Each paper contains its references, editable simulation setup, recipe modules, and comparisons of the legacy and new Libra dynamics implementations.
Start with Tully1990: its README covers environment setup, production runs, individual/summary plotting, and adding methods, initial conditions and models. Repository conventions are recorded in AGENTS.md.
<paper>/
config.py
recipes/<method>.py # load(dyn_general), including integrator choices
simulate_old.py, simulate_new.py, run.py, plot.py
reference/
<model>_<method>/
<initial_condition>/
old/runs/<tag>/ # production
old/smoke/<tag>/ # bounded checks
new/runs/<tag>/
new/smoke/<tag>/
comparisons/<production|smoke>/<tag>/
checks/<tag>/
A case combines a model and initial-condition variant with a scan. Select several methodologies for that case through recipe modules; keep scan points within each run. Configuration is defined once, with resolved copies saved alongside results. Keep original literature data independent of computed outputs.
For a new paper, create a new paper-named folder, document the source and conditions, and adapt the tutorial-style drivers and observables to that problem. The current Tully drivers assume one-dimensional, two-state adiabatic scattering; they are not a drop-in runner for arbitrary models. The model-extension guide lists the required simulation and analysis changes.
Run smoke checks before production. Production requires explicit --production;
choose a fresh tag for each attempt. Plotting reads saved data and never runs
simulations. Preserve current output, including incomplete attempts, unless removal is requested.
Keep the repository focused on the current recipe-based layout. A completed execution does not establish numerical or statistical convergence.