Two commands, and none of them require trusting this page.

Release

python install/mmrf.py --project-root . verify-release
python install/mmrf.py --project-root . doctor

The flag goes before the subcommand. --project-root is a top-level argument, so putting it after verify-release fails to parse — worth knowing, because parts of the shipped documentation show it the other way round.

release id

MMRF-1.0.0

signed release manifest

release_v10/stable_release_manifest_v1.0.json

release public key

release_v10/stable_release_signing.public.pem

stable manifest

stable_data/stable_manifest_v1.0.json

shards checked by doctor

20

Expect valid: false, and read why

Run against a clone of the runtime repository, verify-release reports valid: false. That is the correct answer. The repository is the published package plus two repaired files, and the check names them:

{
  "valid": false,
  "checks": {
    "signature_and_document_hash": true,
    "schema_ok": true,
    "release_id_ok": true,
    "version_ok": true,
    "safety_ok": true,
    "payload_ok": false
  }
}

hash_mismatch  lake/mmrf_data_lake.py
hash_mismatch  lake/mmrf_lake_cli.py

The signature still verifies, and every safety and semantic check passes. Only the payload hashes for the two repaired files differ, and the repairs are described in REPAIRS.md. To verify the release exactly as published, check out the import commit — it is byte-identical to the package — and run it there.

The dataset is untouched either way. No shard byte changed, and the stable manifest hash below is the one the release published.

Reproduce the manifest hash

The manifest hash is not a hash of the file. It is the SHA-256 of the canonical JSON of the manifest object with the manifest_sha256 field removed — a self-referential field cannot be inside the thing it describes. Reproducing it takes four lines:

import json, hashlib
m = json.load(open('stable_data/stable_manifest_v1.0.json', encoding='utf-8'))
core = {k: v for k, v in m.items() if k != 'manifest_sha256'}
canon = json.dumps(core, ensure_ascii=False, sort_keys=True,
                   separators=(',', ':'))
print(hashlib.sha256(canon.encode()).hexdigest() == m['manifest_sha256'])

expected

a5caea22a57efaac915c00dd92c655b1126e0b6d9b2b93790e48bc167733e0d1

Count the records yourself

The shards are plain NumPy archives. Nothing stops you counting them without any MMRF code at all:

import numpy as np, pathlib
total = 0
for p in sorted(pathlib.Path('stable_data/shards').glob('*.npz')):
    with np.load(p) as z:
        total += len(z['prime'])
print(total)   # 148,933

Check it against the literature

Three values in this dataset are independently known, which makes them the cheapest external check available:

A table where both columns agree is only worth printing because the columns come from different places: the left from the shards, the right from published values. If they ever diverge, this page is the thing that should change.

Machine-readable

Static, no JavaScript, no content negotiation: