Curation log — hypertension¶

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

2026-09-01 — Independent audit (auditor: codex; author: claude)¶

Audited every one of the 20 canonical wiki pages and all 14 literature artifacts. This was an independent audit: claude authored the material and codex audited it. The audit checked 1,872 claim–citation pairs (body citations, excluding each page's reference list and the bibliography), 3,728 PMID mentions representing 794 unique PubMed records, and 84 NCT mentions representing 61 unique ClinicalTrials.gov records. Every unique PMID was fetched live from PubMed E-utilities in this audit session; every unique NCT identifier was fetched live from the ClinicalTrials.gov v2 API. All 794 PubMed records and all 61 trial records resolved. The 1,062 wiki reference-list entries (768 unique records) were also checked for closure and identity; every PMID used in a page body appears in that page's References section. The bibliography contains all 794 records exactly once. Eighteen unique external literature URLs were checked: 16 resolved directly, and two official organisation pages that rejected automated retrieval with HTTP 403 were confirmed through live indexed-web results.

Errors found and fixed¶

Area	Error in authored material	Correction made
NICE thresholds and targets	The NICE values carried an unresolved verification flag, and the build log said the full guidance had not been retrieved.	Retrieved the current NG136 recommendations live and replaced the flag with the supported paired clinic/out-of-office diagnostic thresholds and the under-80/80-plus clinic and out-of-office treatment targets; added the indexed NICE record to the page references and bibliography.
Pregnancy treatment targets	The page said no randomized trial had tested a target below 140/90 mm Hg.	Added CHIPS (randomized diastolic target 85 versus 100 mm Hg), narrowed the remaining gap to a systolic target below 140 mm Hg, and recorded the live GOALPOST trial.
Labetalol versus nifedipine in pregnancy	The page said no adequately powered direct trial existed.	Added the 894-participant severe-hypertension trial (primary-control success 84% versus 77%; p=0.05) and the 323-participant postpartum trial (readmission 1.2% versus 8.1%; adjusted odds ratio 0.12), with the claim limited to each study's population and endpoint.
Renin-guided treatment	Several pages said it had never been randomized.	Added the 44-person comparative randomized pilot (similar pressure change, control 62.5% versus 25%, fewer medications), the RETAME-PA protocol and live trial record, and the live SPIRO-First record; narrowed the open question to definitive outcome-scale evidence.
Deprescribing in frail older adults	The pages said the strategy had never been randomized.	Added RETREAT-FRAIL (1,048 participants; death 61.7% versus 60.2%, hazard ratio 1.02, 95% CI 0.86–1.21; systolic pressure 4.1 mm Hg higher after deprescribing) and the completed 522-participant OptimizeBP registry record, for which no indexed results publication was found at the audit date.
Absolute-risk communication	The patient-experience material said absolute-risk framing had never been tested in hypertension.	Added the 210-person randomized decision-aid study (medication use 65.0% versus 57.9%, odds ratio 1.35, 95% CI 0.77–2.36; adherence 43.7% versus 40.2%, odds ratio 1.15, 95% CI 0.67–2.00) and the 159-person EFFRICO trial, which did not significantly change the risk score; narrowed the evidence gap to isolated absolute-versus-relative framing.
Community health-worker outcomes	The CRHCP landmark note and implementation page stopped at early blood-pressure outcomes and said outcome follow-up was awaited.	Added the three-year cardiovascular result (1.62% versus 2.40% per year; hazard ratio 0.67, 95% CI 0.61–0.73) and seven-year result (2.4% versus 3.0% per year; hazard ratio 0.76, 95% CI 0.72–0.81), including the reported hypotension and mild-hypokalaemia signals, and corrected the landmark note and statistics table.
Salt-substitution characterization	The SSaSS note called it the only lifestyle intervention with outcome evidence.	Replaced the absolute claim with the supported, narrower statement that it is one of few hypertension-focused environmental interventions with hard cardiovascular outcomes.
Trial-registry drift	Eleven completion dates and two enrolment values were stale.	Re-fetched every registry record and corrected the displayed completion dates; corrected OptimizeBP enrolment to 522 and SPIRO-First enrolment to 30.
Unsupported absence claims	Multiple pages used undated, categorical “no trial/no study” language.	Re-ran targeted PubMed and ClinicalTrials.gov searches and replaced the surviving gaps with positively stated, search-dated formulations. These now cover inpatient noncardiac treatment outcomes, diagnostic-threshold outcome trials, intensive targets in frailty, deliberate variability reduction, implementation-model head-to-head trials, a diastolic floor, young low-risk stage-1 event trials, ACCOMPLISH replication with chlorthalidone, sex-specific thresholds, nocturnal targets, processed-food salt substitution, randomized hypertension definitions, adherence testing before devices, resistant-hypertension cardiovascular outcomes, measurement-quality clinical endpoints, and validated treatment-burden instruments.
Citation metadata introduced during correction	Initial draft corrections carried incorrect author/journal metadata for four newly located records.	Re-fetched their PubMed summaries and corrected author, journal, year, volume and page/article strings everywhere before promotion.

Closure and disposition¶

No unresolved verification marker remains. All evidence gaps are positively stated and dated 2026-09-01 rather than asserted as timeless absences.
The retracted ambulatory-monitoring paper and the chronotherapy paper carrying an expression of concern remain present only in explicit publication-status warnings; neither is used as affirmative evidence. Their PubMed publication-status records were re-fetched during this audit.
No unresolved substantive claim–citation mismatch remains. Local links resolve, all body citations close to page reference lists, all bibliography entries are unique, and every cited PMID and NCT record exists and matches the work or trial described.
20/20 pages promoted from draft to curated. The condition status in CONDITIONS-ROADMAP.md was advanced to audited.

2026-09-01 — Full build (engine: claude)¶

Built all 20 canonical pages, the complete literature layer, a tiered open-questions agenda and this index update, from live literature searches run in this session. Every one of the 785 distinct PMIDs cited anywhere in this condition was resolved by a live PubMed E-utilities query during this session, and all 56 ClinicalTrials.gov identifiers were resolved by live v2 API queries. No identifier was written from memory. Both sets were re-verified programmatically at the end of the build: the cited-PMID list was diffed against the session's retrieval log with zero unmatched identifiers, and each NCT ID was fetched individually and confirmed to resolve.

What was built¶

20/20 wiki pages, 2,855 lines total, 37–79 cited PMIDs per page (median ~49), all left at status: draft.
literature/BIBLIOGRAPHY.md — 785 records in 19 domain sections, cross-indexed to every file that cites them. Metadata was regenerated programmatically from live esummary output rather than transcribed, so author/journal/year/volume/pages strings match the database exactly.
literature/notes/ — 7 landmark deep notes: NCD-RisC 2021, BPLTTC 2021, SPRINT 2015, Brown 2020 (primary aldosteronism prevalence), SYMPLICITY HTN-3, SSaSS, CRHCP. Choice recorded: the playbook suggests 3–6; I wrote 7 because the seed identified implementation as this condition's defining problem, and CRHCP is the strongest single demonstration of it.
literature/guidelines/REGISTRY.md — 40 documents with supersession chains (JNC7 → JNC8 → 2017 ACC/AHA → 2025 AHA/ACC; ESH/ESC 2013 → ESC/ESH 2018 → the 2023/2024 European split; JSH 2004 → 2014 → 2019 → 2025; NICE CG127 → NG136; Hypertension Canada 2020 → 2025; ISSHP 2014 → 2018), per-document notes, a numerically stated disagreements table and a watch list.
literature/statistics/STATISTICS.md — 15 tables, every figure carrying source, year, population and method, with a 12-point known-conflicts section.
literature/patient-voice/ — README (method and ethics), organizations.md (18 organisations, each verified by retrieving its site on 2026-09-01), themes.md (9 themes, each with ≥2 independent sources, plus 3 sub-threshold observations recorded honestly), sources.md (annotated, with an 8-point coverage-limits section).
OPEN-QUESTIONS.md — replaced the seed set with 27 questions (12 Tier 1, 15 Tier 2) and a 16-row "Dots not yet connected" table.
INDEX.md — rewritten with at-a-glance counts, six reading paths, page statuses and citation counts, and the anchor-record re-verification table.

Anchor re-verification¶

All five seed anchors were re-fetched and confirmed before use: 34450083 (NCD-RisC, Lancet 2021;398:957-980), 33933205 (BPLTTC, Lancet 2021;397:1625-1636), 26551272 (SPRINT, N Engl J Med 2015;373:2103-16), 32449886 (Brown, Ann Intern Med 2020;173:10-20), 39210715 (McEvoy, Eur Heart J 2024;45:3912-4018). Author, journal, year, volume and pages matched the seed's record in every case. As the seed instructed, they were treated as orientation only; every topic was searched afresh.

What the seed questions became¶

Seed	Outcome
SQ-1 (why control stays low)	Became the spine of adherence-and-implementation.md and OQ-10/OQ-18. The implementation trials (CRHCP, barbershop, HOPE 4, Kaiser) were assembled as instructed; the finding that none has been compared with another is now OQ-10.
SQ-2 (how low, in whom, at what cost)	Became blood-pressure-targets.md. The seed asked for the trials side by side with populations, methods and harms rather than a winner; that is the seven-row table. The benefit–harm ledger (NNT 58 vs NNH 55) is the honest answer.
SQ-3 (universal PA screening)	Answered, and the answer had moved. The 2025 Endocrine Society guideline (PMID 40658480) now suggests screening all hypertensive people — published after the seed was written. The question was re-cast as OQ-1 (the false-positive burden nobody has modelled) and OQ-2 (whether subclinical disease should be treated, now testable via NCT07727252).
SQ-4 (does measurement change who is treated)	Became OQ-3 and the measurement section of definition-measurement-and-diagnosis.md. The seed asked for offsets; the honest finding is that pooled attended-vs-unattended heterogeneity (I² 97%) makes a conversion factor unusable, which is a stronger result than a number would have been.
SQ-5 (lifestyle evidence that survives real life)	Became lifestyle-and-dietary-management.md. Salt-substitute outcome trials were retrieved as instructed. The organising finding — environment-level interventions outlast behaviour-level ones — is stated explicitly and carried into the patient-voice layer.
SQ-6 (did renal denervation fail)	Became device-and-interventional-therapy.md, with the trial sequence in chronological order including every negative result, and both errors named: over-claiming from unblinded data, and over-generalising from one negative trial.

Searches run¶

Approximately 75 PubMed E-utilities searches and about 30 targeted efetch abstract retrievals across: measurement and device validation; ambulatory/home/nocturnal prognosis; white-coat and masked phenotypes; global and national epidemiology; care cascades; sex differences and life-course tracking; sub-Saharan African burden; GBD attributable burden; dose–response and variability; J-curve; pressure natriuresis and its critics; RAAS, sympathetic, immune, microbiome, tissue-sodium and obesity mechanisms; GWAS, rare variants, PRS, monogenic forms and pharmacogenomics; primary aldosteronism prevalence/outcomes/subtyping/surgery; renovascular disease; OSA and CPAP; phaeochromocytoma, coarctation, fibromuscular dysplasia, drug-induced hypertension; every major target trial and its subgroup analyses; first-line drug classes and Cochrane reviews; ALLHAT/ACCOMPLISH/LIFE/VALUE/ONTARGET/DCP; combination-first strategies; chronotherapy; aldosterone synthase inhibitors, endothelin antagonism and siRNA; PATHWAY-2 and resistant-hypertension epidemiology; the full renal-denervation sequence plus baroreflex and AV-anastomosis devices; pregnancy (CHAP, CHIPS, ASPRE, POP-HT, BUMP 2) and long-term maternal risk; CKD/dialysis/transplant, diabetes, frailty, deprescribing, ethnicity, HIV, paediatrics; adherence measurement and implementation trials; hypertensive emergency and inpatient over-treatment; guidelines from 12 bodies including historical ones; and qualitative patient-experience literature. Plus about 10 ClinicalTrials.gov v2 queries and 30 organisation-website retrievals.

Judgement calls, recorded¶

Seven landmark notes rather than six, for the reason above.
PMID 29669232 (Banegas, N Engl J Med 2018, Spanish ambulatory registry) was deliberately not used anywhere as evidence: PubMed flags it as a retracted publication. The 2023 Lancet analysis of the same registry (PMID 37156250) is used instead, and the exclusion is stated in definition-measurement-and-diagnosis.md and in the statistics caveats so a later reader does not "restore" it.
The Hygia Chronotherapy Trial (PMID 31641769) is presented as a contested claim, not as evidence. Its journal Expression of Concern (PMID 32318736), the investigators' rebuttals, and the neutral TIME trial (PMID 36240838) are all cited so the reader can see the shape of the dispute.
Three organisations could not be described from their own websites — the American Heart Association and US CDC returned HTTP 403 access-control responses, and the Japanese Society of Hypertension's English page redirected to a Japanese error page. Rather than describe them from memory, they are listed in a separate "referenced but not directly verifiable" table with the limitation stated. One candidate organisation (a national heart foundation in India) resolved to an unrelated NGO consultancy and was dropped entirely.
The 36-month SYMPLICITY HTN-3 result is not presented alongside the 6-month randomised figure in the statistics tables, because unblinding and crossover with imputation mean it is not the same kind of estimate. This is stated in both the wiki page and the statistics caveats.
literature/BIBLIOGRAPHY.md topic assignment is algorithmic — each record is filed under the earliest-in-sequence domain among the pages that cite it. A record cited by several pages therefore appears once, under one heading, with all citing files listed. This is deliberate: one record, one line.
Where a guideline's primary document is not indexed in English (2024 Chinese guidelines; JSH 2025), it is catalogued through the indexed highlights or comparative papers and flagged on the registry watch list rather than cited as if the primary text had been read.

Could not be verified / deliberately omitted¶

No NICE guideline full text was retrieved. NG136 is catalogued via its PubMed guideline record (PMID 31577399) and the update summary (PMID 32001477); its specific numeric recommendations are described only to the extent those indexed sources support them.
JNC 8's non-endorsement by NHLBI is stated in the registry as a fact about its status; it rests on the guideline's own published account (PMID 24352797) and general knowledge of the episode, and an auditor may wish to source it more tightly or soften it.
No non-English database was searched for the patient-voice layer; this is stated as a coverage limit in literature/patient-voice/sources.md.
The May Measurement Month campaign's published results papers were not retrieved — a title-word search returned unrelated records and the campaign is cited only via its verified website. An auditor should search for the ISH-authored MMM results series by author or journal rather than by campaign name.
No figure in this build was taken from a source's abstract-free record. Where only title-level metadata was available, the record is used for existence and attribution, not for a number.

Follow-ups for the auditor, in priority order¶

Re-fetch and check every numeric claim in the statistics tables and the TL;DRs. These carry the highest density of extracted numbers and the highest cost if wrong.
Re-run the searches behind every asserted absence. The pages state a number of "no trial has ever…" claims — inpatient BP targets (OQ-4), renal denervation outcome trials (OQ-9), variability-reduction strategies (OQ-14), renin-guided treatment (OQ-12), sex-specific thresholds (OQ-19), absolute-risk communication (OQ-11). Stale absences are the commonest real error, and these are the load-bearing ones.
Confirm the ClinicalTrials.gov statuses, which drift. All 56 were live at 2026-09-01; the phase 3 outcome trials (NCT06742723, NCT07181109) are the ones most likely to have moved.
Check the guideline registry's supersession chains, particularly the JSH sequence and the 2018→2023/2024 European split, which is described from the guideline documents rather than from a source that states the split explicitly.
Check the ALLHAT race-interaction description in special-populations.md — the point that the interaction exists for lisinopril but not amlodipine is doing real interpretive work and should be read against PMID 15811979 directly.
Verify the three unverifiable-website organisations from a different network position; if they resolve, they should be described properly.
Promote nothing without checking the two flagged publications (PMID 29669232 retracted, PMID 31641769 expression of concern) are still handled correctly.

Next: independent audit by an engine other than claude, then publication. Status set to built in CONDITIONS-ROADMAP.md.

2026-09-01 — Seeded¶

Created the condition scaffold per CLAUDE.md's add-a-new-condition workflow, and registered it in CONDITIONS-ROADMAP.md under Cardiovascular.
Ran live PubMed scoping searches and recorded five resolved anchor records. Every PMID came from a live query in this session; none was written from memory.
Measured the literature first, as with uterine adenosarcoma, and got the opposite answer: 724,578 records, the largest in this repository and roughly 2,800 times the size of the adenosarcoma corpus. The failure mode here is sprawl, not thinness.

The scoping decision. Hypertension borders four conditions already curated, so the seed defines the border explicitly in INDEX.md: this condition owns the exposure and its management — measurement, diagnosis, epidemiology, mechanism, secondary causes, targets, drug and lifestyle therapy, resistant disease, devices, pregnancy, implementation, emergencies — and does not own the end-organ diseases. Cardiac end-organ damage stays in hypertensive-heart-disease; stroke in stroke; coronary disease in ischemic-heart-disease; blood-pressure lowering as dementia prevention in vascular-dementia, which already carries SPRINT MIND. The rule given to the build is cross-link, never restate: where a trial matters to two conditions, each cites it for its own endpoint.

A framing instruction for the build. With three-quarters of a million records, almost any loosely phrased question has already been studied. "Open" must mean a decision that specific retrievable evidence leaves unresolved, not a topic nobody has written about; questions failing that test should be dropped rather than dressed up. The seed anchors are orientation points only and explicitly not a foundation — the build should search each topic afresh rather than building outward from five papers.

Next: full build from live literature, then an independent audit by a different model than the one that wrote the pages.