Reproducibility
The whole analysis is one script plus a few helpers, and every step from a smoke test to the final run is a single command. This section tells you what to install, where the data go, what each run mode costs in steps, and which output files are authoritative.
Three levels of rerun
| Level | Command | Steps | Purpose |
|---|---|---|---|
| Smoke test | --quick | 500 | check that the pipeline runs |
| Medium | --nsteps 5000 --burn-in 1000 --nwalkers 64 | 5,000 | exploratory look |
| Final | --nsteps 100000 --burn-in 20000 --nwalkers 96 | 100,000 | the published run |
Minimal path
- Install the package and check the environment.
- Place the four data files in
data/raw/(Quick Run). - Run the quick smoke test.
- Run the final command; expect a long run (see Quick Run).
- Read the outputs: the curated folder, not the raw run folder, is the reference.
What reproducibility covers here
| Covered | Not covered |
|---|---|
| Code, commands and the exact final command | Random seeds: the walker initialisation is not seeded, so a rerun gives statistically — not bit-for-bit — identical chains |
| Priors and run settings (MCMC Setup) | Package versions of the original run (not recorded) |
Result tables regenerated by scripts/export_final_tables.py | The survey data themselves (external products) |