REPRODUCING THE THIRD REVISION

Use Python 3.12 in an isolated environment and install analysis/requirements.txt. Run the following from the extracted package root:

python analysis/analyse_study.py --source source_data --out recomputed_primary
python analysis/audit_supplements.py --source source_data --out recomputed_supplement
python analysis/deepen_analysis.py --source source_data --out recomputed_deepened
python analysis/make_figures.py
python analysis/make_extra_figure.py

The first three commands read the unchanged source files and create analytical outputs; they do not alter participant data. The two figure commands use the delivered JSON results and write figures_rebuilt. If plotting recomputed estimates, first preserve the delivered JSON files separately, then put copies of the recomputed JSON files into analysis under the original names.

PRIMARY ANALYSES
Behavioural n=405; baseline-adjusted n=354; common path models n=297; all-item CFA n=195. Composite scores require all items in each block. Available-item scoring is an explicitly labelled sensitivity, not multiple imputation. Participant-bootstrap draws are repeated estimates from existing records, not additional students. See analysis_results.json for exact versions, seeds, coefficients and intervals.

SUPPLEMENTAL AUDIT
Workbook 05 is a main-study summary, not an independent pretest. Its printed t statistics are X=0 minus X=1; its mean differences are X=1 minus X=0. The recomputed tables orient contrasts X=1 minus X=0 and distinguish Welch and pooled inference. Holm correction covers 14 exploratory item tests. Initial-position interaction estimates are exploratory.

NEW CONDITIONAL CHOICE ANALYSIS
Each of 405 participants contributes one observed four-element set out of 70 possible sets. There are 14 initial-position-specific material indicators (m8 is the reference in each position) plus X times opposing. The score and Hessian are computed from the exact conditional likelihood. Participant-robust and school-clustered covariance estimates are reported separately; 3240 participant-material alternatives are not counted as independent participants. An independent statsmodels ConditionalLogit implementation checks parameter and likelihood calculations, not the substantive assumptions.

NEW MECHANISM SENSITIVITY
Using the same n=297 linear path sample, b(rho)=b_OLS-rho*s_e/[s_u*sqrt(1-rho^2)] under the stated homoscedastic structural-error specification. The two residual RMS values share denominator n. The indirect component is a*b(rho). Rho is hypothetical, not estimated. The output is a deterministic sensitivity of point estimates, not a confidence band or a simulated participant dataset. This analysis was newly computed; it does not adopt undocumented sensitivity entries from workbook 03.

REPORTING STATUS
All new analyses were added retrospectively and are labelled exploratory. S1-S5 in the main manuscript are illustrative procedural/declaration completions; they have no role in the numerical estimates. Unresolved stimulus provenance and institutional ethics applicability remain disclosed. This is an author-review simulation package, not a submission-ready claim that those missing records have been verified.
