Skip to content

Curation log — Alzheimer's 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.

2026-09-03 — Independent audit of the evidence-breadth pass (Claude; author of the additions: Codex)

Scope. This audit targets only the material Codex added on 2026-09-03; the body checked by the two earlier audits was not re-litigated. The new material was isolated by diffing every page against the site build published on 2026-09-02, which predates the deepening pass, so "new" here means new to the condition, not merely new to a page. Every PubMed record was re-fetched live through E-utilities in this session and every ClinicalTrials.gov record through API v2; nothing below is quoted from memory.

Audit check Result
Pages audited (all 20 left at draft) 20/20
Records added by the pass, live-fetched 56/56 resolved
Reference lines for those records checked against the live record (journal, year, volume, pages) 56/56 matched
Newly used NCT records live-fetched 3/3 resolved; asserted statuses matched
Claim-level checks of the new material every added trial, meta-analysis, guideline and cohort result checked number by number against its abstract
Asserted absences re-run 4
Corrections made 11 content items + 16 structural
Records added by this audit 2 (PMID 40207637; PMID 41307609)
Pages promoted to curated 20/20

Citation breadth — the measure this pass existed to move.

Measure Before (published 2026-09-02) After the pass After this audit
Distinct records cited on wiki pages 308 364 366
Distinct records per page (20 pages) 15.40 18.20 18.30
Mean distinct PMIDs per page 23.6 27.1 27.1
Mean page length (lines) 143 154 154

The pass widened the evidence base rather than padding it, and this is verifiable rather than asserted. Of the 70 page-level citation additions (56 distinct records, some legitimately used on two or three pages), every one was a record the condition did not previously cite anywhere — zero were recycled from the existing landmark pool onto a second page. Distinct records per page rose 19% while mean page length rose 7%. Each page gained between 1 and 6 new records sourced to that page's own topic (fluid biomarkers gained five assay-comparison studies, vascular gained five pericyte/CAA studies, care gained three care-model trials and a discharge-prescribing cohort), and the recurrence that does exist is confined to records that genuinely belong on two pages (New IDEAS on imaging and on diagnostic criteria; the two cost reviews on epidemiology and on care). No page was padded with the condition's existing landmark set, and no page is promoted on that basis.

No fabricated citation was found. All 56 identifiers exist, every reference line names the paper the record actually is, and every quoted effect size, confidence interval, sample size and percentage in the new material matches its abstract except where corrected below. The three newly used trial records were re-queried individually: NCT03887455 ACTIVE_NOT_RECRUITING (1,906 actual), NCT01760005 ACTIVE_NOT_RECRUITING, NCT06424236 TERMINATED with 73 actual — the last matching the 73 participants treated in the gantenerumab open-label extension.

Errors and overstatements corrected.

  1. Wrong analyte attribution (overview). "A blood p-tau217 test spans assay AUCs from 0.642 to 0.947" attributed the bottom of the range to a p-tau217 assay. In Janelidze 2023 (PMID 36087307), 0.642 belongs to a p-tau181/p-tau231 immunoassay; the p-tau217 assays occupy the top of the range. Rewritten as "a blood p-tau test", naming which assays sit at each end.
  2. Stale absence — the experience of anti-amyloid treatment. The patient-experience page and OQ-27 asserted that no qualitative evidence existed on anti-amyloid therapy. A live search on 2026-09-03 found two: 22 people with confirmed AD interviewed after discussing lecanemab with a clinician, reporting information sources, hope, and risk–benefit weighing shaped by family, insurance and trust (Parks 2025, PMID 40207637); and 22 Japanese care partners on donanemab and treat-to-clearance, reporting mental burden, relief at confirmed clearance and continuing worry after stopping (Katayama 2026, PMID 41307609). Both study people considering or supporting treatment. The page and OQ-27 now carry both, and the gap is narrowed and dated to what remains genuinely unstudied: the treatment period itself — infusion attendance, surveillance MRI, and the experience of an ARIA event.
  3. Study population misdescribed (neuropsychiatric symptoms). The suvorexant CSF experiment was described as "a single dose in 12 cognitively unimpaired adults". It is a 38-person randomised trial (placebo 13, 10 mg 13, 20 mg 12) in adults aged 45–65; the quoted effects are the 20 mg arm against placebo (PMID 36897120). Corrected, and the phosphosites named as the record names them.
  4. Cohort split conflated (imaging). Groot 2024's AUC increments were presented as coming from 448 people. They are discovery-cohort figures (n=331; 117 external validation). The AD-dementia base model was also given the all-cause base model's AUC: the base was 0.75 (0.69–0.82), not 0.71. Corrected, and the amyloid-PET increment that did not replicate was added, since "only tau PET replicated" is only meaningful beside what else was significant in discovery (PMID 38857029). The corresponding STATISTICS.md row was corrected too.
  5. Pooled estimate attributed to the wrong denominator (imaging). The AI-imaging SROC-AUCs were attributed to "38 AI-imaging studies"; 38 were included but 28 moderate-to-high-quality studies were pooled (PMID 41061249). Corrected, and the P=0.02 for the intermodal difference added.
  6. Population scope dropped (clinical presentation). Wilkosz's six trajectories come from 201 Caucasian patients with possible or probable AD (PMID 19781112). Labelled, consistent with the treatment the earlier audit gave the Mühlbauer and Katz populations.
  7. One-sided summary (risk reduction). The MAPT blood-amyloid subgroup analysis was summarised as losing significance after multiplicity adjustment, omitting that the per-protocol positive-amyloid comparison (n=154) remained significant at 12 and 36 months after adjustment, and that the authors call the finding a trend requiring confirmation (PMID 37872582). Both results now stand, with the 60-month null in both analyses.
  8. Number without the context that makes it meaningful (care). Post-discharge psychotropic initiation was framed as a safety problem at 6.6%, omitting that 63% of the same cohort was already on a psychotropic before admission and that the authors therefore call new prescribing "relatively uncommon" (PMID 36514208). The paragraph now leads with the baseline, and rests the concern where the paper rests it: 51% persistence beyond 90 days and delirium as the predictor.
  9. Unsupported independence claim (vascular). CSF angiopoietin-2 was called "an independent vascular signal". The study does not test independence from sPDGFRβ — it reports correlation with it (PMID 38182581). Restated as a second barrier-related marker, with what the record actually reports.
  10. Recommendation strength over-assigned (guidelines, registry). The Korean behavioural-symptom guideline grades antipsychotics as a conditional recommendation; the retrieved record states no strength grade for the citalopram advice (PMID 39944528). Both the wiki synthesis and registry row 42 now say so.
  11. Duplicated figures across two pages (care, epidemiology). The same two cost reviews were printed with the same headline ranges in the same framing on both pages. The epidemiology page keeps the severity-graded ranges (its remit); the care page now carries the health-system-specific point — that formal accommodation shifts as much as 67.3% of the burden into direct cost — and cross-links rather than reprinting.

Structural defect fixed on 16 of 20 pages. Every new section was inserted immediately after an existing ## heading and before that heading's own content. The effect was that 16 pages had a heading with no body, and the older section's content appeared to belong to the new one — for example the tau page's therapeutics table sat under "Structure and network propagation", and the trigger-list table on the red-flags page sat under "Two safety transitions". Each new section was relocated to sit before the heading it had displaced, which also puts the new evidence ahead of the summary sections it was inserted into. No prose was changed by the move.

Absences re-run. Four were re-searched live on 2026-09-03. Three were upheld and are now dated to this session: no longitudinal study of insurance, employment or legal-capacity outcomes after a biological AD diagnosis (OQ-25 — the new clinician-documentation study closes only the narrower question of whether practice is uniform, and it is not); no qualitative study of the anti-amyloid treatment period (the narrowed OQ-27); and no study of downstream patient outcomes after direct-to-consumer blood testing. One failed, and is item 2 above.

Nothing verified was deleted. Diffing against the 2026-09-02 build confirms the pass was purely additive: the only non-added lines that differ are link paths rewritten by the site build, not content.

Patient-voice layer. The two records this audit added are filed in literature/patient-voice/sources.md as primary qualitative studies 21 and 22 with their limitations stated, the section-C items renumbered, and the coverage-limits note on the treatment-experience gap rewritten to say what is now covered and what is not.

Literature layer. The 56 records are all present in BIBLIOGRAPHY.md with correct metadata and cited-by paths; 373 records now, after this audit's two additions. STATISTICS.md's new sections match their sources (one row corrected, item 4). REGISTRY.md rows 40–42 (EFNS–ENS/EAN 2015, two Korean Dementia Association 2025 guidelines) match their records, with the correction at item 10; the coverage count of 42 is right. OQ-31 (assay portability) is well-founded on the two head-to-head comparisons it cites. Every wiki PMID appears in its page's reference list and in the bibliography; no cited record is missing from it.

Result. All 20 pages promoted from draft to curated. INDEX.md counts and the curation-state paragraph reconciled. CONDITIONS-ROADMAP.md already reads audited and was not changed.

Follow-ups.

  1. The pass filed all 56 records in BIBLIOGRAPHY.md under a single session-provenance heading ("Evidence-breadth additions — 2026-09-03") with the tag 2026-09-03-deepening, rather than distributing them into the file's 17 topic sections with topical tags as CONVENTIONS §3 specifies. The records are complete and correctly cross-indexed, so nothing is lost; re-filing them by topic is a tidy-up for the next sweep.
  2. clinical-trials-landscape.md reference lines 3 and 4 (PMID 39780249; PMID 41522368 — the evoke design and baseline-characteristics papers) are not cited anywhere in the page body. They predate this pass and were left alone; either cite them or drop them at the next sweep.
  3. Two source records added to this knowledge base's running list of internally inconsistent abstracts: the Care Ecosystem record prints a point estimate that is not centred in its own interval, and the Thyrian record calls a caregiver-burden effect significant (P=0.045) while printing an interval that crosses zero. Both are reported here as the records print them, as with S-CitAD, Graff and DIADS-2. A standing note in CONVENTIONS is now overdue.
  4. Diranersen's phase 2 (NCT05399888) passed its 2026-03-11 primary completion date and should be re-queried for results.

2026-09-03 — Page-by-page evidence-breadth deepening (Codex)

Objective and result. Extended rather than rebuilt all 20 canonical wiki pages. The condition-wide PubMed pool increased from 315 to 371 distinct records (+56; 17.8%), moving the condition-level density from 15.75 to 18.55 distinct records per page. Distinct PMIDs on individual pages now range from 21 to 38; every touched page is status: draft and dated 2026-09-03 for independent re-audit.

What was deepened. Added human Aβ-dimer and presenilin kinetic evidence; long-term lecanemab and gantenerumab extension results with design limitations; quantified collaborative-care and caregiver-component trials; trajectory heterogeneity and operational LATE criteria; management utility versus outcome utility for amyloid PET; multi-ethnic incidence trends and severity-specific costs; rare coding immune variants; tau-fold and connectivity evidence; complement/astrocyte/microglia division of labour; pericyte, barrier and CAA autopsy data; assay-specific plasma-biomarker accuracy; tau-PET prognosis and functional-imaging comparisons; graded symptomatic-treatment effect sizes; replicated and discordant brexpiprazole results; positive-versus-null lifestyle trials; diagnosis, discharge and driving safety transitions; and young-onset family experience. Controversies were preserved explicitly, with evidence on both sides where available.

Shared literature layer. Added every new record to literature/BIBLIOGRAPHY.md in the one-line house format with cited-by mappings. Added 15 rows to literature/statistics/STATISTICS.md covering assays, imaging, treatment, care, safety, copathology and cost. Added three graded documents to literature/guidelines/REGISTRY.md: EFNS–ENS/EAN combination therapy (weak recommendation despite significant effects) and the Korean Dementia Association 2025 cognitive-enhancer and behavioural-symptom guidelines (strong/moderate-evidence and conditional recommendations respectively). Sharpened OQ-17, OQ-23, OQ-25 and OQ-29, and added OQ-31 on assay portability.

Live searches and identifier integrity. Issued 54 topic-specific PubMed searches spanning all page domains, then fetched candidate abstracts and metadata through E-utilities. At validation, all 371/371 distinct PMIDs used anywhere in the condition resolved live; seven records omitted from an initial batched response were re-queried individually and all resolved. All 58/58 distinct NCT IDs used anywhere in the condition resolved live through ClinicalTrials.gov API v2. The three NCT records newly used in prose were also queried directly: NCT03887455 was ACTIVE_NOT_RECRUITING, NCT01760005 ACTIVE_NOT_RECRUITING, and NCT06424236 TERMINATED on 2026-09-03.

Deliberately left alone. Preserved all previously verified claims, tables, references and page structure. Did not inflate complete sections with generic reviews, create new pages or paper notes, alter patient-voice source files, or modify WHATS-NEW.md. The absence of patient-centred qualitative evidence on anti-amyloid infusion/MRI/ARIA experience remains open; the 2026-09-03 search found management-impact studies but no qualifying experience study. Actual longitudinal insurance, employment and legal-capacity outcomes also remain unlocated; the new clinician-interview study closes only the narrower documentation-practice question.

Validation. Every inline wiki PMID appears in that page's References section; no wiki PMID is absent from the master bibliography; all 20 page frontmatter blocks are draft; no new [unverified] marker was introduced; the index counts and statuses match the files. The next session should independently re-check the 56 additions claim-by-claim before promoting any page.

2026-08-31 — Independent audit of the deepening pass (Claude; author of the additions: Grok)

Scope. This audit targets the material added by Grok's substance-deepening pass, not the previously curated body that Codex checked earlier the same day. All 15 pages left at status: draft by that pass were audited. Every PubMed record was re-fetched live through E-utilities in this session and every ClinicalTrials.gov record through API v2; nothing below is quoted from memory.

Audit check Result
Pages audited (all pages left at draft) 15/15
Distinct PMIDs cited on those pages, live-fetched 239/239 resolved
Reference lines cross-checked against the live record (journal + year) 393/393 matched
Distinct NCT records live-fetched 55/55 resolved; asserted statuses matched
Claim-level checks of the newly added material every added trial, meta-analysis and cohort result checked number by number against its abstract
Newly asserted or load-bearing absences re-run 5
Corrections made 21
Records added by this audit 2 (PMID 9110909; PMID 41885326)
Pages promoted to curated 15/15 (condition now 20/20)

No fabricated citation was found. Every identifier the deepening pass introduced exists, and every reference line names the paper the record actually is. The INVOKE-2 numbers the previous log asked the auditor to re-check (week-96 CDR-SB LS differences −0.31, +0.13, −0.17, all P>0.05) match the published abstract exactly, as does NCT04592874 (COMPLETED, 356 actual). Walsh 2002 is confirmed as Nature 416:535-9. NCT01843075 was re-queried and is still UNKNOWN in the registry despite the 2026 publication; the page already says so.

Errors and overstatements corrected.

  1. Husebo trial mis-described. The stepwise-pain-protocol row said "60 nursing homes". The trial randomised 60 nursing-home units within 18 nursing homes in western Norway (PMID 21765198). Corrected, and the 8-week treatment period and the null ADL/cognition results were added, since a 17% CMAI reduction means little without its duration.
  2. Stale absence — vitamin E. The page claimed the TEAM-AD vitamin E result "has not been replicated at this dose in a non-VA population". A live search found the ADCS trial of selegiline and alpha-tocopherol: 341 patients, 2,000 IU/day, two years, non-VA, reporting a delay in a composite endpoint (median 670 vs 440 days, P=0.001) — but only in analyses adjusted for a baseline MMSE imbalance, with no significant unadjusted difference (Sano 1997, PMID 9110909). The text now carries both results and says why two positives at the same dose are not a replication of each other.
  3. Overbroad absence — ChEIs and ARIA. "No retrieved study addresses this" was replaced with the study that does bear on it: a 2026 real-world cohort of 240 antibody-treated patients versus propensity-matched non-monoclonal-treated patients (ARIA hazard ratio 4.65, 95% CI 3.77–5.73; antithrombotic use 1.87, prior stroke 1.59, hyperlipidaemia 1.41) (Dangpiaei 2026, PMID 41885326). The remaining gap is now stated positively and dated: no retrieved study reports whether continuing a cholinesterase inhibitor during anti-amyloid therapy changes ARIA risk or outcome.
  4. Two internally inconsistent source records reproduced as if sound. The Graff occupational-therapy record prints the patient Dqol interval as "0.6-.1" and the DIADS-2 record prints a CSDD interval of "1.65-4.05" that does not contain its own point estimate of 1.2. The deepening pass silently repaired both into plausible intervals (0.6–1.1 and −1.65 to 4.05). Neither repaired bound is in any record. Both are now reported as point estimates with the record defect disclosed — the same treatment the earlier audit gave the S-CitAD abstract.
  5. EXPEDITION3 duration wrong. Given as 76 weeks on the anti-amyloid page and in STATISTICS.md; the trial's primary comparison is at week 80 (PMID 29365294). Corrected in both.
  6. GRADE certainty inflated. Galantamine's effect on global function was described as "probably" improving; the Cochrane review rates CIBIC-plus as low-certainty ("may"). Corrected, and the high-certainty DAD and NPI estimates it omitted were added.
  7. Goate 1991 overstated. Described as "the first demonstration that APP mutation is sufficient". The paper reports co-segregation in one autopsy-confirmed kindred plus a second unrelated family and claims only that some cases of AD could be caused by APP mutations. Rewritten to the paper's own strength. Sherrington 1995 was likewise described as "the more common early-onset locus", which that record does not establish; it now reports what the record says.
  8. Barthélemy 2024 strengthened by one word. "Superior to CSF for tau-PET" in two places; the paper says "generally superior". Corrected on both pages.
  9. Unverifiable study descriptors. "Single-centre" was asserted for the Graff and nabilone trials; neither record states site count. The nabilone record's actual recruitment setting (one long-term-care facility plus geriatric psychiatry clinics) replaced it, and the nabilone row gained the null CGIC result (47% vs 23%, P=0.09) that the positive scale results were quoted without.
  10. Population scope silently dropped. The Mühlbauer Cochrane review covers Alzheimer's and vascular dementia (17 of 26 comparisons in AD specifically); Katz 1999 enrolled 73% AD, 15% vascular, 12% mixed; HARMONY enrolled psychosis in five dementia aetiologies. All three are now labelled on an Alzheimer's-specific page.
  11. Speculation presented as inference. The INVOKE-2 ARIA-like MRI finding was glossed as "consistent with a phagocytic attack on vascular amyloid". The trial does not test that; it is now marked as a hypothesis. The minocycline adverse-effect list said "vestibular"; the record says dizziness.

Padding removed. Three passages quoted a number without the context that makes it meaningful, and are the reason this audit did not simply confirm the pass:

  • The sleep section quoted Herring's 28-minute suvorexant gain and then the Cochrane orexin-antagonist pooled estimate of 28.2 minutes as if the second corroborated the first. The Cochrane estimate derives from a single study of n=274 — that same trial. The text now says so.
  • The two-step plasma p-tau217 workflow was summarised as "88–92% accuracy while avoiding 61–86% of confirmatory tests", pairing the best of each range. Accuracy and test-avoidance move in opposite directions: 85.9% of CSF tests avoided at 88.2% accuracy, 61.2% avoided at 92.0%. Stated as the trade-off it is.
  • The music-therapy row called the social-behaviour advantage over other activities one of the "reliable" effects; it is low-certainty evidence from 4 studies and 84 participants, against a moderate-certainty null for agitation. Rewritten to the certainty grades.
  • One further miscalibration: donepezil and galantamine were said to sit "at or just inside" the ADAS-Cog MCID boundary. Against Muir's published thresholds both fall just short of the +3 mild-AD MCID while exceeding the lower bound of the +2 to +3 MCI band. Stated numerically instead.

Absences re-run and upheld. Four asserted absences survived a live re-search and were left standing, now dated to this session: no randomised trial has reduced dementia incidence through lifestyle modification (the retrieved multidomain literature is mediation analyses and cognitive-composite outcomes); the APA 2016 antipsychotic guideline has not been superseded (the only later record is a 2017 Focus reprint of the same document); no PART consensus record was invented — the tau page names primary age-related tauopathy only as one candidate explanation inside an open question; and the ChEI-during-antibody question above, now narrowed rather than removed.

Literature layer. BIBLIOGRAPHY.md 313 → 315 records with the two additions above; no orphan entries and no cited PMID missing from it. STATISTICS.md carried the EXPEDITION3 duration error and was corrected. REGISTRY.md row 39 (APA 2016) is accurate and honest about the record having no abstract. OQ-20 and D5 match the INVOKE-2 record. Headline counts in INDEX.md reconciled.

Result. All 15 deepened pages promoted from draft to curated; the condition is 20/20 curated. CONDITIONS-ROADMAP.md already reads audited and was not changed.

Follow-ups. (1) The Graff and DIADS-2 records join S-CitAD as source-record defects this knowledge base has now documented three times — worth a standing note in CONVENTIONS if a fourth appears. (2) Whether continuing a ChEI during anti-amyloid therapy changes ARIA risk is a clean, designable question and belongs in OPEN-QUESTIONS.md as a Tier 1 candidate at the next sweep. (3) Diranersen's phase 2 (NCT05399888) has a primary completion date of 2026-03-11 and should be re-queried for results.

2026-08-31 — Substance deepening (Grok; no structural restart)

Job. The condition was fully built and audited but thin (mean ~135 lines and ~20 distinct PMIDs per page). This pass extended existing pages; it did not delete verified content, did not rewrite from scratch, and did not promote anything (all touched pages are status: draft for a different engine to re-audit).

Pages deepened (15 of 20). Added landmark-trial effect sizes with CIs, competing guideline positions, quantified epidemiology, mechanism detail, and both-sides controversies.

Page What was added Deliberately left
symptomatic-and-supportive-therapy.md Per-drug Cochrane (donepezil 2018, galantamine 2024, rivastigmine 2015) with GRADE; Tariot 2004 vs DOMINO on combination; TEAM-AD vitamin E; O'Regan discontinuation meta-analysis; Graff OT; music Cochrane 2025; LipiDiDiet 24/36-month; AAN/CCCDTD5 MCI positions No restart of the AD2000/DOMINO synthesis
neuropsychiatric-symptoms.md Mühlbauer 2021 Cochrane class effects; Katz 1999 risperidone; APA 2016 guideline; HARMONY pimavanserin; DXM-quinidine; nabilone; Husebo pain protocol; DIADS-2; McCleery 2020 sleep Cochrane vs melatonin meta-analysis Did not inflate with unvalidated agents
neuroinflammation-and-glia.md Ewers 2019 sTREM2; INVOKE-2/AL002 (null primary, ARIA-like MRI, NCT04592874); Heneka NLRP3; Howard 2020 minocycline Did not pad with unselected CD33 review papers
fluid-biomarkers.md Olsson 2016 CSF fold-changes; Barthélemy 2024 plasma vs FDA CSF; Brum 2023 two-step workflow; Karikari 2020 p-tau181; Salvadó 2026 plasma staging; 2025 CPG thresholds Did not add every p-tau phospho-site paper
tau-biology-and-spread.md Hutton 1998 MAPT; TANGO gosuranemab; Tauriel semorinemab (individual trial numbers behind the network meta-analysis) PART consensus title search returned 0 hits; not invented
amyloid-biology.md Glenner 1984; Goate 1991; Sherrington 1995; Walsh 2002 oligomers; EPOCH and APECS verubecestat with harm in prodromal disease Did not rewrite the cascade controversy
anti-amyloid-immunotherapy.md EXPEDITION3 solanezumab; TRAILBLAZER-ALZ phase 2 donanemab (iADRS 3.20 already present at n=257) Did not re-argue meaningfulness
risk-reduction-and-prevention.md Barnes 2011 and Norton 2014 PAF lineage; Framingham; Kivipelto CAIDE; Whitmer midlife obesity Did not re-quote Livingston 2024 numbers (still no abstract)
vascular-and-metabolic-contributions.md ELAD liraglutide (null primary); Whitmer midlife obesity Did not treat observational GLP-1–dementia papers as causal
diagnostic-criteria-and-biological-definition.md Jack 2010 hypothetical cascade as the AT(N) predecessor Criteria table left intact
imaging-and-neuropathology.md Plasma vs PET staging (Barthélemy, Salvadó) as the changing comparator for AUC documents IDEAS pair left intact
clinical-trials-landscape.md TANGO, Tauriel, ELAD, INVOKE-2, minocycline, EPOCH/APECS, HARMONY, AVP-923; live NCT statuses 2026-08-31 Pipeline-shape table not recomputed
epidemiology-and-burden.md Framingham four-epoch incidence GBD tables not rewritten
guidelines.md APA 2016 antipsychotic guideline; Cochrane class effects as the evidence behind it NICE/FDA/EMA block not re-fetched
overview.md INVOKE-2 as the first TREM2-agonist RCT Five-sentence field map left intact

Pages left alone (5 of 20). clinical-presentation-and-staging.md (26 PMIDs — already at the citation target), genetics.md (23 PMIDs, 154 lines), care-caregiving-and-health-systems.md (the D-CARE/REACH II/WHELD synthesis is already the controversy), red-flags-and-safety-concerns.md (24 PMIDs, 150 lines), patient-experience-and-advocacy.md (23 PMIDs, 149 lines). These were not padded.

Literature layer. Bibliography 265 → 313 live-resolved PMIDs. STATISTICS.md gained rows for Framingham, CSF fold-changes, plasma-vs-CSF, two-step workflow, EXPEDITION3, TRAILBLAZER-ALZ phase 2, INVOKE-2, ELAD, EPOCH/APECS, TANGO, per-drug ChEI Cochrane, TEAM-AD, atypical-antipsychotic SMDs, HARMONY, Husebo, Norton PAF, Kivipelto, Whitmer, and a 13th known-conflict on TREM2 agonism. REGISTRY.md gained the APA 2016 antipsychotic guideline (row 39). OPEN-QUESTIONS.md: OQ-20 narrowed after INVOKE-2; D5 updated. No new landmark notes.

Searches. ~70 PubMed E-utilities esearch queries plus batched efetch of 59 records; ClinicalTrials.gov API v2 for NCT04592874, NCT03352557, NCT03325556, NCT01584440, NCT05399888, NCT01739348, NCT01953601, NCT01900665, NCT03367403, NCT01843075, NCT03289143, NCT00235716. Title-exact searches were used when keyword searches failed. A PART-consensus title search returned 0 hits and was not back-filled from memory.

Decisions. (1) LipiDiDiet's 36-month completer analysis is reported as a selected-survivor result of a negative 24-month primary, not as a treatment for established dementia. (2) APA 2016 is catalogued from the live PubMed record; class/level tables were not reproduced because the record has no abstract. (3) evoke/evoke+ was already on the clinical-trials page from the audit; it was not re-litigated.

Follow-ups for the auditor. Re-fetch the 48 newly added PMIDs; check INVOKE-2 CDR-SB intervals against the full paper (abstract used); re-query NCT01843075 (registry status UNKNOWN despite a 2026 publication); confirm Walsh 2002 volume/pages (Nature 416:535-9 from live efetch).

2026-08-31 — Independent audit (Codex; author: Claude)

Scope and result. Codex independently audited all 20 files in wiki/ and all 13 Markdown artifacts in literature/ written by Claude. The audit checked 310 identifier records (265 distinct PubMed records and 45 distinct ClinicalTrials.gov records) and 1,445 contextual claim–citation pairs outside bibliography/reference lists. Every PMID was re-fetched through live PubMed E-utilities during this audit session; every NCT ID was re-fetched individually through ClinicalTrials.gov API v2. All 265 PMIDs existed, all 45 NCT IDs returned the same identifier, and the dated NCT statuses used in the text matched the live records. All citations resolved and matched the subject claimed after the corrections below. All 20 wiki pages were therefore promoted from draft to curated, and CONDITIONS-ROADMAP.md was changed from built to audited.

Audit check Result
Wiki pages read and checked 20/20
Literature artifacts read and checked 13/13
Distinct PMIDs live-fetched 265/265 resolved
Distinct NCT records live-fetched 45/45 resolved and identifier-matched
Contextual claim–citation pairs checked 1,445
Bibliography coverage 265/265 cited PMIDs present; no orphan or missing record
Bare [unverified] markers 0
Pages promoted to curated 20/20

Errors found and fixed. Each item below records a content correction, not merely an audit observation.

  1. Stale semaglutide absence. The build said evoke/evoke+ primary results were unavailable. A live PubMed search found the 2026 primary Lancet report (PMID 41865758). Both trials were null at week 104: CDR-SB differences −0.08 (95% CI −0.35 to 0.20; P=0.57) and +0.10 (−0.17 to 0.38; P=0.46). The clinical-trials page, statistics file, bibliography, guideline watch list and OQ-30 were updated; OQ-30 is now resolved rather than open.
  2. Reversed WHELD effects. The neuropsychiatric page printed positive CMAI and NPI-NH point estimates beside negative confidence intervals. The point estimates were corrected to −4.27 and −4.55, respectively, matching PMID 29408901.
  3. Unsafe S-CitAD transcription. PubMed's abstract for PMID 40133524 reports a risk-difference point estimate outside its own confidence interval. The inconsistent numeric estimate was removed from all summaries; the supported null primary conclusion and drug-related QT prolongation remain, with the abstract inconsistency disclosed.
  4. False LMIC evidence absence. The build described LMIC qualitative evidence as almost absent. Live searching located five relevant records: Colombia (PMID 41717228), Iran (PMID 39097934), two Ugandan studies (PMID 37680685; PMID 33043153) and Malaysia (PMID 34368037). The patient-experience page, patient-voice sources/themes/method file, bibliography and OQ-26 now describe the small but real evidence base and narrow the remaining gap to cross-country, patient-led priority setting.
  5. False positive-caregiving absence. The themes file said no second synthesis had been located. The audit found a 41-study integrative review (PMID 29128685) and a separate 17-study synthesis (PMID 30466499). Positive meaning, mutuality, competence and growth are now an established fourteenth theme, explicitly presented as coexisting with—not cancelling—caregiver burden.
  6. False direct-to-consumer guideline absence. The registry said no guideline addressed direct-to-consumer AD blood tests. The 2025 Spanish Society of Neurology positioning document (PMID 40685136) explicitly recommends against such use. It was added to the guideline synthesis and registry; the remaining gap was narrowed to empirical downstream patient outcomes.
  7. Overbroad anticholinergic-trial absence. The red-flags page said no deprescribing trial existed. The audit found AgeWell.de (PMID 40817636) and a 12,787-person high-risk-medication trial in people with dementia (PMID 39432286). The text now distinguishes trials of prescribing outcomes from the still-unanswered question of dementia incidence.
  8. Other absolute absence language. Assertions that no disclosure trial and no agitation/psychosis trajectory trial existed were replaced after focused live PubMed re-queries with dated, outcome-specific evidence gaps: retrieved studies do not establish whether a structured diagnostic encounter changes the listed harms or whether short-term symptom control changes long-term progression.
  9. Regulatory staleness. The guideline page and registry omitted EMA's Kisunla authorisation (24 September 2025), FDA's 2025 subcutaneous lecanemab maintenance and 13 July 2026 at-home starting-regimen approvals, and subsequent NICE appeal/reconsultation activity. Both NICE appraisals remain in development with publication TBC; their latest negative drafts are not final guidance. Current FDA, EMA and NICE pages were directly fetched on 2026-08-31 and the care-capacity discussion was updated to separate infusion demand from continuing biomarker, MRI and emergency-response requirements.
  10. Stale web-access labels. NICE and FDA pages previously labelled 403/404 now fetched successfully and were relabelled. Twelve patient organisations were directly fetched; Alzheimer's Disease International remained HTTP 403 to command-line retrieval but its current official homepage/member directory was present in the live search index. ARDSI returned HTTP 406 and alzheimers.org.in failed DNS; both remain quarantined rather than treated as verified.
  11. Missing and untidy citation wiring. The clinical-trials page gained the live-resolved PMIDs for CitAD (24549548), DIAN (22784036), intranasal insulin (32568367) and the pivotal antibody trials where their results were invoked. Duplicate reference numbering was corrected. The bibliography was expanded from 254 to 265 records and its cited-by paths and all headline counts were reconciled.
  12. Overstatement in synthesis prose. “Burden grows almost entirely through demography” was softened to “largely”; headings and prose implying that amyloid initiation or immune causality were established across all AD were narrowed to what monogenic genetics and pathway association support; and a definitional claim was separated from prognostic evidence.

Remaining dated evidence gaps and access limitations. No bare [unverified] marker remains, and none of the following prevents page curation: (a) the 2026-08-31 PubMed query found no outcomes study of insurance, employment or legal-capacity consequences after a biological diagnosis (OQ-25); (b) the audit query found no qualitative study centred on receiving anti-amyloid treatment, surveillance MRI or ARIA (OQ-27); (c) direct-to-consumer blood-test recommendations exist, but no retrieved study measured downstream patient outcomes; (d) the S-CitAD abstract's internally inconsistent estimate remains a source-record defect, so the number is deliberately not reproduced; (e) the NICE appraisals remain unfinished; and (f) the two South Asian organisation domains above remain in quarantine, while ADI is search-index rather than direct-fetch verified.

Validation. Final checks found 265 unique PMIDs in both the condition corpus and bibliography, no cited PMID missing from the bibliography, 45 unique NCT IDs, no unresolved identifier, no [unverified] text, sequential numbered reference lists, and no broken relative Markdown links within the condition. No file outside conditions/alzheimers-disease/, CONDITIONS-ROADMAP.md, or the requested log/index metadata was modified.

2026-08-31 — Full build (Claude)

What was built. All 20 canonical wiki pages from the INDEX page plan, the complete literature layer, a replacement tiered open-questions agenda, and an updated INDEX. Every page is status: draft; nothing was promoted, because the model that wrote these pages must not audit them.

Deliverable Result
Wiki pages 20/20 written from live literature; ~54,200 words; mean 2,708 words per page (stroke reference: 2,882)
Distinct PMIDs 254, every one resolved through live PubMed E-utilities in this session
Distinct NCT records 45, every one resolved through live ClinicalTrials.gov API v2 queries in this session
literature/BIBLIOGRAPHY.md 254 records in 17 topic sections, metadata regenerated from live efetch output rather than transcribed, cross-indexed to every citing file
literature/notes/ 6 landmark notes: Braak 1991, Jack 2024 (revised criteria), van Dyck 2023 (CLARITY AD), Sims 2023 (TRAILBLAZER-ALZ 2), Boyle 2018 (person-specific pathologies), Palmqvist 2024 (blood biomarkers in primary care)
literature/guidelines/REGISTRY.md 36 records (31 journal-indexed, 5 web-based), supersession chains, a 7-row disagreements table, 5 identified gaps, a 7-item watch list
literature/statistics/STATISTICS.md 13 sections of quick-reference tables with source, year, population and method on every row; a 12-item known-conflicts section
literature/patient-voice/ Four files; 13 organisations verified by direct fetch across six regions; 12 aggregate themes each with ≥2 sources; 2 fetch failures quarantined
OPEN-QUESTIONS.md Replaced the six seed questions with 30 questions (16 Tier 1, 14 Tier 2) and a 14-row "dots not yet connected" table
INDEX.md At-a-glance counts, nine reading paths, page statuses, literature-layer table, and a prioritised list of what the audit should check first

Seed anchors re-verified. All nine anchor PMIDs recorded at seeding (38934362, 29653606, 36449413, 37459141, 37280110, 40156286, 34998485, 39096926, 17568013) were re-fetched at the start of this session; all nine resolved with matching author, journal, year and content. One caveat: Livingston 2024 (PMID 39096926) has no abstract in PubMed, so it is cited for framing only and the population-attributable-fraction numbers on the prevention page are taken from Stephan 2024 (PMID 38824956) instead.

Searches run. Roughly 90 PubMed E-utilities searches across every planned page topic, plus targeted efetch retrieval of about 200 abstracts; about 20 ClinicalTrials.gov v2 queries by intervention name, condition and NCT ID; and 17 direct web fetches of guideline, regulatory and patient-organisation pages. Search terms are recorded per domain in literature/patient-voice/README.md for the patient-voice layer and are reconstructable from the bibliography's topic grouping for the rest.

Deliberate scoping and framing decisions.

  1. The AA-versus-IWG definitional dispute is presented as unresolved content, not adjudicated. Both positions are given with the progression data that support each, because the data genuinely support both readings at different biomarker profiles (amyloid alone: 11–20% five-year progression; amyloid plus tau: 57%).
  2. The clinical-meaningfulness dispute over anti-amyloid effect sizes is likewise presented with numbers on both sides, including the explicit finding that trial effects fall below published MCIDs.
  3. The shared border with vascular dementia is carried by vascular-and-metabolic-contributions.md, which opens with a note stating that the two conditions are curated separately, cross-links the reciprocal page, and explicitly declines to assert a clean separation. Mixed-pathology data are treated as central content rather than as a caveat.
  4. wiki/overview.md was written fresh because the seed had no wiki/ files; the seed's cited overview paragraph in INDEX.md was preserved and extended rather than discarded.
  5. Non-journal sources (guidelines, regulators, patient organisations) are cited in the repository's (Publisher — "Title", URL, accessed date) form, and every one records whether it was directly fetched or only search-result verified.
  6. Forward links to vascular dementia are not hyperlinks. That condition is seeded and its wiki/ is empty, so links to mixed-pathology-and-alzheimer-overlap.md, cerebral-amyloid-angiopathy.md, cerebral-small-vessel-disease.md and nosology-and-diagnostic-criteria.md would have been broken. Those pages are named in prose, marked as planned, and the hyperlink points to the companion condition's INDEX instead. The audit or the vascular-dementia build should convert them to real links once those pages exist. A link check across the whole condition returns zero broken relative links.

What could not be verified, and is flagged rather than hidden.

  • NICE guidance pages (NG97 and the lecanemab/donanemab final draft guidance) returned HTTP 403 to direct fetch on 2026-08-31; the content used is search-result verified and labelled as such in wiki/guidelines.md and the registry.
  • FDA approval and labelling pages returned HTTP 404 to direct fetch through this session's proxy; the content used is search-result verified and labelled. The EMA Leqembi page fetched successfully.
  • evoke / evoke+ (semaglutide) primary results were not retrievable: both trials show COMPLETED status with a September 2025 primary completion date on ClinicalTrials.gov, but no primary-results publication was returned by PubMed searches. Recorded as OQ-30.
  • Two patient-organisation sites (ardsi.org, alzheimers.org.in) failed to fetch (connection reset; DNS failure) and are quarantined in literature/patient-voice/organizations.md. No South Asian organisation is therefore verified in this build.
  • Four asserted absences are recorded explicitly in OPEN-QUESTIONS.md so that the audit can re-run them: empirical work on insurance/employment/capacity consequences of a biological AD diagnosis (OQ-25); qualitative research on the experience of anti-amyloid treatment, infusion and ARIA (OQ-27); evoke primary results (OQ-30); and low- and middle-income-country patient-voice evidence (OQ-26).
  • Four reference lines initially carried article identifiers (afad128, afaf344, eaau5732, eabd1327) that had not been returned by the live records; these were removed mid-build and the references now carry only the verified volume.

Follow-ups for the next session.

  1. Independent audit by a different engine — the immediate next step. Priorities are listed in INDEX.md under "Curation state": numeric transcription on the three highest-density pages, the four asserted absences, the three search-result-verified web records, NCT statuses, and whether the Lancet Commission can be given quantitative content.
  2. Re-query evoke/evoke+ for results and close or update OQ-30.
  3. Retry the two failed organisation fetches and add South Asian, Chinese, Middle Eastern and Spanish-language Latin American organisations; this is the largest single gap in the patient-voice layer.
  4. Search for non-English national guidelines (German S3, Japanese, Chinese, Latin American) absent from the registry.
  5. Re-verify all 45 NCT statuses; several trials have primary completion dates in 2026–2028 and will change status.
  6. Once audited, add the condition to tools/build_frontend.py and publish.

2026-08-31 — Seeded

  • Created the condition scaffold (INDEX.md, OPEN-QUESTIONS.md, LOG.md, wiki/, literature/) per CLAUDE.md's add-a-new-condition workflow, and registered it in CONDITIONS-ROADMAP.md.
  • Defined the canonical 19-page list in INDEX.md (the canonical list for every condition after the two founding ones is its own INDEX Pages table, per CONVENTIONS.md §5). The list was designed against the shape of this condition's literature, not copied from another condition.
  • Ran live PubMed scoping searches and recorded the resolved anchor records in INDEX.md. Every PMID in the seed came from a live query in that session; none was written from memory.
  • Wrote six seed open questions, each tied to an anchor record, to be replaced by a full tiered agenda plus dots table at build time.

Deliberate scoping decision. Alzheimer's disease and vascular dementia are curated as two conditions with an explicit shared border rather than as one. They have distinct literatures, criteria sets and trial pipelines, but mixed pathology is the common case at autopsy, so each condition carries a dedicated overlap page cross-linking the other, and neither may assert a clean separation the pathology literature does not support.

Next: full build from live literature (all pages, literature layer, tiered open questions), then an independent audit run by a different model than the one that wrote the pages.