POLICYSHIFT-AML SUPPLEMENTARY MATERIAL
======================================

Contents
--------

experiments/
  Corrected sanity and Stage 1-4 experiment scripts.
  MANIFEST.md records the definitive artifact directories and corrections.

experiments/artifacts/
  Compact CSV result files and JSON run manifests for all four definitive
  ten-seed stages.

Makaleler/Academic Study/experiments/
  TransXion feature/baseline support code and IBM stratified-cache preparation.

paper/make_figures.py
  Recreates all four manuscript figures from the definitive artifacts.

Data
----

Large datasets and processed Parquet caches are excluded. Obtain TransXion from:
  https://github.com/chaos-max/TransXion

Obtain IBM AMLWorld data from:
  https://github.com/IBM/AML-Data

The scripts expect:
  Makaleler/Academic Study/experiments/cache/transxion_causal_features.parquet
  Makaleler/Academic Study/experiments/cache/ibm_hi_small_typology_sample.parquet

The IBM cache contains 216,315 stratified rows: all retained positives plus
sampled background. It is not the full 5.08-million-row source table.

Environment
-----------

The definitive runs used Python 3.13.2 on Windows 11. Required packages are
listed in requirements.txt. MiKTeX is needed only to compile the manuscript,
not to reproduce the experiments.

Reproduction order
------------------

  python experiments/run_sanity.py
  python experiments/run_stage1_policyshift.py --seeds 10
  python experiments/run_stage2_estimated_pu.py --seeds 10
  python experiments/run_stage3_graph_invariance.py --seeds 10
  python experiments/run_stage4_typology_bounds.py --seeds 10
  python paper/make_figures.py

The scripts write timestamped artifact directories. Oracle quantities are
simulation diagnostics and must not be interpreted as deployable information.
