is not quantified as causal mediation; earlier CKD-MBD biochemical trials are distinguished from absent patient-important-outcome trials; and the exercise heading and adherence explanation were de-hyped.

Citation breadth and promotion

The deepening pass did widen the evidence base rather than merely lengthen pages with the landmark set: 130 distinct wiki PMIDs before deepening (5.4 distinct condition records per page) → 306 after Claude's pass (12.8 per page) → 310 after this audit's four corrective records (12.9 per page). The 176-record widening is substantive, although 12.9 remains below the 15–24 reference-condition range logged by the author; further additions should remain gap-driven rather than metric-driven.

22 pages promoted to status: curated. guidelines.md and red-flags-and-safety-concerns.md remain draft for the pre-existing unresolved reason: central recommendation and red-flag claims depend on KDIGO full text not present in the PubMed records, so those statements still require direct guideline-PDF verification. All body/reference PMID sets match, reference numbering is consecutive, no duplicate reference number exists, and no unresolved [unverified] marker remains on the 22 promoted pages.