Validation Data for processbehavior Package
============================================

What lives here and why
-----------------------
- PBTESTDATABASE_T100.csv - Tom Bishop's reference database (4000 rows;
  FACTOR 1 x FACTOR 2 x 100 time periods; response columns PM SDS 1-6 and
  PM INERT, each structured to exercise one observed design state). Ground
  truth for the CI validation gate (e2e_bishop_report.py) and the
  Bishop Reference Validation tutorial.
- PBTESTDATABASE_T20.csv - the 20-period variant, consumed by
  e2e_T20_report.py.
- e2e_bishop_report.py - runs the full 280-assertion check of computed
  results against Bishop's Minitab numbers; executed in CI and regenerates
  docs/reference/validation.md (drift-guarded there). Writes
  e2e_bishop_report.html locally (untracked).
- e2e_T20_report.py - the same idea against the T20 database.
- effects_visualization.ipynb - effects-chart exploration against the
  reference data.
- sds1_data.csv - synthetic design-state-1 dataset (make_design-style
  structure, seed 42); loaded by the Complete Design-State-1 Analysis
  tutorial (docs/tutorials/sds1-complete-analysis.ipynb).
- electrical_resistance_in_megohms_204.csv - fixture for
  tests/test_plotting.py.
- GROSSREVATTENDERRORDATABASE.csv - garbage-character fixture for
  tests/test_garbage_character_handling.py.
- aco_per_capita_expenditure.csv - the synthetic 24-organisation x 4-year
  dataset (96 rows, ADS 2) built with Tom Bishop for GitHub issue #114, from
  the reporter's description of Medicare shared-savings series. Input for
  mr_permutation_invariance.py.
- short_series_sampling.py, short_series_bands.py - contributed by the #114
  reporter (PR #119). Self-contained: regenerate the limit-width-vs-T table
  and the sampling distribution of MRbar/d2 for n = 3..30 under a stable
  process. No data files needed.
- mr_permutation_invariance.py - contributed by the #114 reporter (PR #130).
  Holds every observation of the ACO file fixed and shuffles only the lane
  order of the combined by=[] chart, reporting which results move (boundary
  mR signals) and which do not (the X chart's "not one process" conclusion).
  Run from the repo root:
      python validation/mr_permutation_invariance.py validation/aco_per_capita_expenditure.csv

Removed in the 2026-08 cleanup (recoverable from git history):
sds1-6_analysis.xlsx (Excel exports generated by the package itself - not
external reference truth; the Minitab ground truth lives in
e2e_bishop_report.py's expected values), sds2-6_data.csv (regenerable
synthetic data; only sds1_data.csv is consumed by anything),
sds1.ipynb, and sds_raw_vs_tidy.py (a one-off diagnostic for raw-vs-tidy
detection drift, now covered by the design-state lineage tests and docs).

Regenerating synthetic data
---------------------------
Synthetic design-state datasets come from the canonical generator:

    from processbehavior.datasets import synthetic
    df = synthetic.make_design(state, seed=42)   # state in 1..6

(The make_sds() dispatcher mentioned in older revisions of this file was
removed; make_design is its replacement.)