BRAIN STRUCTURE AND FUNCTION — VERSION 1.8, FINAL AUDITED RELEASE
Read README_BSF_v1_8.txt for figure commands and FINAL_AUDIT_2026-10-27.txt
for the successful end-to-end verification. The scientific analyses below
remain version 1.7. A S900 CSV newline-handling fix and two synthesis provenance
hash corrections are explicitly recorded; no reported statistical value or
analysis specification changed. The original AOMIC files are preserved.

SUPPLEMENTARY CODE 1
Brain Tissue Proportionality and General Cognitive Ability Across Three Adult Cohorts
Version 1.7 | 27 October 2026

CONTENTS AND ANALYSIS STATUS

HCP_DLBS contains the original primary model definitions, cohort scoring and
specified sensitivities, with full-precision aggregate outputs. The companion
HCP_DLBS_extensions contains supplementary size-covariate comparisons, DLBS
HC3 and bootstrap inference, six-part anatomical contrasts, and the reproduced
official S900 unrelated-participant sensitivity.

AOMIC retains the original analysis plan, internal receipt, source manifest,
primary and sensitivity code, separate numerical verification and aggregate
outputs byte-for-byte. The internal plan preceded the first AOMIC association
analysis. It was not a public preregistration. The original timestamps and
receipt are preserved; no new retrospective registration has been created.

AOMIC_extensions contains the additional volume, pairwise-balance, cognitive
domain, cerebellar-decomposition and size-covariate analyses. These extensions,
the new DLBS uncertainty checks, and the two-cohort synthesis were performed
after primary results were available. They are exploratory supplementary work.

Synthesis reports DLBS and AOMIC inverse-variance fixed-effect and REML
random-effects summaries with conventional and HC3 standard errors. Modified
Knapp-Hartung intervals with one degree of freedom illustrate the substantial
uncertainty in inference across only two cohorts. The cognitive batteries and
cohort-specific standardizations differ; pooling is a descriptive synthesis,
not a third independent confirmation or a universal biological effect.

Figures regenerates the four-panel Figure 3 from model-adjusted participant
residuals and exact aggregate coefficients. Both focal coordinate and g are
residualized on age, sex, log total analyzed tissue volume and the other three
ILR coordinates. Cohort slopes are separate. HCP intervals and independent-
observation confidence bands are omitted.

DATA ACCESS AND DISTRIBUTION

HCP-YA data are subject to HCP open-access terms. DLBS original workbooks must
be obtained under the dataset's applicable conditions. This archive does not
redistribute any HCP/DLBS participant-level source data or residual data.
The seven input filenames required for HCP/DLBS are listed in run_all.py.
Source filenames, hashes and software versions are recorded in the outputs.

The official S900 public unrelated list is downloaded only during reproduction
and must match its pinned SHA256. It contains 339 unique public IDs; 337 meet
this study's complete-case criteria. The S900 sensitivity remains null and is
fully reported. It overlaps the HCP discovery cohort and is not an independent
replication. It does not supply the missing family structure for the full HCP
sample. Its source URL and checksum are in fetch_s900.py and the manifests.

AOMIC ID1000 is public CC0 (OpenNeuro ds003097 v1.2.1;
https://doi.org/10.18112/openneuro.ds003097.v1.2.1). The original minimal AOMIC
analysis dataset is included. Its complete public inputs are fetched from a
fixed source revision and checked against unchanged hashes on reproduction.
Citation: Snoek et al. (2021), Scientific Data 8, 85,
https://doi.org/10.1038/s41597-021-00870-6.

FULL REPRODUCTION

Use a Python environment matching requirements.txt and software_environment.json.
The code has no dependence on the authors' local workspace paths.

1. Verify the archive:
   python verify_package.py

2. Install dependencies in your chosen environment:
   python -m pip install -r requirements.txt

3. Place the seven original authorized HCP/DLBS input files in one directory.
   Inspect the execution plan without downloads or model fitting:
   python run_all.py --hcp-dlbs-inputs /path/to/inputs \
     --output-dir /path/to/new_run --dry-run

4. Run all analyses, both 2,000-resample bootstraps, checks and Figure 3:
   python run_all.py --hcp-dlbs-inputs /path/to/inputs \
     --output-dir /path/to/new_run

The output directory must be empty or new. It contains a separate working copy
of the original AOMIC program and plan, leaving the distributed inputs/results
unchanged. Logs are written by analysis stage.

For offline use, supply the four original pinned AOMIC sources and official
S900 CSV explicitly; all source hashes are still verified:
   python run_all.py --hcp-dlbs-inputs /path/to/inputs \
     --aomic-source-dir /path/to/aomic/source \
     --s900-list /path/to/Unrelated_S900_Subject_multilist1_with_physio.csv \
     --output-dir /path/to/new_run

Each extension, synthesis and figure script also has an individual --help CLI.
The unchanged original AOMIC analysis is run inside its own directory; use the
wrapper to create that directory automatically.

INFERENCE AND REPORTING

Full-HCP point estimates and R2 values are discovery descriptions. Its OLS
confidence intervals and probabilities are nominal, because family identifiers
are unavailable. Neither HC3 nor participant-pairs bootstrap is used to pretend
that HCP family dependence has been resolved. Full-HCP standard errors never
enter the two-cohort synthesis.

DLBS/AOMIC HC3 intervals use residual-df t distributions. The DLBS bootstrap
samples 418 complete cases with replacement. Original baseline cognitive
domain composites are fixed, while domain standardization, PC1, g scaling,
age-spline knots/boundaries and all ILR standardizations are refit per draw.
AOMIC similarly refits PC1 and standardizations from its three released domains.
Both bootstraps use 2,000 draws and seed 20261027. Bootstrap standard deviations
and percentile intervals are supplied; no bootstrap p value is inferred.

All five-part models include four orthogonal ILRs. Six-part decompositions
include all five orthogonal ILRs. The size model including log(T) and log(eTIV)
is algebraically equivalent to log(eTIV) plus log(T/eTIV). T is the sum of the
five selected analyzed tissue compartments; T/eTIV is therefore called the
selected analyzed-tissue fraction. This identity does not establish why the
AOMIC association attenuates under eTIV replacement, nor a causal mechanism.

Column names distinguish OLS, HC3 and multiplicity-adjusted probabilities.
The supplemental six-part outputs include a direct cortex/cerebellar-WM
contrast. Magnitudes or significance of two separately standardized effects
are not themselves a formal test of their difference.

Numerical verification uses separate computational implementations; it does
not signify independent human review. Included audit files document exact
numerical comparisons, input hashes and model-rank/rotation checks.

Generated runs include participant-level source caches and private residual
CSVs solely for local analysis and figure construction. Exclude those files
from public uploads. The distributed archive includes only aggregate HCP/DLBS
results, code, metadata, and the explicitly identified public AOMIC data.
