#!/usr/bin/env python3
"""Compare the Inspect-pinned GDPval table and input metadata with two snapshots.

Only table hashes, aggregate metadata and comparison results are printed.
No task text, individual paths, URLs or input-file contents are retained.


Sapien1 makes this script available to download and run locally. No additional
license to modify or redistribute it is stated. Upstream terms are separate.
The later pinned GDPval dataset card includes a benchmark canary; check the
applicable upstream terms before use."""

import hashlib
import json
import urllib.request


REVISIONS = {
    "earlier": "4bf642692ef091918ef20797ecca070924815ec1",
    "inspect": "a3848a2a812d5d4d0f08003fac3c8eac40805962",
    "later": "11e7900cdcac61bc4daf59e65feb238acda98fbf",
}
EXPECTED_TABLE_HASH = "e26aeea5542ece91356d51cfb4d2f46cdb0ac86ab311b825b8d67bc06cbe4d9f"
EXPECTED_METADATA_HASH = "d016d8408cf871f260c4014894aad44b0b2f13f2b81118cd4912f459e0cd55f8"
TABLE = "data/train-00000-of-00001.parquet"
BASE = "https://huggingface.co/datasets/openai/gdpval/resolve/"
API = "https://huggingface.co/api/datasets/openai/gdpval/revision/"


def read(url, limit):
    request = urllib.request.Request(url, headers={"User-Agent": "Sapien1Research/1.0"})
    with urllib.request.urlopen(request, timeout=30) as response:
        data = response.read(limit + 1)
    if len(data) > limit:
        raise ValueError("source exceeded expected size")
    return data


def metadata(revision):
    document = json.loads(read(API + revision + "?blobs=true", 2_000_000))
    if document.get("sha") != revision:
        raise ValueError("repository revision mismatch")
    entries = [entry for entry in document["siblings"]
               if entry.get("rfilename", "").startswith("reference_files/")]
    result = {entry["rfilename"]: (entry.get("blobId"), entry.get("size"))
              for entry in entries}
    if len(entries) != len(result) or any(not blob or not isinstance(size, int)
                                          for blob, size in result.values()):
        raise ValueError("duplicate paths or incomplete input metadata")
    canonical = json.dumps(sorted((name, blob, size) for name, (blob, size) in result.items()),
                           ensure_ascii=False, separators=(",", ":")).encode("utf-8")
    return result, hashlib.sha256(canonical).hexdigest()


def main():
    table_results = {}
    for label in ("earlier", "inspect"):
        raw = read(BASE + REVISIONS[label] + "/" + TABLE, 1_000_000)
        digest = hashlib.sha256(raw).hexdigest()
        table_results[label] = {"bytes": len(raw), "sha256": digest}
        if digest != EXPECTED_TABLE_HASH:
            raise ValueError(label + " task table hash differs from reviewed snapshot")
        del raw

    snapshots = {label: metadata(revision) for label, revision in REVISIONS.items()}
    earlier, earlier_hash = snapshots["earlier"]
    inspect, inspect_hash = snapshots["inspect"]
    later, later_hash = snapshots["later"]
    if earlier != inspect or earlier_hash != EXPECTED_METADATA_HASH or inspect_hash != EXPECTED_METADATA_HASH:
        raise ValueError("Inspect-pinned input metadata differs from reviewed snapshot")
    shared = set(inspect) & set(later)
    changed = sum(inspect[path][0] != later[path][0] and inspect[path][1] != later[path][1]
                  for path in shared)
    print(json.dumps({
        "revisions": REVISIONS,
        "table_checks": table_results,
        "input_metadata": {
            label: {"count": len(entries), "sha256": digest}
            for label, (entries, digest) in snapshots.items()
        },
        "earlier_and_inspect_input_metadata_equal": earlier == inspect,
        "inspect_paths_present_in_later": len(shared),
        "shared_paths_with_changed_blob_and_size": changed,
        "scope": "Source table hashes and provider metadata only; no task text, paths or input contents retained.",
    }, indent=2))


if __name__ == "__main__":
    main()
