code/verify_cumulative.py

#!/usr/bin/env python3
"""Fail-closed recheck of the preserved, previously accepted Claim 1–3 evidence."""
from __future__ import annotations

import csv
import hashlib
import io
import json
import math
import zipfile
from collections import defaultdict
from pathlib import Path

import numpy as np
from scipy import stats


ROOT = Path(__file__).resolve().parents[2]
EXPECTED_HASHES = {
    "outputs/claim1_accuracy/full_n512.csv":
        "3016dfb245deb56b3d0d3d06f902471bc415f80196d61ecf4d49eb773f1edd76",
    "outputs/codenet/full_gpu_n200.csv":
        "7843f75c38b9ee0bf3f15d796f151808fc01a526c704f429049016d57538f390",
    "outputs/colab/regresslm_evidence_bundle.zip":
        "b1d33e20922194cab7dfd20526cac13680f1ca16ce51ad7836a358feee5ebd1f",
    "outputs/colab/table3_results.json":
        "173dbce239ca51269d3b231dd851b7fb299dffc82d2b0eede75f61ae6a835dc2",
}
EXPECTED_ACCURACY = {
    "NASBench101": 0.4065993610622034,
    "ENAS": 0.24946117751369043,
    "NASNet": 0.20673752343866375,
}
EXPECTED_CODENET_LANGUAGES = {
    "C++", "Python", "Java", "C", "Ruby", "C#", "Rust", "Go", "Haskell",
    "Kotlin", "JavaScript", "PHP", "D", "Scala", "OCaml", "Perl", "Fortran",
}


def sha256(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for block in iter(lambda: handle.read(1024 * 1024), b""):
            digest.update(block)
    return digest.hexdigest()


def check_hashes() -> dict[str, str]:
    observed = {}
    for relative, expected in EXPECTED_HASHES.items():
        path = ROOT / relative
        if not path.is_file():
            raise AssertionError(f"missing preserved artifact: {relative}")
        observed[relative] = sha256(path)
        if observed[relative] != expected:
            raise AssertionError(f"SHA-256 mismatch: {relative}")
    return observed


def check_accuracy() -> dict:
    grouped: dict[str, list[tuple[float, float]]] = defaultdict(list)
    identifiers = set()
    path = ROOT / "outputs/claim1_accuracy/full_n512.csv"
    with path.open(newline="") as handle:
        for row in csv.DictReader(handle):
            identifier = row["identifier"]
            if identifier in identifiers:
                raise AssertionError(f"duplicate accuracy identifier: {identifier}")
            identifiers.add(identifier)
            draws = np.asarray([float(row[f"draw_{i}"]) for i in range(8)])
            prediction = float(row["prediction"])
            if not math.isclose(
                prediction, float(np.nanmedian(draws)), abs_tol=1e-12
            ):
                raise AssertionError(f"stored prediction is not draw median: {identifier}")
            grouped[row["space"]].append((float(row["target"]), prediction))
    if set(grouped) != set(EXPECTED_ACCURACY):
        raise AssertionError(f"accuracy spaces differ: {sorted(grouped)}")
    rows = []
    for space, expected in EXPECTED_ACCURACY.items():
        pairs = grouped[space]
        if len(pairs) != 512:
            raise AssertionError(f"{space}: expected 512 rows, got {len(pairs)}")
        target, prediction = map(np.asarray, zip(*pairs))
        rho = float(stats.spearmanr(target, prediction).statistic)
        if not math.isclose(rho, expected, abs_tol=1e-12):
            raise AssertionError(f"{space}: Spearman changed: {rho}")
        rows.append({"space": space, "n": len(pairs), "spearman": rho})
    return {
        "rows": len(identifiers),
        "raw_draws": len(identifiers) * 8,
        "per_space": rows,
        "mean_spearman": float(np.mean([row["spearman"] for row in rows])),
    }


def check_codenet() -> dict:
    grouped: dict[str, list[tuple[float, float]]] = defaultdict(list)
    path = ROOT / "outputs/codenet/full_gpu_n200.csv"
    with path.open(newline="") as handle:
        for row in csv.DictReader(handle):
            grouped[row["language"]].append(
                (float(row["y_true"]), float(row["y_pred"]))
            )
    if set(grouped) != EXPECTED_CODENET_LANGUAGES:
        raise AssertionError(f"CodeNet languages differ: {sorted(grouped)}")
    per_language = {}
    for language, pairs in grouped.items():
        if len(pairs) != 200:
            raise AssertionError(f"{language}: expected 200 rows, got {len(pairs)}")
        target, prediction = map(np.asarray, zip(*pairs))
        per_language[language] = float(
            stats.spearmanr(target, prediction).statistic
        )
    average = float(np.mean(list(per_language.values())))
    if not math.isclose(average, 0.5234034026121069, abs_tol=1e-12):
        raise AssertionError(f"independent CodeNet mean Spearman changed: {average}")
    return {
        "provenance": "independent CPU run",
        "languages": len(grouped),
        "rows_per_language": 200,
        "average_spearman": average,
        "per_language_spearman": per_language,
    }


def check_table3_summary() -> dict:
    summary = json.loads(
        (ROOT / "outputs/colab/table3_results.json").read_text()
    )["table3"]
    apps = summary["APPS"]
    kbss = summary["KBSS"]
    if apps["n"] != 512 or not math.isclose(
        apps["spearman"], 0.9268067718469594, abs_tol=1e-15
    ):
        raise AssertionError("APPS preserved summary changed")
    if kbss["n"] != 512 or not math.isclose(
        kbss["spearman"], 0.5352789637599933, abs_tol=1e-15
    ):
        raise AssertionError("KBSS preserved summary changed")
    return {"APPS": apps, "KBSS": kbss}


def check_bundle() -> dict:
    path = ROOT / "outputs/colab/regresslm_evidence_bundle.zip"
    with zipfile.ZipFile(path) as archive:
        bad_member = archive.testzip()
        if bad_member:
            raise AssertionError(f"corrupt evidence ZIP member: {bad_member}")
        summary = json.loads(archive.read("summary.json"))
        observed = []
        for expected in summary["claim_3_codenet"]["per_language"]:
            rows = list(csv.DictReader(io.TextIOWrapper(
                archive.open(expected["task"] + ".csv")
            )))
            if len(rows) != 200:
                raise AssertionError(
                    f"{expected['task']}: expected 200 rows, got {len(rows)}"
                )
            targets = []
            predictions = []
            for row in rows:
                draws = np.asarray([float(row[f"draw_{i}"]) for i in range(8)])
                prediction = float(row["prediction"])
                if not math.isclose(
                    prediction, float(np.nanmedian(draws)), abs_tol=1e-12
                ):
                    raise AssertionError(
                        f"{expected['task']}: prediction is not draw median"
                    )
                targets.append(float(row["target"]))
                predictions.append(prediction)
            rho = float(stats.spearmanr(targets, predictions).statistic)
            if not math.isclose(rho, expected["spearman"], abs_tol=1e-12):
                raise AssertionError(
                    f"{expected['task']}: stored Spearman changed"
                )
            observed.append(rho)
        primary_average = float(np.mean(observed))
        if not math.isclose(primary_average, 0.5298501742814378, abs_tol=1e-12):
            raise AssertionError(
                f"primary CodeNet mean Spearman changed: {primary_average}"
            )
    return {
        "members": 19,
        "model_commit": summary["model"]["checkpoint_commit"],
        "parameters": summary["model"]["parameters"],
        "transformers": summary["environment"]["transformers"],
        "seed": summary["environment"]["seed"],
        "primary_codenet": {
            "provenance": "Colab evidence bundle",
            "languages": len(observed),
            "rows_per_language": 200,
            "raw_draws": len(observed) * 200 * 8,
            "average_spearman": primary_average,
        },
    }


def verify() -> dict:
    return {
        "status": "PASS",
        "scope": (
            "independent recomputation from preserved row-level data for ONNX "
            "accuracy and CodeNet; integrity and summary recheck for APPS/KBSS"
        ),
        "artifact_sha256": check_hashes(),
        "claim_1_accuracy": check_accuracy(),
        "claim_2_table3": check_table3_summary(),
        "claim_3_codenet": check_codenet(),
        "bundle": check_bundle(),
        "limitations": [
            "This baseline rechecks preserved evidence; it does not rerun model inference.",
            "Full APPS/KBSS row-level draws were not present in the judged repository.",
        ],
    }


def main() -> None:
    result = verify()
    print("CUMULATIVE_RESULT " + json.dumps(result, sort_keys=True))


if __name__ == "__main__":
    main()