Curation log — anorexia nervosa¶

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 draft build (Codex)¶

Built all 18 canonical wiki pages at status: draft: overview; diagnosis/classification including atypical AN; epidemiology; mortality/outcome; genetics; neurobiology/cognition; medical complications; refeeding; adolescent and adult treatment; pharmacotherapy; service settings; experimental therapy; severe/enduring illness, capacity and compulsory treatment; patient experience; guidelines; clinical-trials landscape; and red flags/safety.

Built the literature layer: a topic-grouped bibliography; four landmark notes (Arcelus mortality, Watson GWAS, Zipfel ANTOP and Garber refeeding); global guideline registry with supersession chains; quantitative statistics tables; and the four-file patient-voice layer. Replaced the seed open questions with 15 stable tiered questions and ten “Dots not yet connected” junctions. Rewrote the index with reading paths, real page statuses and literature-layer links.

Live PubMed work. Re-verified all 14 seed anchors by ESummary. Ran topic searches for diagnosis/atypical AN; epidemiology/prevalence/incidence; mortality/suicide/recovery; genetics/GWAS; neurobiology/reward/cognition; medical complications; refeeding/hypophosphataemia; adolescent family treatment; adult CBT-E/MANTRA/SSCM/focal psychodynamic treatment; pharmacotherapy/olanzapine; inpatient/day/outpatient settings; neuromodulation/DBS/rTMS/psilocybin; compulsory treatment/capacity/terminal or palliative proposals; qualitative/lived experience/stigma; guidelines; bone/estrogen; and medical management. Retrieved abstracts for quantitative anchors. Final integrity fetch submitted every PMID found anywhere in the condition tree to PubMed ESummary: 87/87 unique PMIDs resolved live on 2026-09-02.

Live trial work. Queried ClinicalTrials.gov API v2 for condition Anorexia Nervosa, tabulated 13 selected records with live status/enrollment, then individually fetched every NCT identifier written in the condition tree: 13/13 resolved live on 2026-09-02.

Guideline and patient-source work. Fetched or search-result-verified NICE NG69 and its 2024 surveillance decision, the 2023 APA guideline, the RANZCP guideline, Beat, ANAD, F.E.A.S.T. and the Australian Government directory entry for Eating Disorders Families Australia. Public sources only; themes were paraphrased in aggregate; no private names or identifying details were recorded.

Scoping choices. Retained atypical AN throughout diagnosis, complications, refeeding and safety. Kept bulimia nervosa, binge-eating disorder and ARFID only in the differential table. Treated “terminal AN” as a live dispute and sourced autonomy/palliative and prognostic/capacity objections without resolving it.

Follow-up for the independent auditor. Check every claim against full text where abstracts were insufficient; normalize incomplete volume/page metadata in bibliography entries; assess whether the intentionally shorter small-literature pages need additional primary studies; re-run guideline currency checks; and verify trial statuses after 2026-09-02. No page was promoted.

2026-09-02 — Seeded (Claude)¶

Created the scaffold: INDEX.md with an 18-page plan and a 14-record anchor table, a seed OPEN-QUESTIONS.md, this log, and the empty wiki/ and literature/ layers. No wiki pages written.

Interpretation recorded. The condition was requested as "Anorexia". That word also names a symptom (loss of appetite, as in cancer anorexia-cachexia), which is a different subject with a different literature. Scoped to anorexia nervosa because the request sat in a list of named diseases and the symptom sense would not support a condition page set. Recorded in INDEX.md so the assumption is visible and reversible; if the symptom sense was intended, this seed should be rebuilt rather than adapted.

Searches run (live PubMed E-utilities, 2026-09-02). Scoping: anorexia nervosa 21,841; anorexia nervosa[MeSH Major Topic] 13,227. Per-page topic counts in INDEX.md, including epidemiology 3,179, severe and enduring/chronic 1,491, mortality 928, family-based treatment 587, refeeding 261, GWAS 219, named adult psychotherapies 200, medical complications 183, compulsory treatment 179, deep brain stimulation 124, lived experience 106, olanzapine 101, neurobiology 85, atypical/ARFID 24.

Query-hygiene note. Three initial queries were over-specified — ANDing many words together returned implausibly small counts (e.g. CBT-E psychotherapy adult returned 39). Re-run with proper alternation and marked † in INDEX.md.

Anchor records resolved live: 14, listed in INDEX.md.

Scoping decisions. Not density-exempt (13,227 MeSH-major records support the target), but several sub-literatures are genuinely small and the builder is instructed to let those pages be shorter and better-sourced rather than padded. Bulimia nervosa, binge-eating disorder and ARFID are excluded and recorded as candidate future conditions; atypical anorexia nervosa is included, because excluding it would reproduce a weight-based diagnostic bias the literature is actively contesting. The ethics material — capacity, coercion, and the contested "terminal anorexia" proposal — gets a dedicated page and is presented as an unresolved dispute with both sides sourced.

Flagged for the build pass. The MANTRA-versus-SSCM query returned nothing and must be re-run with better terms. No anchor yet for NICE/APA guidelines, bone-density and hormonal trials, day-patient versus inpatient comparisons, atypical AN outcome cohorts, or patient-organisation material.

Follow-ups. Build with tools/build-condition.sh anorexia-nervosa; auditor must differ from the writer.

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

Full audit of every wiki page and every literature artifact in conditions/anorexia-nervosa/. The build was written by Codex; this session did not write any of the material it checked.

Scope of checking¶

Check	Volume	Result
Unique PMIDs in the condition tree at audit start	87	87/87 resolved live via PubMed ESummary
Abstracts or full text retrieved and read	71 records	Used to check claim direction and numbers
Claim–citation pairs checked against source	~210	9 substantive errors found (below)
NCT identifiers re-fetched individually from the ClinicalTrials.gov v2 API	13 written + 4 supporting = 17	17/17 resolved; every status and enrollment figure in `clinical-trials-landscape.md` matched exactly
Asserted absences re-searched live	11	6 stale, rewritten
Orphan references (in a `## References` list but never cited in the page body)	26	All 26 resolved — each integrated into the body with a verified claim, none deleted
Relative links	all	0 broken
Guideline and organization URLs re-requested	9	6 resolved; 2 APA URLs returned HTTP 403 to automation; 1 government directory URL timed out
Unique PMIDs at audit close	114	114/114 resolved live

Errors found and fixed¶

Wrong paper (neuromodulation). PMID 37488290 (Majić 2023, Nat Med 29:1906-1907) is a two-page Comment, not a study. It was cited twice as the source of psilocybin feasibility and safety findings. Replaced with the actual phase 1 trial it comments on — Peck 2023, PMID 37488291 (n=10 adult females, single 25 mg dose, NCT04661514) — with Majić retained and labelled as commentary.
Direction inverted (ethics, patient voice). Both Asaria letters (PMID 37400874, 38082348) argue against the “terminal anorexia” label as harmful. They were cited among the arguments in favour of the proposal on severe-enduring-illness-and-compulsory-treatment.md, and used to support a "lived experience is divided" claim on patient-experience-and-advocacy.md and in patient-voice/themes.md. Corrected in all three places; genuine plurality is now sourced to a stakeholder study (Robb 2026, PMID 42410953).
The primary proposal was never cited. A page devoted to the terminal-AN dispute cited neither Gaudiani 2022 (PMID 35168671), which proposed the four criteria, nor Riddle 2022 (PMID 35710504), the principal published objection. Both added, with the criteria stated in full.
Source contradicted (overview, epidemiology). Winkler 2014 (PMID 24857566) was cited for quality-of-life impairment "with variation by diagnosis" and for combining "heterogeneous … instruments". The meta-analysis restricted inclusion to a single instrument (SF-36) and states it was not possible to establish any difference between diagnostic groups. Both statements corrected; the seven-study sample size added.
Misattributed finding (medical complications). "Physiologic transdermal estrogen improves bone density … without fully restoring it" was cited to Misra 2011 (PMID 21698665). The incomplete-catch-up claim is in Misra 2014 (PMID 24731664); Misra 2011 reports the BMD Z-score increase. Attribution split correctly.
Negative result omitted (medical complications). The Singhal 2021 row (PMID 33693703) said only "bone endpoints studied". The trial was negative for the addition of rhIGF-1 — lumbar areal BMD in fact increased more on oestrogen plus placebo (p=0.004). Result now stated.
Unsupported attribution (neurobiology). Toor 2021 (PMID 34052810) — a general review of orexin, sleep and cognition — was cited to support the claim that malnutrition changes endocrine signalling, sleep, attention, cardiovascular function and brain structure. Reframed to what the source shows (orexin's role and its reported alteration in AN); the malnutrition claim now rests on Misra 2014.
Overstated synthesis (adolescents, guidelines). The Cochrane review (PMID 31041816) was summarized as "supports family approaches". It found low-quality evidence of advantage over treatment as usual from two studies and 81 participants (RR 3.50, 95% CI 1.49–8.23), not maintained at follow-up, and no clear advantage over other psychological interventions (RR 1.22, 0.89–1.67). Numbers and certainty ratings added on both pages.
Overstated attribution (ethics) and unsupported trial claim (services). Carney 2019 (PMID 31046931) was cited for coercion causing "trauma" and damaging "trust"; the review argues a rehabilitation/harm-minimization case and reports quality-of-life impairment comparable to depression or schizophrenia, and says nothing about trauma or trust. Separately, the ANDI trial (PMID 24439238) was credited with having "reduced costs"; the trial reports no cost analysis, only that day treatment "might be … less costly". Both corrected, and ANDI's eligibility boundaries (female, 11–18, below tenth BMI percentile, first admission, randomized after three weeks inpatient) added because they carry most of the interpretive weight.

Publication types were also mislabelled by omission in six further places — Bainbridge 2014 (letter), De Vloo 2021 (research letter), McAdams 2017 (comment), Norris 2016 (editorial), Le Grange 2013 (editorial) and Himmerich 2024 (opinion article) were presented as though they were primary studies. Each is now labelled in both the page reference list and the bibliography. These are not errors of fact but they were inflating the apparent evidence base.

Stale absences re-searched and rewritten¶

"Do faster inpatient gains improve one-year remission and readmission?" — answered, and negatively. Golden 2021 (PMID 33753542) reports the StRONG 1-year outcomes: no difference in clinical remission (p=0.42), rehospitalization 32.8% vs 35.4% (p=0.84).
Atypical AN and refeeding — Garber 2024 (PMID 38179719) shows the fixed kcal/day protocol underfeeds atypical AN (32.4 vs 43.4 kcal/kg), producing slower heart-rate restoration and triple the odds of hypomagnesaemia. This reverses the reassuring reading the build had given.
"24 comparative publications" on atypical AN — superseded by an invited 2026 update covering 64 publications (Lee 2026, PMID 42557659), which confirms the cross-sectional picture and still finds no course evidence.
"Compulsory treatment outcome evidence is sparse" — a 4,425-person Danish register cohort now exists (Bager 2026, PMID 42383339: all-cause HR 2.21, suicide HR 5.30), though confounded by indication.
"Terminal AN cannot be answered by consensus alone" — it was not answered by consensus. It was tested empirically in 782 patients (Robison 2024, PMID 38619462, finding improvement rather than terminal decline) and the term was formally disavowed by the original proposal's lead author in 2025 (Gaudiani 2025, PMID 40361218).
Fracture evidence "sparse" — a matched cohort of 7,332 AN patients and 73,215 controls quantifies it (Cyrenne-Dussault 2026, PMID 41109616: osteoporotic-fracture HR 7.50, 95% CI 5.62–10.01, present in men as well as women).

Also added from live search where the build had a silent gap rather than a stated one: Semchishen 2026 (PMID 41536100) and Lai 2026 (PMID 41277145), which reproduce the 2011 mortality benchmark on much larger samples; Song 2026 (PMID 41927769), the integrative GWAS; Larsen 2025 (PMID 40670905), the register-based severity index; Maguire 2024 (PMID 38520886), the negative oxytocin phase II; Filiz 2026 (PMID 42116676), the OPEN olanzapine feasibility study in young people; and the actual rTMS and DBS primary literature (McClelland 2016/2018, Lipsman 2013/2017, Villalba Martínez 2020, Shaffer 2023).

Every surviving gap is now written as a dated, positively-stated evidence-gap sentence ("as of September 2026, no …") rather than a bare marker. There are no [unverified] markers left in the condition tree.

Searches run at audit (live PubMed E-utilities and ClinicalTrials.gov v2, 2026-09-02)¶

AN GWAS 2021–2026 and title-restricted AN GWAS across all years; refeeding RCTs 2021–2026; adult AN psychotherapy RCTs 2021–2026; atypical AN outcome/course 2023–2026; compulsory/involuntary treatment 2020–2026; AN and fracture; AN network meta-analyses 2022–2026; AN suicide cohorts/registers 2020–2026; psilocybin and AN; rTMS and AN randomized; DBS and AN trials; "terminal anorexia nervosa" and the Gaudiani criteria. Guideline and organization pages re-requested individually.

Verification that failed, and what was done about it¶

APA 2023 guideline. Both psychiatryonline.org URLs returned HTTP 403 to automated requests, and the indexed summary article (Crone 2023, PMID 36722117) carries no abstract. The APA recommendation content could not be verified against source. wiki/guidelines.md is therefore held at status: draft, with the reason stated on the page itself, and the APA rows in both the page and the registry are marked as unverified in this session. A companion 2024 APA practice-assessment tool (PMID 38988459) was located and added to the registry.
EDFA directory link. The Australian Government directory URL recorded at build did not respond within 40 s on two attempts. The organization's own site resolved cleanly and is now the primary verified source for that row; the observation is recorded in patient-voice/organizations.md.
Harrop 2023 (PMID 36577133) and Verma 2024 (PMID 37897094). Neither carries a PubMed abstract, and neither full text was reachable (Wiley, no PMC deposit). Their arguments are now cited only at the level their titles support, and that limitation is stated inline on both pages that use them.
NICE NG69 and RANZCP were both retrieved and read in full. Twelve numbered NICE recommendations are now quoted in substance in wiki/guidelines.md, including two that had not been used at all: 1.2.8 (do not use single measures such as BMI or duration of illness to decide whether to offer treatment) and the §1.11 warning that a long admission may institutionalise a person and that lack of change may indicate inpatient treatment is harmful.

Promotions¶

17 of 18 pages promoted to status: curated: overview, diagnosis-and-classification, epidemiology-and-incidence, mortality-and-long-term-outcome, genetics, neurobiology-and-cognition, medical-complications, refeeding-and-nutritional-rehabilitation, treatment-in-adolescents, treatment-in-adults, pharmacotherapy, service-models-and-setting, neuromodulation-and-experimental-therapy, severe-enduring-illness-and-compulsory-treatment, patient-experience-and-advocacy, clinical-trials-landscape, red-flags-and-safety-concerns.

Held at draft: guidelines.md — for the APA verification failure described above, and only for that reason. Its NICE, RANZCP and SAHM content is verified.

Remaining flags for the next sweep¶

Page length. Pages run 69–119 lines against the 150–400 target in BUILD-PLAYBOOK.md, and citations per page run roughly 8–24 against the 25–60 target. The audit raised both (total wiki lines 1,545, up from 1,192; unique PMIDs 114, up from 87) but did not close the gap; this is a build-depth shortfall, not a correctness problem, and it needs a deepening pass rather than another audit.
Retry APA 2023 verification from a session able to reach psychiatryonline.org, then promote guidelines.md.
Retry Harrop 2023 and Verma 2024 full text; if obtained, strengthen the atypical-AN classification-critique claims beyond title level.
OQ-17 is the highest-value open question in this condition: two systematic reviews three years apart both close by saying longitudinal course and outcome evidence for atypical AN is missing, and none has appeared. Watch for the first cohort.
Re-check trial statuses after 2026-09-02; NCT05368844 changed to Terminated with a 2026-08-03 update, so this set is moving.
Watch the SE-AN literature: the "terminal AN" term is disavowed by its proposer, and the terminology that replaces it is actively being negotiated (Bauschka 2025, PMID 40760030; Connor 2026, PMID 42312722).

2026-09-02 — depth pass (Claude), all 18 wiki pages¶

Purpose. The audit entry above closed by naming a build-depth shortfall: 85 lines and ~12 citations per page, but — the measure that mattered — only 6 distinct PubMed records per page condition-wide, against reference conditions running 17–24. The problem was not page length; it was that eighteen pages recycled one small shared pool of 113 records. This pass widened the evidence base rather than lengthening the existing text.

Result.

Measure	Before	After
Distinct PMIDs condition-wide	113	278
Distinct PMIDs per page (mean)	6.3	15.4
Total wiki lines	1,545	2,443
Lines per page (mean)	86	136
Bibliography entries	114	278

Every one of the 165 newly cited PubMed records came from a live E-utilities esearch/esummary/efetch call in this session, and every quantitative claim attributed to one was taken from the abstract or structured record retrieved here. ClinicalTrials.gov figures came from live v2 API calls on 2026-09-02: NCT02488109, NCT00288574, NCT00592930, NCT00140426, NCT03097874, NCT05834010, NCT05814458 and NCT04378101 were each fetched individually and resolved with the status and enrolment quoted; the portfolio counts and the recruiting-trial table came from live count and search queries against the same API.

What was deepened, page by page¶

genetics.md (8 → 29 distinct records). Added the pre-2019 GWAS wave (Duncan 2017 h²_SNP 0.20; Huckins 2018 exome-chip null for large-effect low-frequency variants), age-of-onset genetics (Watson 2022: early-onset AN genetically distinct, MR link from younger menarche), polygenic-score prognostics and the BMI-direction paradox (Johansson 2022), rare-variant work (BBOX1, NNAT, nanopore structural variants), the epigenetics literature with its reversibility finding and its NR1H3 directional contradiction, gene–environment interaction from MoBa (no multiplicative interaction), and the germ-free microbiota transfer experiments plus the first human FMT trial. Added a full Twin and family evidence section with the Bulik 2006 interval (a² = 0.56, 95% CI 0.00–0.87) and the continuum-versus-category dispute (Dinkler 2021).
neurobiology-and-cognition.md (9 → 31). This page previously named domains without reporting findings. Added structural imaging with numbers and reversal rates (King 2015 >85% of cortical surface thinned; Bernardoni 2016 0.06 mm/month; Kaufmann 2020 age not duration predicts restoration; Bahnsen 2022 mega-analysis and virtual histology), the reward-prediction-error series with its trait-versus-BMI-readout problem, the habit model with its supporting (Foerde 2015) and contradicting (Godier 2016) evidence, set-shifting/central-coherence effect sizes and their non-specificity, the autism overlap, interoception, and the leptin/temperature contradiction in the activity-based anorexia model.
medical-complications.md (13 → 25). Added the 354-patient extreme-malnutrition prevalence table, the cardiovascular section rebuilt around Sachs 2016's finding that arrhythmic sudden death is asserted without supporting data and Krantz 2012's finding that QTc does not correlate with severity, starvation hepatitis and hypoglycaemia as a terminal-malnutrition syndrome distinct from refeeding syndrome, SMA syndrome, gelatinous marrow, and two population registry studies of pregnancy and neonatal outcomes.
refeeding-and-nutritional-rehabilitation.md (9 → 19). Added the ASPEN definition and the 0–62% incidence range that survives applying it, the ICU series, prophylactic electrolyte supplementation, the nasogastric route literature, three StRONG secondary analyses (mealtime distress, renal function, pre-admission energy balance), weight-suppression-indexed targets and meal support therapy's effect on tube-feeding rates.
treatment-in-adolescents.md (7 → 19). Added the founding Russell 1987 trial with its actual restriction to non-chronic, early-onset illness; Eisler 2007's five-year expressed-emotion moderator; multifamily therapy (OR 2.55) and its convergence by 18 months; the replicated early-response predictor and the single failed attempt to act on it; the two setting trials; distance delivery and guided self-help; and cost-effectiveness.
treatment-in-adults.md (9 → 18). Added SWAN (28.3% remission at 12 months across three therapies), McIntosh 2005 — the control condition outperforming specialist therapy, which the field has never absorbed — Dare 2001 pointing the other way, Pike 2003 relapse prevention (the largest effect in the adult literature, unreplicated for two decades), the SE-AN trial, the Cochrane setting review, DAISIES, and dropout predictors.
pharmacotherapy.md (10 → 21). Added the fluoxetine relapse-prevention trial and its 2025 depression-moderated re-analysis, the modelled relapse hazard peaking at 60 days, the 1986 cyproheptadine subtype interaction, the null adolescent olanzapine and risperidone trials, dronabinol, and two Danish register studies showing what is actually prescribed — including higher prescribing in severe AN among patients with no diagnosed comorbidity.
service-models-and-setting.md (7 → 20). Added the Cochrane setting review, TOuCAN in full (49% inpatient adherence; no effectiveness difference; specialist outpatient dominant economically), DAISIES, the HoT protocol, three cost-effectiveness analyses, the intensive-community/home-treatment scoping review (46 sources, zero RCTs), residential-care outcome evidence, and the FREED early-intervention series.
guidelines.md (12 → 16, page 110 → 173 lines). Added the RANZCP published version with its harm-minimisation and all-ages manualized-therapy positions, GRADE-graded Canadian guidance (only two strong recommendations, one of them about setting), the Canadian transitions guideline, the German S3 revision, ANZAED workforce standards, the AED "Nine Truths" evidence review, and the international concordance study.
clinical-trials-landscape.md (18 → 21). Added live portfolio counts (286 interventional records; 49 phase 2/3 across the register's entire history; 35 terminated or withdrawn), a themed table of currently recruiting studies, the publication funnel (201 registered → 41 published → 8 prospectively registered → 7 replicated) and the diversity-of-recruitment problem.
diagnosis-and-classification.md (9 → 18), epidemiology-and-incidence.md (11 → 20), mortality-and-long-term-outcome.md (14 → 20), neuromodulation-and-experimental-therapy.md (14 → 23), severe-enduring-illness-and-compulsory-treatment.md (20 → 25), patient-experience-and-advocacy.md (16 → 24), red-flags-and-safety-concerns.md (13 → 20), overview.md (24 → 28) — see the per-page reference lists for detail.

Controversies added with citations on both sides¶

Suicide SMR in AN: 18.1 after methodological re-analysis, 49% lower than earlier estimates, while all-cause SMR barely moved (Keshaviah 2014, PMID 25214371 vs Arcelus 2011, PMID 21727255).
Diagnostic crossover: majority crossover on weekly prospective interview versus infrequent threshold-to-threshold transition in a 9,622-person register (Eddy 2008, PMID 18198267 vs Schaumberg 2019, PMID 29911514).
Leptin versus ambient temperature as the driver of starvation-induced hyperactivity (Exner 2000, PMID 11032380 vs Fraga 2020, PMID 32210308).
The habit model: dorsal-striatal engagement during food choice versus intact general outcome-devaluation performance (Foerde 2015, PMID 26457555 vs Godier 2016, PMID 27497292).
Specialist versus non-specific psychotherapy in adults (McIntosh 2005, PMID 15800147 vs Dare 2001, PMID 11230031).
Compulsory treatment: elevated register mortality versus no quality-of-life or mortality difference in a small four-year pathway comparison, with rising retrospective endorsement of necessity (Bager 2026, PMID 42383339 vs Abry 2024, PMID 37690079).
Adjunctive antipsychotics: a weak positive Canadian GRADE recommendation against NICE silence, with zero randomized aripiprazole trials behind the international guidance that names it (Couturier 2020, PMID 32021688 vs Thorey 2023, PMID 36928656).

Stale absence corrected¶

OQ-8 and wiki/severe-enduring-illness-and-compulsory-treatment.md both asserted that no study had compared compulsory treatment with an alternative pathway on survival, quality of life and trust. A live search found Abry 2024 (PMID 37690079): 23 involuntarily and 25 voluntarily admitted patients compared at ~4 years, with no significant difference in quality of life, BMI or mortality, higher weight restoration in the voluntary group, and increased perceived necessity of the involuntary treatment among those who received it. The absence claim has been rewritten to specify what is actually missing (an adequately powered matched comparison), and the finding is now reported on the page alongside the register cohort it contradicts.

Literature layer¶

BIBLIOGRAPHY.md: 165 new entries in house format, grouped into the existing seven topic sections, each tagged and carrying its cited-by list and a **[deepening pass 2026-09-02]** marker. Bibliography and wiki PMID sets now match exactly (278 = 278), with no orphan references and no page citing a PMID absent from its own reference list.
statistics/STATISTICS.md: five new sections — comorbidity/familial risk and heritability; medical complications at extreme malnutrition; pregnancy and neonatal outcomes; economic burden; trial and registry landscape — plus new rows in prevalence/incidence and mortality. Eight new caveats added, including that suicide-specific SMRs are far more model-dependent than all-cause SMRs, that register and twin heritability answer different questions, and that maladaptive-exercise prevalence varies three-fold by instrument within one sample.
guidelines/REGISTRY.md: eight bodies added to the master table (RANZCP published version, Canada 2020, Canada transitions 2025, German S3 revision, ANZAED standards ×3, USPSTF, ASPEN, AED Nine Truths), six new per-guideline entries, six new disagreement rows and three watch-list items.
OPEN-QUESTIONS.md: four new junctions (D11 microbiome causality, D12 indirect cost, D13 the failed mechanism-to-treatment chain in cognitive remediation, D14 irremediability judgements) and eight new questions (OQ-18 screening, OQ-19 the unpublished 80%, OQ-20 microbiome, OQ-21 crossover, OQ-22 neuropsychological targets, OQ-23 premorbid-indexed weight targets, OQ-24 indirect burden, OQ-25 evidence standards for irremediability). OQ-3, OQ-8 and OQ-15 were rewritten against new evidence.

Deliberately left alone¶

No verified content was deleted or rewritten. Every addition extends existing sections or adds new ones; the audit's corrections, its [audit 2026-09-02] markers, and its inline verification notes are untouched.
The APA 2023 verification gap. Still unresolved — psychiatryonline.org was not retried in this pass, since the failure was network-level and this session had no better access. guidelines.md keeps its draft status note and its unverified markers.
Harrop 2023 and Verma 2024 remain cited at title level; no new access route was found.
patient-voice/ was not modified. It is a distinct ethics-governed layer, its sources are non-journal, and widening it needs live retrieval of organization pages rather than PubMed searches. Its four files remain as the audit left them.
WHATS-NEW.md and INDEX.md counts were updated only where they state page-level facts; the INDEX audit-standing table describes the 2026-09-02 audit and has been left as the record of that pass, with the depth-pass numbers added beneath it.
Landmark notes. No new note was written. Several papers added here are arguably landmark (Russell 1987, Pike 2003, McIntosh 2005, Duncan 2017), but templates/paper-note.md reserves notes for papers that changed practice or opened a field, and the existing four notes already cover the mortality, refeeding, genetics and adult-psychotherapy anchors. This is a judgement call and a candidate for the next sweep.

Status¶

All 18 pages set to status: draft for re-audit by a different engine, as required for a deepening pass. guidelines.md was already draft.

For the next session¶

Audit this pass with a different engine. The highest-risk claims are the quantitative ones taken from abstracts: effect sizes, confidence intervals and prevalence percentages should be re-fetched and re-checked one by one.
Two numbers to check specifically: the EDGI enrolment (ClinicalTrials.gov lists 17,991 actual against 31,671 analysed in Watson 2026, PMID 41115789 — the paper describes recruitment across four countries and the discrepancy is probably scope, but it is stated rather than resolved here), and the FMT pilot (registry lists 17 actual enrolment against 22 recruited in Panah 2026, PMID 41535289 — flagged inline on genetics.md).
Consider whether Russell 1987 (PMID 3318754) warrants a landmark note; it is the trial the entire family-therapy literature descends from, and its actual restriction to non-chronic, pre-19-onset illness is routinely dropped in summary.
Page lengths now average 136 lines against the 150–400 playbook target. Three pages remain under 125 (epidemiology-and-incidence, mortality-and-long-term-outcome, neuromodulation-and-experimental-therapy); they are dense rather than thin, but a further pass could reach the target without padding.

2026-09-02 — independent audit of the depth pass (Codex; additions authored by Claude)¶

Scope. This audit examined the material Claude added in the immediately preceding depth pass. Previously curated material was not treated as new work, except where a new statement depended on it or an audit rewrite would otherwise leave an orphan.

Citation integrity and breadth¶

The bibliography contained 167 distinct PMIDs marked as depth-pass additions at audit intake, not the 165 reported in the preceding log entry. All 167 were re-fetched from PubMed through live E-utilities calls in this audit session. Every record existed; author/year/title checks matched the cited paper. PMID 36722117, the APA guideline summary, still had no PubMed abstract and could not verify the carried-forward recommendation wording.
One depth-pass PMID was removed as padding: Rodgers 2019 (PMID 31084423) is about participant diversity in technology-based eating-disorder prevention research, not diversity in AN treatment trials. The attached AN-treatment diversity claim and bibliography entry were removed.
Six PubMed records were added after live retrieval because the absence searches disclosed relevant evidence or the depth-pass account had become stale: Ward 2015 (PMID 25545619), Brieva-Toloza 2024 (PMID 38389169), Cardi 2024 (PMID 38841708), Couturier 2026 (PMID 42009386), Ewis 2026 (PMID 42226264), and Boehme 2026 (PMID 42386712).
Val-Laillet 2015 (PMID 26110109) became an orphan when the neuromodulation section was rewritten around direct AN evidence and was removed from that page and the bibliography.
Citation breadth therefore moved from 278 distinct condition-wide PubMed records at audit intake (15.4 records/page using the repository's condition-wide ÷ 18 convention) to 282 (15.7/page). For clarity, the actual mean of page-local distinct PMID counts is 22.6 because records cited on several pages count once on each page. The pre-depth-pass figure was 113 (6.3/page). The depth pass did widen the evidence base substantially; it did not merely add lines from the existing landmark set, although the irrelevant Rodgers citation was padding and was removed.
The final 282 page-cited PMIDs and 282 bibliography PMIDs match exactly.

ClinicalTrials.gov verification¶

Every NCT identifier now present in the condition was queried individually through the live ClinicalTrials.gov v2 API: 46 distinct records, all resolving with the displayed study identity, status, and enrollment.
The newly asserted portfolio count was re-run. The live interventional query returned 337, not 286: 159 completed, 55 recruiting, 23 terminated, 22 not yet recruiting, 14 active not recruiting, 12 withdrawn, 8 enrolling by invitation, and 44 unknown. The page and statistics table were corrected; terminated plus withdrawn is 35/337, about 10%, not 12%.
The claim that the recruiting portfolio had only one phase-designated study was false. Six current phase-designated studies are now shown, including the live phase 1/2 records actually cited; the interpretation is narrowed to a thin rather than absent development pipeline.

Claim and absence corrections¶

Twenty-one targeted PubMed absence searches were run live. The principal corrections were:

Adult setting trials: DAISIES did randomize adults, but stopped at 15 participants. The absence is now an adequately recruited adult comparison, not any adult randomized comparison (PMID 39943786).
Olanzapine in younger people: a small placebo-controlled trial did exist and was null—20 participants spanning adolescence/young adulthood, 15 completers—so the categorical absence was removed (Kafantaris 2011, PMID 21663423).
Transitions: TRIANGLE randomized 371 patient-carer dyads and found no benefit on its primary carer-distress outcome. The page now reports the trial and narrows the residual question to admission/readmission outcomes (Cardi 2024, PMID 38841708).
Compulsory treatment: the page now reports the 81-versus-81 matched comparison with approximately 20 years of follow-up rather than asserting no such comparison exists; its early mortality difference attenuated over follow-up (Ward 2015, PMID 25545619).
Core outcomes: a 2024 review found 196 measures across 91 RCT reports and proposed, but did not complete, a Delphi core set. The absence is a completed consensus set, not any core-outcome work (Brieva-Toloza 2024, PMID 38389169).
Microbiome intervention: a placebo-controlled adolescent FMT protocol exists, but no outcome results were available; the claim was updated accordingly (Couturier 2026, PMID 42009386).
Neuromodulation: the account based on one older rTMS RCT was stale. A 2026 double-blind proof-of-concept study and a 13-study meta-analysis were added. The meta-analysis found no BMI benefit and only a modest depression signal, so the page does not imply established efficacy (Boehme 2026, PMID 42386712; Ewis 2026, PMID 42226264).

Other new absence claims were retained only after live searches did not locate contrary evidence, including an FBT component-dismantling trial, randomized refeeding below 60% median BMI, a fracture-primary-endpoint trial, atypical-AN longitudinal outcome/incidence cohorts, premorbid-indexed target-weight randomization, target-trial-emulation mortality work, blinded DBS-off testing, prophylactic-phosphate randomization, a validated treatment-selection biomarker, and a genome-wide-significant non-European AN discovery study.

Additional overstatement was removed or narrowed across the mortality, neurobiology, epidemiology, adolescent/adult treatment, pharmacotherapy, service-model, guidelines, red-flag, genetics, and patient-experience pages. Examples include removing unsupported superlatives, not treating cross-sectional starvation correlates as causal traits, not treating an ADOS cutoff as a confirmed autism diagnosis, and not inferring a secular mortality trend from overlapping heterogeneous syntheses. A claim that almost any admission-reducing model would be cost-effective was removed because the cited analysis did not justify it.

Status and validation¶

Promoted to status: curated: 17 of 18 wiki pages.
Held at status: draft: guidelines.md. Claude's new guideline sources resolved, but the pre-existing APA recommendation text remains unverifiable: the publisher blocks automated retrieval and PMID 36722117 has no abstract. Promotion would therefore violate the playbook.
Final checks found zero broken relative links, no PMID mismatch between pages and bibliography, and no page-local reference mismatch. The 18 wiki pages contain 2,451 lines after correction.