Consultation 01: scope and architecture of a living NMA of GLP-1 receptor agonist containing drugs

You are consulted as an independent senior methodologist (evidence synthesis, Bayesian and frequentist NMA, living systematic reviews) and senior engineer. Do NOT edit any files. Answer in plain prose with numbered points. Be concrete and critical; flag anything wrong or missing. Use American English and do not use dashes as sentence punctuation.

Project

Repo: the repository root . Owner request is in documentation/prompt-2026-10-12.md. Background report (partly stale) in documentation/deep-research-report.md.

Goal: a living network meta-analysis of GLP-1 receptor agonist containing drugs, with:

• An inventory of ALL trials and observational studies from PubMed E-utilities, Europe PMC REST, OpenAlex, plus registrations and protocols from ClinicalTrials.gov v2 and other registries (WHO ICTRP, EU CTR/CTIS, ISRCTN, ANZCTR, ChiCTR, jRCT, CTRI, DRKS, etc.). Unit of analysis is the study; records (publications, registry entries, regulatory documents) link to studies.
• Full text plus supplements plus shared code and data for every identified study (copyright: PDFs kept in a gitignored local store; committed manifest with checksums and licenses).
• Inventory of already published systematic reviews and meta-analyses (with full texts, supplements, code, data, and their protocols in PROSPERO, OSF, or inside the paper) for comparison.
• Bayesian NMA (rstan via multinma, custom Stan via cmdstanr where needed) and frequentist NMA (netmeta), all diagnostics, RoB 2 and ROBINS-I v2 / ROBINS-E, GRADE and CINeMA.
• Interactive plots via the owner's R package ggextreme (ggiraph htmlwidgets: ggnma network, ggmeta forest, ggfunnel, ggleague, cinema_judge/cinema_league/cinema_network/cinema_contribution/cinema_incoherence/cinema_clinical, ggresponder, ggsensitivity, ggmultiverse, ggkm, etc.).
• Reporting: a classic scientific paper plus a website on GitHub Pages built in JS/TS with first-class UI/UX. All code R only (site is the only JS/TS). Everything public on GitHub, machine readable.
• A full protocol (PRISMA-P, PRISMA-NMA, living review reporting guidance) written before any screening.

Probes done today (2026-10-12): PubMed tiab query for 5 agents + MeSH GLP-1RA: 17,243 hits. Europe PMC unfielded query: 44,258. OpenAlex now has usage-based credits (0.10 USD/day without a key; search= maps to full text search, so use title_and_abstract.search). ClinicalTrials.gov v2 intervention=semaglutide: 779 studies. R 4.6.0, CmdStan 2.40.0, multinma, netmeta, metafor, gemtc installed.

Questions

1. Scope. Which populations should form separate networks (type 2 diabetes; overweight or obesity without diabetes; others such as MASH, CKD, HFpEF, OSA, knee osteoarthritis, PCOS, type 1 diabetes, neurodegeneration, substance use)? Which interventions belong in the node set (mono GLP-1 RAs, dual/triple incretin agonists like tirzepatide, retatrutide, survodutide, mazdutide, CagriSema, oral small molecules like orforglipron; fixed ratio combinations with insulin like IDegLira and iGlarLixi)? Withdrawn agents (albiglutide, taspoglutide)? Comparator scope (placebo, class members, other glucose or weight lowering drugs, lifestyle, bariatric surgery)? Dose and formulation node granularity, and class effects models?
2. Outcomes hierarchy and time windows you would prespecify for each population, including efficacy, cardiovascular/renal/mortality, and harms (GI, gallbladder, pancreatitis, thyroid, NAION, suicidality, aspiration), and the effect measures.
3. Role of observational studies: inventory only, separate synthesis for rare harms, or integration into the NMA (design-adjusted or bias-adjusted)?
4. Search strategy design for each source (field tags, filters, date slicing, dedup keys) and registry coverage given that WHO ICTRP has no open API.
5. Data model for the inventory (tables, keys, link tables) and the full text retrieval pipeline (Europe PMC fullTextXML and supplementaryFiles, PMC OA, Unpaywall, OpenAlex locations, repositories such as Zenodo, Figshare, Dryad, OSF, GitHub).
6. Living mechanics: update cadence, triggers for a new version, versioning, human screening gates, and how the statistical model handles repeated looks.
7. Website: which TS framework and architecture would you choose for a static GitHub Pages site that hosts a classic paper, living dashboards, and embedded ggiraph htmlwidgets, with first-class UI/UX and accessibility? How should R outputs be exchanged with the site (JSON schemas, versioned data)?
8. Anything else a demanding methods reviewer (Cochrane, BMJ, Lancet) would expect that is missing. Keep the answer under 1,800 words.

1. Build one study inventory feeding several population-specific syntheses. Neither “ALL studies” nor universal full-text retrieval can be guaranteed. Report coverage, retrieval failures, and unresolved eligibility. Freeze umbrella protocol and initial network protocols before screening.

   Separate adult T2D from overweight/obesity without diabetes. Within T2D, distinguish treatment-naive, oral-treatment escalation, and insulin-treated settings. Give MASH, CKD, HFpEF, OSA, knee osteoarthritis, PCOS, T1D, each neurodegenerative disease, and each substance-use disorder distinct PICOs. Separate pediatric populations, pregnancy, treatment initiation, and randomized withdrawal/maintenance. Comorbidities can overlap without creating duplicate studies. Mixed populations require eligible subgroup estimates or explicit transportability justification.

   Assess effect modifiers by outcome: baseline HbA1c, BMI, cardiovascular risk, kidney function, background therapy, lifestyle intensity, and trial era. Generic placebo does not make short glycemic trials interchangeable with cardiovascular outcome trials. Connectivity does not establish transitivity. Cochrane guidance.

   Include monoagonists, tirzepatide, retatrutide, survodutide, mazdutide, orforglipron, and GLP-1-containing combinations. CagriSema combines semaglutide with an amylin analogue; calling it a dual incretin agonist is wrong. Retain albiglutide and discontinued-development taspoglutide as historical evidence, distinguished from current treatment choices. CagriSema composition.

   Define nodes by molecule, formulation/route, maintenance dose, and titration strategy. IDegLira and iGlarLixi belong in insulin-intensification networks. Distinguish comparator background treatment and lifestyle intensity. Include relevant glucose/weight drugs; surgery needs its own decision framework. Comparator-only trials require additional searches if contributing network evidence. Class and dose-response models should be secondary; combination models cannot assume additivity.

2. Prespecify critical patient outcomes separately from primary efficacy endpoints. Suggested primary targets: T2D HbA1c change at 52 weeks, accepting 40 to 60, with 26 weeks secondary; obesity percentage weight change at 68 weeks, accepting 52 to 80; MASH histologic response at 72 weeks, accepting 48 to 80. Analyze MASH resolution without fibrosis worsening separately from fibrosis improvement without MASH worsening.

   For CKD prioritize kidney failure, sustained eGFR decline, and renal death; separate eGFR slope and albuminuria. For HFpEF prioritize cardiovascular death/HF events and KCCQ at 52 weeks; OSA AHI and sleepiness at 52 weeks; osteoarthritis WOMAC pain/function at 68 weeks. For PCOS prioritize menstrual/ovulatory outcomes at 26 weeks and live birth per initiated fertility treatment. For T1D prioritize HbA1c, severe hypoglycemia, and ketoacidosis at 26 and 52 weeks. Neurodegeneration requires disease-specific cognition/function, usually 52 and 104 weeks; substance-use disorders require substance-specific abstinence/use days at 12 and 26 weeks.

   Default remaining windows: 12 weeks accepts 8 to 16; 26 accepts 20 to 32; 52 accepts 40 to 60; 104 accepts 88 to 120. Select one observation nearest each target using fixed tie rules. Test sensitivity to broad windows.

   Across populations extract mortality, three-point MACE and components, HF hospitalization, harmonized kidney outcomes, quality of life, serious adverse events, and discontinuation. Event outcomes generally use longest randomized follow-up, retaining duration. Add weight-loss responders at thresholds of at least 5%, 10%, 15%, 20%, and 25%.

   Use mean differences in natural units, including HbA1c percentage points; SMD only for equivalent constructs measured differently. Use RR for fixed-time binary outcomes, HR for time-to-event outcomes, rate ratios for recurrent events, and absolute effects against explicit baseline risks. Consider common-horizon restricted mean survival time when proportional hazards fail. Never relabel OR as RR.

   Harms need separate definitions, exposure windows, and ascertainment: GI symptoms/discontinuation, gallbladder disease, pancreatitis, medullary thyroid cancer versus other thyroid disease, NAION versus retinopathy, suicidality, procedural aspiration, hypoglycemia, and dehydration/AKI. Zero events do not establish safety. Retain suicidality surveillance while acknowledging FDA’s January 2026 review found no increased risk. FDA assessment.

3. Inventory observational studies broadly; synthesize comparative safety and effectiveness separately from randomized evidence. Prefer active-comparator, new-user designs with aligned eligibility, treatment initiation, and follow-up. Extract adjustment sets, adherence estimands, exposure lags, competing risks, immortal-time safeguards, and overlapping database populations. Case reports and spontaneous reports inform signals, not incidence.

   Keep observational integration into NMA secondary and explicitly bias-adjusted. Adding a design indicator does not remove confounding. Use RoB 2 for randomized results and ROBINS-I for observational treatment-effect results. ROBINS-E addresses exposure questions and is not interchangeable. ROBINS-I V2 remains draft and currently focuses on cohort studies; pin its exact release and choose appropriate tools for other designs. ROBINS-I V2, ROBINS-E.

4. Commission PRESS review of a versioned dictionary covering generic names, brands, development codes, regional agents, and receptor terms. Five agents are insufficient. Avoid outcome, language, publication-status, and RCT filters in master inventory. Add Embase, CENTRAL, regulatory reviews, conferences, and backward/forward citation searches. Three overlapping discovery sources cannot establish completeness.

   PubMed: MeSH plus drug/substance headings and [tiab] synonyms. Recursively split date intervals below ESearch’s 10,000-record ceiling, reconcile counts, and retrieve every PMID. Updates need entry/modification dates and overlapping backfills, not publication dates alone. NCBI limit.

   Europe PMC: TITLE_ABS terms, broader searches for registry identifiers, cursor pagination, and first-indexed/update-date harvesting. OpenAlex: prefer current search.title_and_abstract; legacy title_and_abstract.search filter syntax is deprecated. Exhaust cursors, partition long synonym lists, retain work IDs, and periodically refresh older records. Obtain free key and monitor credits. OpenAlex search, pricing.

   ClinicalTrials.gov: v2 query.intr plus query.term for aliases/codes, every study status, pageToken pagination, and LastUpdatePostDate updates. WHO offers requested web-service access despite lacking unrestricted API access. Combine negotiated access or documented exports with direct EU CTR/CTIS, ISRCTN, ANZCTR, ChiCTR, jRCT, CTRI, DRKS, and other WHO primary registries. Record coverage and manual search dates. WHO access.

   Deduplicate records by normalized DOI, PMID, PMCID, and source identifiers; uncertain title/author matches need review. Registry identifiers link studies, not automatically duplicate publications. Run dedicated review searches using systematic-review/meta-analysis terms and publication types, then trace PROSPERO, OSF, and published protocols.

5. Use normalized SQLite through R, with canonical CSV/JSON exports. Tables should cover studies, identifiers, records, record versions, study-record links, arms, regimens, outcomes, results, assessments, screening decisions, searches, search hits, artifacts, reviews, review-study links, and releases. Stable internal IDs must survive identifier corrections.

   Results need population, arm/contrast, time, estimand, denominator, uncertainty, source page/table, and extractor/adjudicator. Preserve original and standardized values. Many-to-many study-record links handle pooled reports; cohort-overlap links prevent observational double counting.

   Retrieve Europe PMC fullTextXML and supplementaryFiles, PMC OA packages, publisher supplements, Unpaywall/OpenAlex locations, and linked Zenodo, Figshare, Dryad, OSF, or GitHub deposits. Freeze repository commits and dataset versions. Track expected artifacts as retrieved, unavailable, restricted, or requested, with URL, timestamp, SHA-256, MIME type, license, and parent record. Europe PMC API.

   Apply identical retrieval to reviews and their protocols. Keep PDFs locally as requested. Gitignore does not confer reuse rights; public availability does not imply public-domain status. Restricted IPD cannot become public merely because project code is public.

6. Harvest weekly; adjudicate monthly; assess releases quarterly, with expedited review for major trials, credible safety signals, retractions, or corrections. Publish dated no-change notices. New eligible evidence, revised certainty, or corrected results should trigger releases independently of statistical significance.

   Require independent human screening, verified extraction, resolved study linkage, result-specific bias judgments, and signed analysis/report review. AI may prioritize and draft, never serve as sole second reviewer.

   Freeze immutable dated releases containing protocol, input hashes, code commit, dependencies, seeds, outputs, and changelog. Distinguish methods changes, evidence additions, and corrections.

   Bayesian refits should use original priors with cumulative data, avoiding double counting through recycled posterior priors. Cumulative confirmatory frequentist declarations require prespecified sequential error control; otherwise label intervals pointwise and avoid significance-triggered publication. Bayesian decision thresholds also need operating-characteristic calibration. PRISMA-LSR supplements PRISMA 2020. Living-review reporting.

7. Choose Astro with TypeScript and small interactive islands, statically built for GitHub Pages. R owns ingestion, analysis, tables, figures, and manuscript generation; version custom Stan source alongside R. Generate paper HTML/PDF from one manuscript source. Astro deployment.

   Exchange versioned JSON validated against committed schemas: study/node IDs, population, endpoint, time, estimand, effect scale/direction, estimate, interval type, certainty, provenance, and release ID. Keep machine precision separate from display formatting. Publish immutable release paths plus a latest manifest.

   Embed ggextreme widgets as lazily loaded, titled HTML documents with bundled dependencies. Supply accessible HTML tables, static figures, keyboard/touch equivalents, visible filter state, shareable URLs, and CSV downloads. Test WCAG 2.2 AA, mobile layouts, screen readers, and GitHub Pages base paths. Hover tooltips are insufficient. Static filtering cannot refit R models; expose precomputed analyses with clear boundaries.

8. Before launch, register protocol and amendments, involve patients and clinical specialists, justify staffing, and define retirement criteria. Prespecify treatment-policy versus on-treatment estimands, rescue medication, discontinuation, missing-data assumptions, cluster/crossover handling, and conflicting-source adjudication.

   Use multinma/rstan for standard Bayesian models, cmdstanr for validated extensions, and netmeta for frequentist sensitivity analyses. Designate primary framework before results. Align estimands without expecting identical estimates from different likelihoods or priors.

   Specify heterogeneity priors, multiarm covariance, and sparse-event methods without automatic continuity corrections. Require prior/posterior predictive checks, R-hat, effective sample sizes, Monte Carlo error, divergences, heterogeneity/prediction intervals, global/local incoherence, influence analyses, and transitivity assessments. Nonsignificant incoherence tests cannot validate sparse networks. Test known numerical examples and simulated networks; agreement between packages does not establish methodological validity.

   Define clinically important thresholds before CINeMA/GRADE judgments, avoid double downgrading, and prioritize absolute effects over ranks. Assess missing evidence and review quality using appropriate ROB-ME and ROBIS/AMSTAR 2 applications. Reviews provide comparison and discovery, not independent duplicate evidence.

   Background report is not executable specification: obsolete ClinicalTrials.gov endpoints, optional human gates, broad licensing claims, and netgrade/netlevel examples need replacement. Current netmeta documentation contains neither grading function. netmeta reference.

   No files changed. Runtime package checks failed under read-only restrictions; retrieval completeness and widget accessibility remain unverified.

1. Scope. Read "GLP-1 receptor agonist containing" literally for the inventory (every mono agonist, dual and triple incretin agonists, amylin co-formulations such as CagriSema, oral small molecules such as orforglipron, insulin fixed ratio combinations, withdrawn albiglutide and taspoglutide, development codes) and more strictly for NMA nodes. Run three primary networks: type 2 diabetes (glycemic and weight outcomes); overweight or obesity without diabetes (weight outcomes); and one cardiovascular, kidney and mortality network that pools the dedicated outcome trials across populations with population as a covariate, because those trials are too few to split. Everything else (MASH, CKD, HFpEF, OSA, knee osteoarthritis, PCOS, type 1 diabetes, neurodegeneration, substance use) is inventoried and synthesized pairwise, with an NMA only once a prespecified connectivity rule is met (at least three interventions and one closed loop). Node definition: drug by maintenance dose by route by frequency; record escalation schedules but do not make them nodes. Prespecify a drug level pooled network and a class effect model as sensitivity analyses, not primary; multinma 0.9.1 installed here exposes class_effects, class_sd and trt_class, so no custom Stan is needed for that. Treat IDegLira and iGlarLixi as nodes in the diabetes network only, and prespecify a component NMA (netmeta::netcomb) for additive structure; otherwise they create intransitive edges. Keep withdrawn agents: they close loops and inform class effects, and their removal is a sensitivity analysis. Comparators: placebo, class members, other glucose or weight lowering drugs (SGLT2 and DPP-4 inhibitors, insulin, metformin, sulfonylureas, phentermine/topiramate, naltrexone/bupropion, orlistat) and standardized lifestyle. Bariatric surgery goes to the inventory only; its populations and open label designs break transitivity. Flag: the PubMed probe named five agents while the node list has at least eight more plus development codes and brand names, and the Europe PMC probe was unfielded, so neither count means anything until the query is harmonized. Flag: estimands. Recent trials report treatment policy and trial product estimands side by side (ICH E9(R1)); they cannot be pooled naively. Prespecify treatment policy as primary and extract both.

2. Outcomes. Diabetes: primary HbA1c change and body weight change at the trial's primary timepoint closest to 26 weeks within 20 to 32 weeks, a 52 week window (40 to 60) as secondary, and longest follow-up as a third slot, each taking the timepoint closest to target; trial durations cluster at 26, 30, 40, 52 and 56 weeks, so a single rigid window would drop whole programs. Secondary: HbA1c below 7 percent, 5 and 10 percent weight loss, fasting glucose, systolic blood pressure, lipids, level 2 and severe hypoglycemia, all cause discontinuation and discontinuation for adverse events. Obesity: primary percent weight change at the primary timepoint (52 to 72 weeks; the semaglutide and tirzepatide programs used 68 and 72), responders at 5, 10, 15 and 20 percent, waist, blood pressure, physical function. Hard outcome network: three point MACE, cardiovascular death, all cause death, heart failure hospitalization, kidney composite, eGFR slope, as hazard ratios. Harms, prespecified for all populations at longest follow-up: nausea, vomiting, diarrhea, constipation, discontinuation for gastrointestinal events, cholelithiasis and cholecystitis, adjudicated acute pancreatitis, thyroid neoplasms (C-cell and medullary carcinoma separately), diabetic retinopathy complications (the brief omits this known signal), NAION, suicidal ideation and behavior, aspiration under anesthesia, gastroparesis and ileus, injection site reactions. Measures: mean difference for HbA1c and weight (pick kg or percent as primary, the other secondary); risk ratio for common binary outcomes; for rare events, odds ratio from a binomial logit model in Stan, where zero cells need no correction, and netmetabin with Mantel-Haenszel on the frequentist side, never continuity corrections; hazard ratios for time to event. Prespecify minimal important differences for cinema_clinical.

3. Observational studies. Inventory all. Synthesize separately, never inside the primary NMA, for rare and long latency harms (thyroid cancer, pancreatitis, NAION, suicidality, gallbladder disease, aspiration) and for persistence, stratified by design and giving priority to target trial emulations with active comparators over self controlled or spontaneous reporting designs. A prespecified sensitivity analysis may integrate them through bias adjusted or design stratified hierarchical models in custom Stan, with bias priors published before use. ROBINS-E for exposure designs, ROBINS-I v2 for comparative intervention designs, and GRADE applied to these bodies separately.

4. Search. PubMed: one saved query with [tiab], [nm] and [mh] terms for every agent, salt, brand and development code plus the pharmacological action heading, no study design filter in the base query, [edat] slices with overlap, retrieval through history with an API key. Europe PMC: fielded TITLE, ABSTRACT and KW terms, SRC and preprint sources included, cursorMark paging, FIRST_PDATE slices. OpenAlex: title_and_abstract.search, publication date windows, cursor paging, a key for credits. ClinicalTrials.gov v2: query.intr per agent plus query.term for codes, pageToken paging, pull resultsSection. WHO ICTRP has no API but its portal exports results per search and mirrors most of the listed registries; verify the per-export record limit and split by registry or date if needed, keep that export as the audit backstop, and query the native portals (EU CTR, CTIS, ISRCTN, ANZCTR, ChiCTR, jRCT, CTRI, DRKS) directly where they offer JSON or CSV. Add regulatory sources (FDA approval packages, EMA EPARs, PMDA), company result registers and medRxiv. Dedup keys in order: normalized registry ID, lowercased DOI, PMID, PMCID, then fuzzy study linking on first author, year, sample size and arm set. Flag: the background report's ClinicalTrials.gov api/query endpoints are retired and its gemtc and Python advice is superseded by the brief.

5. Data model. Core: study (study_id, design, population, registration status), record (record_id, source, source identifier, doi, pmid, pmcid, type: article, preprint, registry entry, regulatory document, protocol, erratum, retraction) and a study_record link table with a role (primary report, secondary report, protocol, SAP, results posting). Then arm (node, dose, route, frequency, n randomized), baseline (age, BMI, HbA1c, diabetes duration, cardiovascular history, background therapy; the brief omits this and the transitivity table needs it), outcome_definition (study definition mapped to the prespecified list and estimand), outcome_result (arm level and contrast level rows, timepoint, statistic type, value, dispersion, n analyzed, source record and page), rob (tool, domain, judgment, rationale, assessor), review (the SR and MA inventory with PROSPERO or OSF protocol IDs), file_manifest (record_id, kind: fulltext, supplement, code, data; sha256, bytes, license, source URL, retrieval date, local path), search_run and screening_decision. Retrieval order: Europe PMC fullTextXML and supplementaryFiles, the PMC OA service, Unpaywall by DOI, OpenAlex locations, ClinicalTrials.gov results, then Zenodo, Figshare, Dryad, OSF and GitHub found through data availability statements. Commit only the manifest.

6. Living mechanics. Monthly searches and screening, quarterly analysis updates, plus triggered updates when a new trial exceeds a prespecified size, adds a node, reports a hard outcome, or a regulatory action occurs. Versions are dated and tagged, each archived on Zenodo with a DOI, its own PRISMA flow and a change log. Gates: dual independent screening at both stages with a calibration set; any AI prioritization validated against a recall target on the previous version's inclusions; single extraction with second verification; dual RoB. Repeated looks: the Bayesian model is coherent under updating and carries the inference; state plainly that the frequentist analysis applies no alpha spending and its intervals are estimation only, and do not promise trial sequential analysis for a network. Prespecify what counts as a changed conclusion (direction, certainty level, rank crossing) and a stopping rule. Compute: Stan fits do not belong on hosted Actions runners; fit locally or on a self hosted runner, commit posterior summaries rather than fit objects, and let Actions only validate schemas and build the site.

7. Website. Astro with TypeScript, static output, MDX for the paper, islands for dashboards; SvelteKit static adapter is the alternative. Keep Quarto, which the .gitignore already anticipates, for the paper PDF only. Exchange: R writes versioned JSON to site/public/data/vYYYY-MM-DD/ with a latest pointer and a checksum manifest; JSON Schema files in the repo are validated in R (jsonvalidate) and in TS (zod) at build, in separate Actions jobs. Widgets: do not call graph_save(.html) per plot; it writes selfcontained = TRUE and embeds the 368 KB inst/www bundle plus ggiraph dependencies in every file. Instead call htmltools::renderTags(graph_widget(x)) to get the fragment (widget div plus the data-for JSON script) and the dependency list, copy dependencies once with htmltools::copyDependencyToDir, and mount each fragment from a TS island with HTMLWidgets.staticRender; the zero code alternative is saveWidget(selfcontained = FALSE, libdir = shared) with thin pages the island fetches or iframes. Set data-theme="dark" on the html element, because ggextreme's theme follows that attribute. Accessibility: ggiraph tooltips are hover only and the SVG has no semantics, so each widget needs an aria-label, a data table fallback, keyboard focus and prefers-reduced-motion. Flag: every NMA facing ggextreme function (ggleague, cinema_judge and all cinema_* plots) hard checks inherits(x, "netmeta"); there is no multinma or Stan path, so Bayesian results cannot reach the CINeMA plots. Add a multinma method or a generic input (relative effects, intervals, contributions, split estimates) to ggextreme; rebuilding a netmeta shaped object from posteriors is fragile for incoherence and contributions.

8. Missing. Protocol registration on PROSPERO and OSF before screening, with a deviation log. A statistical analysis plan with priors, their justification and prior predictive checks. A transitivity table with named effect modifiers and an ML-NMR plan if covariate data allow. A registry to publication discrepancy audit for outcome switching. Sponsor as a covariate and funding bias assessment. Comparison adjusted funnel plots and an inventory of registered but unpublished trials for reporting bias. Threshold analysis for robustness of conclusions to bias. PRISMA-S for search reporting. A structured comparison table against existing reviews. A statement on AI use. Flag: the brief asks for two models to be consulted at every step; a reviewer will want the human decision record, not the consultation transcripts.