Skip to content

Curation log — generalized anxiety disorder

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

2026-09-02 — Full build (Claude). Awaiting independent audit.

What was built. All 19 planned wiki pages, the complete literature layer, a rewritten OPEN-QUESTIONS.md, and an updated INDEX.md. Every page is status: draft; nothing was promoted. The build engine was Claude — the auditor must be a different engine, per CLAUDE.md.

Artefact Result
Wiki pages 19 of 19, 126–171 lines each, 24–41 verified citations each (~2,830 lines total)
Distinct verified papers 387, every PMID retrieved live from PubMed E-utilities during this session
Trial registry records 36 NCT IDs retrieved live from the ClinicalTrials.gov v2 API
Non-journal sources 21 organisation/government/WHO pages fetched by HTTP with status codes recorded; 6 further organisations recorded as unverifiable (HTTP 403/timeout)
literature/BIBLIOGRAPHY.md 387 records grouped by section, with tags and cited-by lists
literature/notes/ 6 landmark notes (Spitzer 2006, Ruscio 2017, Ruscio 2024, Slee 2019, Kendler 1992, Bschor 2024)
literature/guidelines/REGISTRY.md 26 documents with superseded-by chains, 7 disagreements, 5 structural gaps, 7-item watch list
literature/statistics/STATISTICS.md ~13 topic tables, every figure carrying source/year/population/method, plus 13 stated conflicts
literature/patient-voice/ 4 files; 24 organisations catalogued (1 closed, 6 unverifiable); 7 themes each on ≥2 sources; 8 stated coverage limits
OPEN-QUESTIONS.md 40 tiered questions with stable IDs plus an 18-row "dots not yet connected" table

Anchor re-verification. All nine seed anchors were re-fetched live and all nine resolved with matching first author, year and journal. None was discarded. Carpenter 2018 (PMID 29451967) is used only with its pooled scope stated explicitly, as the seed flagged.

Citation integrity. Every PMID in every file was cross-checked programmatically against the session's own fetch store: 387/387 present, 0 fabricated. Every numbered reference line was additionally checked for year and journal match against the retrieved record: 0 mismatches. Roughly a dozen page/volume errors caught by that check were corrected during the build. All markdown tables were checked for column-count consistency: 0 malformed rows.

Searches run (live PubMed E-utilities, 2026-09-02). Approximately 70 distinct queries across: diagnosis and DSM-5/ICD-11 classification; reliability and structured interviews; taxometrics and latent structure; HiTOP/RDoC; network analysis of anxiety–depression symptoms; epidemiology (WMH, NCS-R, NESARC, India NMHS, Bangladesh, Singapore, urban China, Iran); GBD burden; age of onset; treatment gap; twin and molecular genetics; worry models (avoidance, intolerance of uncertainty, metacognitive, emotion dysregulation, contrast avoidance); neuroimaging, HRV, ERP, executive function, inflammation; GAD-7/GAD-2/HADS-A/PSWQ/HAM-A/PHQ-4 psychometrics and cultural bias; USPSTF and WPSI screening; CBT meta-analyses and networks, applied relaxation, low-intensity CBT, delivery formats, mechanisms of change; metacognitive therapy, ACT, MBSR, psychodynamic therapy, emotion regulation therapy, motivational interviewing augmentation; SSRIs/SNRIs (Cochrane, networks, individual trials, dose-response, cost-effectiveness); pregabalin, benzodiazepines, quetiapine, buspirone, agomelatine, silexan, kava, cannabidiol, vortioxetine; internet CBT, apps, conversational agents, adherence; treatment resistance, augmentation, rTMS/cTBS/tDCS; course, remission, relapse prevention, mortality; children/adolescents (CAMS, CAMELS, escitalopram, duloxetine), older adults, perinatal; primary care recognition, collaborative care, guideline concordance, substance-use comorbidity, perceived helpfulness; guidelines (WFSBP, NICE, BAP, German S3, Canadian, CANMAT, Brazilian, AACAP, USPSTF, nutraceuticals); placebo response, trial recruitment, core outcome sets; safety (falls, hyponatraemia, bleeding, discontinuation, paediatric suicidality, benzodiazepine deprescribing, gabapentinoid misuse, thyroid and pheochromocytoma mimics, caffeine); qualitative and lived-experience literature.

ClinicalTrials.gov queries (v2 API, 2026-09-02). Recruiting GAD studies; completed phase 2/3 GAD studies sorted by start date; targeted queries for MM120, ITI-1284, ABBV-932 and psilocybin; and individual record lookups to verify the five NCT IDs cited from published trials (NCT03522844, NCT00426426, NCT02818751, NCT03924323, NCT00052078 — all confirmed COMPLETED with matching enrolment). A search-hygiene note for future sweeps: "GAD" also abbreviates glutamic acid decarboxylase, so naive registry queries return type 1 diabetes GAD-alum immunotherapy trials (e.g. NCT01785108, NCT05351879). These were excluded explicitly and the caveat is written into clinical-trials-landscape.md.

Scoping decisions taken during the build.

  1. Pooling is labelled at the point of use, not once. Every effect size taken from a source that pools anxiety disorders (Carpenter 2018, Hofmann 2008, Bandelow 2015, Pauley 2023, Hoge 2023, Jakubovski 2019, Seppala 2018, Henssler 2024, Kalfas 2025, Haller 2021, Hyde 2022, Black 2019, Wagner 2026's non-GAD strata) carries an inline "pooled" marker. GBD figures are flagged as anxiety-disorder figures because GBD does not model GAD.
  2. Comorbid depression is cross-linked, never absorbed. The GAD–MDD genetic correlation and the pure-vs-comorbid dispute are treated as content on the-diagnostic-boundary.md and comorbidity-and-primary-care.md, with depression curated elsewhere.
  3. PTSD is cross-linked as a neighbour from diagnosis-and-classification.md, epidemiology-and-burden.md, patient-voice/README.md and overview.md; nothing about PTSD is restated here.
  4. Four different prevalence frames are separated explicitly — interview-diagnosed GAD, ICD-11 algorithm self-report, administrative claims, and screening-scale symptoms — because conflating them is the commonest error in secondary reporting of GAD statistics.
  5. Conflicting syntheses are presented as conflicts, not averaged. Digital vs face-to-face CBT; benzodiazepine tolerability within-class vs across-class; quetiapine augmentation; clinic vs community prognosis; responder vs continuous analyses; drug vs psychotherapy effect sizes across different control types.

Things that could not be verified, carried as [unverified]. GAD-specific κ from the DSM-5 field trials (Regier 2013 reports only the distribution across 15 adult diagnoses); pregabalin's US regulatory status for GAD; agomelatine hepatic-monitoring requirements; a GAD-specific HAM-A MCID; a systematic audit of GAD trial reporting; DIF studies of the GAD-7 outside one US undergraduate sample; treatment-response data for criteria-excluded (non-excessive, <6-month, 2-symptom) cases; a German S3 revision later than 2013/2014; the 2002 WFSBP first edition (referenced by its 2008 successor but not separately retrieved); published results for four completed GAD phase 2/3 programmes (troriluzole NCT03829241, ENX-102 NCT05749055, SEP-363856 NCT05729373, extended-release lorazepam NCT02305797). Each is flagged in place and dated.

Positively-stated absences, re-searched and dated 2026-09-02. No current APA practice guideline for GAD (the 1998 anxiety guideline covers panic disorder, PMID 9585731). No GAD-specific GBD estimate. No screening trial showing outcome benefit. No comorbidity-stratified treatment estimate. No fixed-versus-indefinite continuation trial. No dismantling trial of collaborative care. No preference-based drug-versus-therapy design in GAD. No adult GAD core outcome set (one is in development for paediatric anxiety, PMID 37244653). No GAD-specific patient priority-setting partnership. Only one located qualitative study of people diagnosed with GAD (Hurtado 2020, PMID 31908140).

Notable findings that reshaped the build. (i) GAD's placebo pre-post effect is d_av 1.23, second of nine psychiatric disorders (Bschor 2024, PMID 38809560) — this became the organising fact of clinical-trials-landscape.md and is cross-referenced from every treatment page. (ii) GAD carries 48% excess all-cause and 55% excess natural-cause mortality (Wagner 2026, PMID 42136520), which moved mortality from a footnote to a section. (iii) 70% of treated patients find treatment helpful but only ~30% would persist through enough encounters to reach the modelled ceiling (Stein 2021, PMID 34372811) — this reframes the treatment gap as partly a persistence problem. (iv) A commissioned UK RCT of drug versus therapy was terminated because three-quarters of eligible patients refused randomisation to medication (Kalpakidou 2019, PMID 31126337), which makes the central comparative-effectiveness question possibly unanswerable by randomisation. (v) Four independent literatures converge on uncontrollability rather than excessiveness as GAD's discriminating feature, and no revision proposal has been built on that convergence (new OQ-36). (vi) Anxiety Canada's website now carries a closure notice — a national patient organisation has ceased operations.

Follow-ups for the auditor.

  1. Re-fetch all 387 records and re-check every claim→PMID pair; the reference lines were machine-checked for year and journal but not for whether the claim matches the abstract.
  2. Re-run the searches behind every asserted absence listed above and re-date them; stale absences are the commonest real error.
  3. Resolve [unverified] flags where possible — in particular the DSM-5 field-trial GAD κ (needs full text), pregabalin's regulatory status, and whether a German S3 revision exists outside PubMed.
  4. Check the four numbers most likely to be mis-transcribed: Cardoner 2025's 12-week HAM-A MD (reported as +0.99 where every other timepoint is negative — recorded as published, unreconciled); Stein 2021's agomelatine HAM-A difference of 6.30, roughly double the reference network's best estimate; Gundugurti 2024's cannabidiol GAD-7 difference of −7.02 with an implausibly narrow interval for n=178; and Duan 2025's rTMS response OR of 291.40 (95% CI 13.08–6490.21).
  5. Verify the perinatal GAD-7 cut-off claim (13 rather than 10) against Simpson 2014 full text before any clinical use is implied.
  6. Re-verify the patient-voice organisation directory: URLs decay, and one organisation in it has already closed. Attempt the six unverifiable sites again with a different approach.
  7. Promote to curated only pages where every citation resolves and matches with no unresolved substantive issue. Record both engines by name in the audit entry.

2026-09-02 — Seeded (Claude)

Created the scaffold: INDEX.md with a 19-page plan (16 topical + 3 standing house pages) and a 9-record anchor table, a seed OPEN-QUESTIONS.md, this log, and the empty wiki/ and literature/ layers. No wiki pages written.

Searches run (live PubMed E-utilities, 2026-09-02). Scoping: generalized anxiety disorder 17,415; anxiety disorders[MeSH Major Topic] 69,338. Per-page topic counts in INDEX.md, including epidemiology 6,253, comorbid depression 2,809, GAD-7 2,202, CBT 1,640, course/relapse/remission 1,609, worry/intolerance of uncertainty 1,501, digital delivery 1,132, children and older adults 915, treatment-resistant 291, SSRI/SNRI 175, primary-care detection 125, pregabalin/benzodiazepine 67.

Query-hygiene note. Five initial queries were over-specified and returned single- or double-digit counts that would have wrongly suggested empty pages (neurobiology worry returned 9; long-term course relapse returned 18; digital internet-delivered therapy returned 20). Re-run with proper alternation they return 1,501, 1,609 and 1,132. Affected counts are marked † in INDEX.md. This is the mirror of the PTSD problem in the same session: loose Booleans inflate, over-specified ANDs deflate, and both distort a page plan.

Anchor records resolved live: 9, listed in INDEX.md.

Scoping decisions. The gap between 17,415 GAD records and 69,338 anxiety-disorder records is the scoping problem: most apparent GAD literature is anxiety-disorder literature or GAD-with-depression literature. Three borders written into INDEX.md — the condition owns GAD and not panic, social anxiety, phobias or OCD; comorbid depression is cross-linked to depression rather than absorbed; PTSD is a neighbour, seeded in the same session, and cross-links. Where a trial or guideline pools anxiety disorders, the page must say so explicitly, because a pooled anxiety effect size is not a GAD effect size. One anchor is flagged in the table for exactly this reason: Carpenter 2018 (PMID 29451967) covers anxiety and related disorders, so any GAD-specific figure must come from its GAD stratum or a GAD-specific source.

The diagnostic boundary is scoped as content. Whether excessive worry should remain a required criterion is a live published dispute — the same group that produced the cross-national epidemiology has argued for removing it (PMID 28297020; PMID 39364896). It gets a dedicated page and is left unresolved with both sides sourced.

Flagged for the build pass. No anchor yet for NICE/APA/WFSBP guidelines, metacognitive therapy, internet-delivered CBT trials, maintenance and relapse-prevention trials, perinatal and paediatric GAD, or patient-organisation material.

Follow-ups. Build with tools/build-condition.sh generalized-anxiety-disorder; auditor must differ from the writer.

2026-09-02 — independent audit

Authorship and independence. Original pages and literature layer: Claude. Independent auditor and corrective editor: Codex. Codex did not participate in the original authorship.

Coverage checked. Audited all 19 wiki pages and every artifact in literature/, plus INDEX.md and OPEN-QUESTIONS.md. Re-fetched 387 original distinct PMIDs by live PubMed E-utilities calls, then live-fetched 12 additional PMIDs discovered during correction; the final corpus contains 399 distinct PMIDs, all resolving to the claimed record. Re-fetched all 38 distinct NCT IDs individually from the live ClinicalTrials.gov v2 API. Checked 2,003 original claim–citation occurrences across narrative artifacts, including 1,193 inline PMID occurrences in wiki prose, and rechecked that every wiki-body PMID appears in that page's reference list. Checked 27 external URLs live: 20 returned HTTP 200, six returned bot/Cloudflare 403 responses, and one failed to connect; the inaccessible entries remain explicitly labelled rather than treated as verified content.

Errors found and fixed. The corrections below are exhaustive for this audit pass.

Area Error in Claude-authored material Correction made
German guideline The 2014 S3 document was labelled current and a later revision “not located.” Added the first revision (Bandelow 2022, PMID 34609587), which reviewed 92 additional RCTs; updated the registry, synthesis, bibliography and watch list.
Treatment resistance Repeated claims said no consensus definition or staging model existed. Added the 36-expert Delphi consensus with 14 recommendations and a potential trans-anxiety staging model (Domschke 2024, PMID 38214637); retained the narrower gap that GAD-specific validation and adoption remain limited.
Qualitative evidence The patient layer repeatedly claimed exactly one GAD-specific qualitative study. A fresh PubMed search located several. Added GAD-specific work on primary-care implementation, perinatal behaviours, motivational-interviewing experience and dietary-trial participation (PMIDs: 33218311, 35716017, 21644188, 41587136), while retaining the evidence-base limitation.
Patient preference The corpus claimed no preference-based GAD comparison existed. Added the quasi-experimental patient-choice comparison of e-CBT, medication and combination (Stephenson 2023, PMID 38125282); reframed the gap as lack of a design supporting causal comparative inference.
GAD-7 invariance Sex-based measurement invariance was described as unstudied. Added partial strong sex/time/language invariance in young adults (PMID 39880312, already present) and a 165,872-patient sex-invariance study (Saunders 2023, PMID 37118684); narrowed the remaining gap to cross-cultural and diagnosed-GAD generalisability.
Core outcomes The corpus said adult GAD had no core-outcome work. Added the patient-engagement study developing candidate perinatal-GAD outcomes (Stallwood 2026, PMID 42003419); retained the dated absence of a general-adult GAD core set.
Perinatal trials A 2020 review's “no treatment trials” finding was repeated as a current universal absence. A fresh search found a 2024 secondary analysis of a systems-level cluster RCT measuring perinatal GAD symptoms (Zimmermann 2024, PMID 38992743). The remaining dated gap is specifically perinatal GAD pharmacotherapy RCTs.
Fear generalisation Lissek's n=22 signal was described as never replicated. Added the later 28-GAD/30-control study that found no difference on most indices (Tinoco-González 2015, PMID 26459843); the evidence is now described as small and conflicting.
Hydroxyzine Hydroxyzine was left wholly unverified. Added the 334-participant, three-month double-blind trial: HAM-A change −12.16 versus −9.64 with placebo, p=0.019, with response and remission also favouring hydroxyzine (Llorca 2002, PMID 12444816).
Kava An impossible source value was repeated as a GAD null: MD −0.17 with 95% CrI −2.55 to −1.97. Explicitly identified the PubMed abstract's internal inconsistency and removed the directional GAD conclusion pending a corrected source value.
Trial scope NCT07233278 was presented as a GAD rTMS pipeline trial. Live record showed recurrent pregnancy loss with anxiety, not a GAD-defined sample; removed it from the active GAD table and documented its exclusion.
Trial population NCT04895995 was presented as a general GAD digital-CBT programme. Qualified it as a cardiovascular-disease population selected for generalized-anxiety symptoms, not diagnosed-GAD-only.
Trial reporting Four completed programmes were labelled unpublished without distinguishing registry results. NCT-ID PubMed searches found no journal articles; ClinicalTrials.gov has posted results for NCT03829241 and NCT02305797. The pages now distinguish registry results from journal publication.
Regulatory claims Pregabalin licensing, agomelatine monitoring and “quetiapine off-label everywhere” were asserted without retrieved regulatory sources. Removed jurisdiction-wide assertions; efficacy and safety statements now remain within the verified literature, and regulatory status is described as jurisdiction-dependent or outside scope.
Mortality Similar natural-cause and all-cause RRs were used to infer that excess deaths were dominated by physical illness. Removed the unsupported decomposition; overlapping outcome categories and intervals cannot establish causal composition.
Cross-treatment rankings Collaborative-care NNTs, drug mean differences, psychotherapy SMDs and rTMS effects were repeatedly compared as if exchangeable. Removed “best intervention,” “three times more effective” and similar rankings; retained effects with their own population, comparator and endpoint.
Causal language Placebo response was said to explain trial failures; statistical mediation was called causal; persistent thyroid symptoms and later depressive symptoms were treated as diagnostic proof. Reframed each as context, association or differential-diagnosis evidence; removed unsupported causal certainty.
Paediatric follow-up “Treatment type does not matter” and modality interchangeability were inferred from remission analyses. Added that treatment assignment predicted some functional trajectories and reframed the question around joint remission and functional outcomes.
Medical safety A source on SSRI-associated bleeding was presented as evidence for an SSRI–NSAID/antiplatelet interaction; age 40–50 and universal thyroid-testing implications were added without support. Restricted the bleeding claim to SSRI exposure, removed the unsupported age threshold, and changed testing language to feature-guided differential assessment.
Diagnostic reliability The unavailable GAD-specific DSM-5 field-trial kappa was left as a bare unresolved marker. Stated positively that the PubMed abstract reports only the cross-diagnosis distribution and that no GAD-specific value is inferred.
Evidence gaps Bare [unverified] markers covered HAM-A MCID, fixed-versus-indefinite continuation, collaborative-care dismantling, older-adult escitalopram replication, CBT by depressive comorbidity, criteria-excluded treatment response, medical-work-up yield and digital equity. Re-ran targeted PubMed searches and converted each into a positively stated, dated evidence-gap sentence. The MM120 2.5-point HAM-A threshold is correctly labelled a prespecified assumption rather than an anchor-based MCID validation.
Attribution and references Compound surnames and several body citations/reference lists were inconsistent; three body PMIDs were missing from page references. Normalised van Dis, Ter Kuile, Di Nardo, Trenoska Basile, Af Winklerfelt Hammarberg and Ishrat Husain; added missing page references for PMIDs 40130828, 40906494 and 19351943; final body/reference coverage check found zero omissions.
Landmark-note rhetoric Notes described one review as the first at scale, placebo as explaining failures, and GAD-7 as the route by which most people are identified. Removed unverifiable priority and prevalence claims and narrowed interpretation to what the cited studies establish.

Remaining flags and limitations. No bare [unverified] marker remains in current wiki or literature prose. The kava credible interval remains an explicitly documented error in the source abstract, not a repaired number. The GAD-specific DSM-5 field-trial kappa remains unavailable from the PubMed abstract and is not asserted. Six organisation pages returned access-control responses and one failed to connect during the live HTTP pass; their access status remains recorded. Dated negative searches are evidence gaps, not proof of universal absence, and should be rerun at the next curation date.

Promotion result. All 19 of 19 wiki pages now have status: curated. Every current PMID and NCT ID resolved live, every wiki-body PMID is represented in its page references, all relative links resolve, and no unresolved substantive citation mismatch remains. CONDITIONS-ROADMAP.md was intentionally not modified.