real-data-v3/verify_real_table2_cpu.py

#!/usr/bin/env python3
"""Checks for the ten-dataset ridge sensitivity reproduction."""

from __future__ import annotations

import json
from pathlib import Path

import numpy as np
import pandas as pd


ROOT = Path(__file__).resolve().parent
OUT = ROOT / "results" / "real_table2_cpu"


def main() -> None:
    raw = pd.read_csv(OUT / "raw_results.csv")
    summary = pd.read_csv(OUT / "summary.csv")
    report = json.loads((OUT / "verification.json").read_text())
    assert len(raw) == 50
    assert len(summary) == 10
    assert raw.groupby("dataset").seed.nunique().eq(5).all()
    assert report["included_meps"] and report["included_bio"]
    assert report["dataset_count"] == 10 and report["runs"] == 50
    # This is a deliberately retained negative control: with a stronger Ridge
    # point predictor the length reduction is not universal.
    assert report["protocol"]["predictor"] == "ridge"
    assert report["pt_shorter_count"] == 6
    assert np.isfinite(raw.select_dtypes(include="number").to_numpy()).all()
    assert np.allclose(summary.pt_length_mean, raw.groupby("dataset", sort=False).pt_length.mean().to_numpy())
    assert np.allclose(summary.vcp_coverage_mean, raw.groupby("dataset", sort=False).vcp_coverage.mean().to_numpy())
    assert abs(report["mean_vcp_coverage"] - 0.9) < 0.025
    assert abs(report["mean_pt_coverage"] - 0.9) < 0.025
    assert report["max_abs_mean_coverage_difference"] < 0.025
    assert (OUT / "real_table2_cpu.png").stat().st_size > 40_000
    print("PASS: ridge sensitivity control retained (6/10 shorter; nominal coverage)")


if __name__ == "__main__":
    main()