PROUSE RESIDUAL LIABILITY (Lr) MODEL v2.1 - SIMULATION STUDY: CODE AND RESULTS
Richard G. Prouse, Principal Analyst, Confidential Risk Consulting (New Zealand); author of the Prouse Residual Liability (Lr) Model (SSRN, 2026).
Licence: CC BY 4.0. Companion to "The Prouse Residual Liability Model, Version 2.1: Specification, Simulation Evaluation and Validation Programme", section 8 and Appendix E.

FILES
protocol.md   The pre-specified protocol (aims, worlds, models, measures, selection rule), written before results were generated.
sim.py        Main study: 7 worlds x 14 models; expert-error sweep; audit sample-size sweep. Writes results.json.
extra.py      Sobol sensitivity analysis, interval coverage, caps test, Version 1 x impact check, no-noise calibration. Writes extra.json.
cov2.py       Interval coverage including scoring uncertainty. Writes cov2.json.
figs.py       Figures 1-3 of the specification.
results.json, extra.json, cov2.json   Full results (mean, Monte Carlo standard error) for every world, model and measure.

RUN
Python 3.10+ with numpy, scipy, matplotlib.  python3 sim.py && python3 extra.py && python3 cov2.py
All random seeds are fixed, so results reproduce exactly on the same library versions (tested with NumPy 2.5.3, SciPy 1.18.1).

LIMITS
Results are conditional on the assumed worlds, input distributions and error levels. They show how the model behaves if courts and regulators act in these ways; they do not show that real courts do. Empirical validation (Studies 1-5 of the specification) is required for that.
