{"columns":["amendment_id","date","protocol_version","section","change","rationale","before_or_after_results"],"rows":[["A000","2026-10-12","0.1.0","all","Initial draft","Not applicable","before"],["A001","2026-10-12","0.2.0","all","Revision after peer review round 1 (74 comments; see protocol/peer-review/round-1/response.md); search concepts 1.1.0","Peer review","before"],["A002","2026-10-12","0.3.0","all","Revision after peer review round 2 (39 comments; see protocol/peer-review/round-2/response.md); search concepts 1.2.0","Peer review","before"],["A003","2026-10-12","0.4.0","all","Revision after peer review round 3 (25 comments; see protocol/peer-review/round-3/response.md)","Peer review","before"],["A004","2026-10-12","0.5.0","all","Revision after peer review round 4 (16 comments; see protocol/peer-review/round-4/response.md)","Peer review","before"],["A005","2026-10-12","0.6.0","Event outcomes and harms; competing risks; absolute effects; certainty","Revision after peer review round 5 (19 comments; see protocol/peer-review/round-5/response.md)","Peer review","before"],["A006","2026-10-12","0.7.0","Event outcomes and harms; competing risks; certainty; data items","Revision after peer review round 6 (14 comments; see protocol/peer-review/round-6/response.md)","Peer review","before"],["A007","2026-10-12","0.8.0","Event outcomes and harms; competing risks; absolute effects; data items","Revision after peer review round 7 (13 comments; see protocol/peer-review/round-7/response.md)","Peer review","before"],["A008","2026-10-12","0.9.0","Competing risks; absolute effects; data items","Revision after peer review round 8 (15 comments; see protocol/peer-review/round-8/response.md)","Peer review","before"],["A009","2026-10-12","0.10.0","Competing risks; absolute effects; data items","Revision after peer review round 9 (9 comments; see protocol/peer-review/round-9/response.md)","Peer review","before"],["A010","2026-10-13","0.11.0","Competing risks; absolute effects; data items","Revision after peer review round 10 (8 comments; see protocol/peer-review/round-10/response.md)","Peer review","before"],["A011","2026-10-13","0.12.0","Competing risks; absolute effects; data items","Revision after peer review round 11 (8 comments; see protocol/peer-review/round-11/response.md)","Peer review","before"],["A012","2026-10-13","0.13.0","Competing risks; data items","Revision after peer review round 12 (5 comments; see protocol/peer-review/round-12/response.md)","Peer review","before"],["A013","2026-10-13","0.14.0","Competing risks; absolute effects; priors","Revision after peer review round 13 (8 comments; see protocol/peer-review/round-13/response.md); priors PR15 to PR17","Peer review","before"],["A014","2026-10-13","0.15.0","Competing risks; absolute effects","Revision after peer review round 14 (8 comments; see protocol/peer-review/round-14/response.md)","Peer review","before"],["A015","2026-10-13","0.16.0","Competing risks; absolute effects; priors","Revision after peer review round 15 (7 comments; see protocol/peer-review/round-15/response.md); PR02 extended to the reference arm meta-analysis","Peer review","before"],["A016","2026-10-13","0.17.0","Competing risks; absolute effects; priors","Revision after peer review round 16 (11 comments; see protocol/peer-review/round-16/response.md)","Peer review","before"],["A017","2026-10-13","0.18.0","Absolute effects; priors; amendments","Revision after peer review round 17 (9 comments; see protocol/peer-review/round-17/response.md)","Peer review","before"],["A018","2026-10-13","0.19.0","Absolute effects; amendments","Revision after peer review round 18 (6 comments; see protocol/peer-review/round-18/response.md)","Peer review","before"],["A019","2026-10-13","0.20.0","Administrative information; information sources; search strategy; outcomes; certainty; patient involvement; timeline; agents; concepts","Completion for registration: guarantor and contact; no funding and no competing interests; registration as a GitHub release instead of PROSPERO and OSF; Embase not searched with stated gaps; PRESS 2015 review and Chinese term check by two language model reviewers in place of an information specialist and a Chinese-reading team member; importance ratings and thresholds set by the guarantor with a rule for later changes; search concepts 1.3.0 and 33 new agents (31 plus two with ambiguous codes), supaglutide merged into efsubaglutide (see protocol/peer-review/press-2026-10-13/)","Owner decision and PRESS review","before"],["A020","2026-10-13","0.21.0","Administrative information; information sources; search strategy; certainty; timeline; agents; concepts","Corrections after the targeted re-check of A019 (see protocol/peer-review/press-2026-10-13/recheck-response.md): registration stated as future and as self-archiving rather than a registry entry; support and in-kind support declared; Embase omission stated as a resource decision; Chinese name check described as not verified by the team; thresholds described as undocumented operational judgments with a sensitivity analysis multiplying them by 0.5 and 2 (ratio thresholds on the log scale); releases described as tagged GitHub snapshots identified by commit hash, without a Zenodo DOI; checklist marked complete except the harvest; search concepts 1.3.1 (codes kept although not yet indexed, ambiguous codes in the registry term search) and the agent naperiglipron","PRESS re-check","before"],["A021","2026-10-13","0.22.0","Administrative information; selection process; use of automation and large language models; timeline; exclusion reasons","Full text screening of the 6,150 records with an open full text directly, without title and abstract screening, calibration or duplicate human screening, by language model reviewers (opus-5.5) under the coordinating reviewer's oversight, following data/screening/README.md and the fixed reasons in protocol/exclusion_reasons.csv; re-examination of deferred and low confidence decisions, a blind stratified sample of 300 screened again by the coordinator and by a second model, and a seeded sample of 300 kept for human audit; the four-source harvest with concepts 1.3.1 becomes the baseline search after registration; made before any screening decision (consultations: documentation of gpt-6-astra and fable-5.1 in protocol/peer-review/screening-2026-10-13/)","Owner decision: the labor that title and abstract screening and duplicate human screening save does not apply","before"],["A022","2026-10-13","0.22.1","Information sources; search strategy; screening manual","OpenAlex baseline from the public snapshot (2026-09-23 release, JSON Lines works) streamed and matched exactly on title and abstract with the concepts 1.3.1 terms, plus the API for works created after the snapshot; updates stay on the API. The screening manual (registered file data/screening/README.md) also accepts decision lines separated by | and its example line was corrected, before any decision was recorded","The API credit budget (1 USD per day) cannot cover a full harvest; the snapshot gives the same terms without that limit. The manual change is technical","before"],["A023","2026-10-14","0.23.0","Administrative information; study records; data collection; human gates; use of automation and large language models; timeline","Data extraction of every included randomized trial by two language model extractors (opus-5.5) working independently in separate sets, for every field, under data/extraction/README.md with a validating recorder; records linked to studies by registry numbers and otherwise by a language model reviewer; the coordinating reviewer, a language model session (opus-5.5) acting for the guarantor, compares the two extractions, resolves and logs every discrepancy and reports agreement; figure values indexed for later double digitization; nonrandomized studies extracted in a second pass; a seeded 10 percent sample (at least 50 trials, or all when fewer; seed and draw date recorded in data/extraction/human_audit_sample.csv when drawn) kept for field by field human verification, with re-extraction of a field in every trial when its error rate exceeds 5 percent; until the audit is done every release states that extraction is not verified by a human; changes to the extraction manual after launch are logged as amendments; made before any value was extracted for analysis (a pilot on four trials preceded launch) (consultations: documentation of gpt-6-astra and fable-5.1 in protocol/peer-review/extraction-2026-10-14/)","Owner decision: no human extractors are available, and duplicate model extraction with coordinator adjudication of every discrepancy keeps the protocol's duplicate extraction rule","before"],["A024","2026-10-14","0.23.1","Extraction manual (data/extraction/README.md)","Clarifications after the pilot report, about one hour after launch: output paged in 6,000 character pages and never redirected (overflowing output was saved to disk); estimand terms in trial, irrespective of adherence and before rescue mapped; dose left out for flexible and maximum tolerated arms; a fixed target dose that may be lowered stays fixed; run in enrichment for any run in; arm order from the registry when no main publication exists; adverse events reported as preferred terms (term), serious or nonserious only (ae_scope) and hypoglycemia definitions recorded; risk differences (rd) and event percentages accepted; no value inferred from eligibility; event driven trials use the maximum follow up covered by the counts; no coverage line for O36A, O36B and O120; worked example corrected (lifestyle standard, modified intention to treat); added after the second pilot report: a find command for short snippets, retraction of a recorded line by retract=yes, sap_available not_reported, and an invented trial as the worked example so that no study to extract is shown with answers; an arm without a printed randomized count gives n_analyzed instead, and planned or treated counts never stand in for the number randomized; dose_reported=no for a new arm whose dose is never reported; registry withdrawals for adverse events recorded under O26 only with analysis_label study_withdrawal; contrasts are arm_id versus vs_arm_id with printed signs kept; PDFs read in reading order with an optional page layout, and signs garbled by PDF fonts checked against other sources; proportion_method not_stated; Word supplements readable; factorial trials reported by factor coded by the GLP-1 factor; any cancer (O39) only from a printed total, with thyroid and pancreatic cancer terms recorded separately; registry responder yes and no counts recorded as events and nonevents with the printed analyzed n; registry on treatment period analyses are while on treatment; adjudication sets C01 to C20 decide every difference by accepting, correcting or dropping a line, with label only differences paired for comparison; study and arm lines carry every field the sources support, whatever the node; arms and studies with no printed count are recorded with counts_reported=no instead of a computed count; adjudication rulings on registry event scope, end dates, zero episode counts, quality of life totals, trials with pooled reports only, population shares printed as counts and placebo arm fields; percentages without a stated method are not_stated; unlabeled registry contrasts are not_stated with an analysis label; whole trial baseline only when no arm value is printed; ratios never inferred","Pilot extractor report on four trials (SURMOUNT-3, SUSTAIN 7, LixiLan JP-O1, SELECT); running extractors were told of the changes and recorded lines stay valid under them","before"],["A025","2026-10-16","0.23.2","Information sources; search strategy; study records (inventory)","OpenAlex baseline from the public snapshot of 2026-10-14 alone, without the API top up for works created after the snapshot; OpenAlex works typed peer review, dataset, supplementary materials, paratext, software, reference entry, standard, book review or data paper left out of the inventory; the exact title, year and first author match joins a correction or retraction notice only to another notice; a record split from an earlier merge keeps its ID for the part with the lowest PMID and the other part gets a new ID, logged in data/inventory/record_id_splits.csv. The weekly OpenAlex update is unchanged here; its date filter also needs a paid plan and its replacement awaits the guarantor.","The OpenAlex API now requires a paid plan for the from_created_date and from_updated_date filters; the snapshot date equals the search date, so nothing is lost for the baseline. OpenAlex lists some correction notices under the title of the corrected article, which merged notices with articles (for example an included article merged into its correction); its peer review reports and datasets are not reports of studies. Made before screening any OpenAlex record.","before"],["A026","2026-10-27","0.23.3","Data synthesis; living review methods (versions); reporting and dissemination (website)","Interim network meta-analyses of the randomized trials adjudicated so far are published on the website at the guarantor's request before extraction is complete, labeled interim, without risk of bias, CINeMA or GRADE, and refitted as adjudication proceeds (R/nma.R, pipeline/10_nma.R). In these interim analyses: cluster and crossover trials, standardized mean differences (O24B, O24C), rate ratios (O04B), complete status counts of time to event outcomes, meta-regression, incoherence models, dose response, class and component models, absolute effects and contribution matrices are deferred; adverse event counts in the recorded harms analysis are pooled whatever estimand the trial states; results whose trials state no estimand form their own networks beside treatment policy and hypothetical networks; a trial contributes arm level data when they exist for at least two nodes, otherwise one reported contrast, so a multi arm trial reported only as contrasts contributes one contrast; a percentage of responders is converted to a count when its method is observed or not stated; drug level networks (doses pooled) are fitted beside dose level networks; the frequentist check uses netmeta on all within trial contrasts; the interim web pages draw the network graph and forest plot from the published tables, and the ggextreme figures follow with the full analysis.","Owner request to publish the analyses as data accrue. The full analysis of the protocol follows complete extraction, risk of bias and certainty assessment and supersedes these interim results. The rules were written before any network was fitted, apart from one test fit used to check the code.","after"],["A027","2026-10-29","0.23.4","Data synthesis (transitivity); appendix C (effect modifiers)","Each network's comparability is assessed across designs in three domains: eligibility (entry ranges for HbA1c, BMI and age read from the eligibility summaries), outcome definition and assessment (time point, analysis set, definition, instrument, adjudication, proportion method, intercurrent event strategies, blinding and funding), and the baseline values of treatment effect modifiers and prognostic factors listed with their roles, outcomes, imbalance thresholds, literature and rationale in protocol/tem_pf.csv, which extends effect_modifiers.csv. A design whose mean differs from another's by more than the threshold, a factor missing from more than 20 percent of trials, or a classification confounded with design raises a flag, and networks whose heterogeneity exceeds the outcome threshold list their flags as candidate explanations. The trials' subgroup results are summarized per factor with a test of subgroup heterogeneity.","At the guarantor's request, so that every network meta-analysis states whether its trials are comparable and which differences could bias it or explain heterogeneity; the effect modifier list is grounded in published analyses and the trials' own subgroup results.","after interim results (A026); assessment only, no estimate changes"]]}