# Scope of the archived administrative-component results bundle

This document describes the 15-context component bundle and its tested replay tiers. The current 17-context synthesis is under `whole_project_synthesis_3377/`; its completion receipt and reproduction manifest supersede this document for current coverage. The separate whole-synthesis bundle supplies its own portable replay, acquisition list and scope. Do not treat the older workbook or the component bundle as the complete current synthesis.

This is a source-linked results and analysis-code bundle, not a certified one-command reproduction of the entire project.

The `calibration_stress_tests/` directory includes the twelve additional shared-bias scenarios from Supplement S65. With NumPy and SciPy installed, run `python project_results/calibration_stress_tests/replay_3392.py project_results/calibration_stress_tests/3392_calibration_stress_tests.json`. This small replay verifies the posterior summaries from a frozen likelihood; it does not refit raw data, cure the original fit's numerical limitations, or rerun the control-covariance diagnostics.

The `joint_country_3376/` directory includes the joint Brazil/Sweden extension in Supplement S66: 42 synthesis scenarios, eight prior sensitivities, 102 country diagnostics, finer/wider grids, and the saved common-center likelihood. Run `python project_results/joint_country_3376/replay_3376.py project_results/joint_country_3376/3376_joint_country_synthesis.json project_results/joint_country_3376/synthesis_161_6.npz` with NumPy and SciPy to verify all 42 main summaries. The original-cell input panels are retained under `global_integration/`. Full project code is included, but this compact replay does not regenerate the core likelihood or certify causal identification; the separate source-project reconstruction command requires its saved core synthesis cache.

The synthesis represented by this component bundle is `global_integration/hierarchy/hierarchy_final_p18.json`: fifteen pooled contexts, fourteen likelihood blocks and 5,133 outcome cells. Its shared mean ratio is 0.957 (95% credible interval 0.599-1.517), conditional on the model. It is not an individual schedule-risk ratio and is not the expanded 17-context headline.

The bundle includes the complete result register, the checked design table through its hashed source copy, the current PDFs, saved posterior summaries, international block inputs, fitted global block curves, current U.S./Japan frozen inputs, and analysis code. `sources.json` maps original filenames to exact hashed copies. Every stored file is listed in `bundle_manifest.json`.

Large numerical likelihood-draw caches are omitted; code and frozen inputs are retained. Raw survey downloads and some upstream acquisition dependencies remain outside this bundle. Code is collected under `code/`. In a fresh extraction, `reproduce_release.py` restores the relative `processed/` code layout alongside the extracted `project_results/` folder without using the originating machine's project path. Its runnable tiers are:

```text
python project_results/reproduce_release.py verify
python project_results/reproduce_release.py held-international
python project_results/reproduce_release.py current-reference
python project_results/reproduce_release.py alternate-2005
```

Use Python 3.12 and the exact versions in `requirements-release.txt`; a separate environment is recommended. The first command checks every manifest member. The second refits the saved international outcome observations and compares their reported likelihood summaries. The third rebuilds the current U.S./Japan power-18 importance-sampling likelihoods from frozen exposure/outcome inputs and refits **only the 15-context reference shared-mean model**. It compares seven numerical headline and diagnostic fields and writes `reproduction/current_reference_check.json`. Both these replay tiers passed from separate fresh extractions: every then-packaged manifest member verified, 936 international numbers matched exactly, and all seven current-reference fields matched the saved result exactly (shared mean 0.957219; 95% credible interval 0.598796–1.517032). The fourth command is a separate, more expensive replay of the named November-2005 rotavirus-onset sensitivity; its code and frozen inputs are included, but it has not undergone a second independent extraction-and-refit test. These refits can take substantial CPU time and disk space; each uses at most four numerical threads. Use a fresh extracted copy for each replay, since analytical commands may replace saved result files in that copy. These are not one-command reproductions of every historical estimate or sensitivity.

The local results browser works offline with `index.html` and `results_data.js` together. Inspection of saved numbers requires no model refit. A source hash, a passing validator, or a numerical rerun does not certify causal identification or worldwide data completeness. A public raw-download-to-result reconstruction remains unverified: several source-specific reconstructions exist, but the whole current register has not been rebuilt from newly reacquired raw files in an isolated environment.

The named alternative 2005 U.S. rotavirus exposure convention has now been estimated: its shared mean is 0.957612 (95% credible interval 0.599297–1.517320), versus 0.957219 (0.598796–1.517032) under the primary convention. It remains a sensitivity, not a replacement primary. The numerical low-effective-sample-size cases are documented in `finalization_numerics/low_ess_audit.json`; models below the specified ESS floor remain diagnostic or use the specified fallback, not silently discarded. Legacy downstream readers with unresolved historical probabilities are held out of the current claim set. Blocked or unsuitable source designs remain explicitly listed in the global integration files. The website's current_release_manifest.json identifies its downloadable artifact versions; the finalization receipt records publication separately from computational checks.