Reproducibility archive for the GBD 2023 China acute myeloid leukaemia decision model (transplant capacity and treatment accessibility)

This archive is the reproducibility set for the manuscript "Expanding treatment access would avert more deaths than expanding transplant capacity in acute myeloid leukemia in China, and population aging widens that gap: a microsimulation study". It contains the analysis plan as executed, the statistical analysis plan, the model source code, every input file the model reads at run time, every output table the manuscript reports, the scripts that produced the figures, and the literature search record. The five figure files and the graphical abstract can be rebuilt from this deposit alone; all six were rebuilt from the unpacked archive and came out byte-identical to the delivered files. Contents by directory: 00_主稿/ — the manuscript as submitted (Markdown, Word and PDF), the Highlights, the cover letter, the graphical abstract, the online supplementary appendix (Markdown, Word and PDF) and the PRISMA flow diagram. 01_方案与计划/ — the final analysis plan as executed, and the script that produced its last revision. 02_引擎与参数/ — the statistical analysis plan, the acceptance-gate execution record, the record of the non-blocking registry comparison, the supplementary demographic-input check, the parameter extraction cards, the key parameter tables and the transplant-capacity allocation rules. The model source itself sits in 06_repro_inputs/R_work/. 03_主结果/ — every output table cited in the manuscript and the online appendix: scenario results, calibration check, probabilistic sensitivity analysis (2,000 draws), structural sensitivity analyses, data-quality-assurance sweeps, cause-of-death decomposition, threshold scan, acceptance-gate tables and the gate execution logs, the tables of the non-blocking registry comparison, and the tables of the supplementary demographic-input check. 04_图/ — the five figures as submitted (PNG, 7,480 px, 1,000 dpi, RGB), the graphical-abstract script and its layout template, and the figure and consistency checkers. 05_检索留痕/ — the literature search record: 127 structured search queries in total (48 in the first seven rounds plus 79 in round 8), of which 18 went through the Europe PMC machine check and title/abstract screening, per-engine hit counts, per-record screening decisions, the query-semantics audit and the forward-citation coverage audit. The same directory also carries the AI use log (AI使用记录_AI-use-log.md), which records the tool and its version, each category of use (text preparation, code drafting, figure scripts, pixel digitisation of the calibration curves, and two input-source references), the manual verification trail, and the change made on 30 October 2026. The raw hit sets (~150 MB) are not copied into the archive; a per-file SHA-256 inventory of them is provided instead. 06_repro_inputs/ — the input files the model reads at run time, with the directory and file names the source code expects, plus the scripts that generate the forecast case series and that execute the acceptance gates. 07_figure_build/ — a self-contained kit for rebuilding the five figures: the figure scripts, the six output tables they read, the palette-to-RGB conversion step, and the instructions. How to re-run the model: set the environment variable G2_BASE to the directory 06_repro_inputs and run Rscript 06_repro_inputs/R_work/run_v38hc.R. Requirements: R 4.6.1; no other package is needed for the core model. The script writes its output to $G2_OUT_DIR. All random draws use fixed seeds, so the probabilistic analysis is reproducible. This was verified end to end by unpacking this archive into a clean directory and running the model there: the reported scenario tables are reproduced byte for byte (the six result tables of the reported run; the run log and the timing record differ only in their elapsed-time and timestamp fields). The probabilistic run as deposited is reproduced separately, with the draw count raised through the environment variable the model already reads: G2_PSA_N=2000 with G2_PSA_DESIGN=mc. With that setting the 2,000-draw sample table is reproduced byte for byte as well. Note that the probabilistic run predates the half-cycle correction of the life-year bookkeeping, so its life-year columns carry the earlier values; the deaths-averted columns, which are what the probabilistic analysis reports, are identical either way (see section 3 of the README). Integrity: SHA256SUMS.txt lists a SHA-256 checksum for every file in the archive except itself and FREEZE_MANIFEST.json. Verify with sha256sum -c SHA256SUMS.txt. The manifest also records the MD5 checksum and line count of the model source. A note on the acceptance-gate readings: the gate-execution record in 02_引擎与参数/ carries two dates. The readings of 24 October 2026 were taken before a recalibration of the model carried out later the same day; the record of 27 October 2026 supersedes them and gives the readings produced by the model source as deposited (MD5 9f0da612dbe372d609276042d8834a55, 1,453 lines). The passage between the two sets is attributable to that recalibration, not to the later half-cycle correction of the life-year bookkeeping, which left the death-side output unchanged digit for digit. Every gate reading quoted in the manuscript and in the online appendix is the 27 October value. The acceptance criteria themselves were not altered at any point. A fourth, non-blocking comparison is also deposited: the agreement of modeled 5-year survival with the same registry, on direction and magnitude only. Its memorandum is in 02_引擎与参数/ and its output tables are in 03_主结果/ (prefix G2_out_ncr__). The criteria were fixed before execution. The outcome is that direction agreed but magnitude was not met (no modeled value fell inside the registry's 95% confidence interval); it changes no result and reported quantity and cannot support a claim of external validation. It draws on the same registry as the calibration targets and is an internal consistency check. The analysis plan as amended on 27 October 2026 cancelled the submission-strategy constraint that had tied a failed comparison to the journal tier; that amendment was made after the result was known, is recorded in the plan itself, and changed no criterion, no threshold and none of the findings. The memorandum also records why re-targeting the calibration to the tabulated registry values was not adopted. The analysis plan deposited in 01_方案与计划/ is the final version as amended on 27 October 2026. That amendment cancelled one clause only: the submission-strategy constraint under which a failed non-blocking registry comparison would cap the journal tier. The cancellation was made after the comparison result was known and is disclosed as a post-hoc amendment in the plan itself (annex point 10 and freeze order 11). No criterion, no threshold and none of the reported findings were changed by it, and the comparison still cannot support a claim of external validation. A fifth, supplementary check on the demographic input is also deposited. The model has drawn its projected population and its background life tables from the UN World Population Prospects 2024 medium variant since it was first built, but that source was never cited in the manuscript or the plan; this deposit records the citation and, separately, one directional check of the demographic input against the official age composition published by the National Bureau of Statistics of China. Because the official communique does not publish age-by-sex population, the check applies a translation anchoring that preserves total population and the within-band shape; under it the share of new cases aged 60 or more over the reporting window rises from 48.5% to 50.4%, widening the margin to the ranking-flip threshold from 21.2 to 23.1 percentage points. The ranking is unchanged and, on the official age composition, more robust. Its output tables are in 03_主结果/ (prefix G2_out_officialage__) and the probe that produced them is in 02_引擎与参数/. This check inspects one input, not a model output; it does not re-run the model, changes no reported quantity, and cannot support a claim of external validation. The analysis plan as amended on 27 October 2026 records both the citation and the check, discloses that the amendment was made after the result was known, and states that the model source and every locked number are unchanged. Absolute values reported in the manuscript are mechanistic quantities: arithmetic consequences of the model under the stated inputs and settings, not predictions of the real world. Only the ranking, the relative magnitudes and the conditional boundaries are stated as findings. No patient-level identifiable information is present in this archive.