Getting started

Clone it

git clone https://github.com/mcp-tool-shop-org/rnd-data
cd rnd-data

Every file is stored byte-exact (.gitattributes has * -text), so git never converts line endings and every hash in a manifest holds on any platform.

Find a run

Each folder under runs/ is one run or collection, named by its run id: its source folder with / written as --.

ls runs/
cat runs/receipts--colab--rung-prestep-muse-20261019--try1-gate/manifest.json

That run, from the Publisher’s receipts/colab/rung-prestep-muse-20261019/try1-gate, lists three files: gate.jsonl and outcome.json are public, with their hashes and a copy each under files/; window.log is private for identity, listed by path and size with "sha256": null. See Layout for every field.

An experiment in rnd names its runs here by raw_manifest_id, and rnd where finds every copy of a file by its hash or path.

Check a file by hand

A file’s SHA-256 is in its run’s manifest, under files and again under copies. Hash the file and compare:

sha256sum runs/receipts--colab--rung-prestep-muse-20261019--try1-gate/files/gate.jsonl
# df19ad096ad0add77162cce929506d10da112b78c5398e69b86e801532da120b

grep -A1 '"path": "gate.jsonl"' runs/receipts--colab--rung-prestep-muse-20261019--try1-gate/manifest.json
#    "path": "gate.jsonl",
#    "sha256": "df19ad096ad0add77162cce929506d10da112b78c5398e69b86e801532da120b",

Each copy’s location is relative to the repository root, so every git copy can be checked at once with jq:

jq -r '.copies[] | select(.store == "rnd-data") | "\(.sha256)  \(.location)"' runs/*/manifest.json \
  | sha256sum -c --quiet

It prints nothing when every copy matches, and names each one that doesn’t. A file larger than 256 KB has its copy in the blob store ("store": "d", at blobs/<sha256[:2]>/<sha256>), not in git; once the blobs are public on Hugging Face, a downloaded one is checked the same way, against the sha256 its manifest names.

Check it with rnd

Install rnd, then, from the root of this checkout:

bash verify.sh                                   # what CI runs: the public check
RND_DATA="$PWD" rnd raw verify --public-only     # the same, by hand
RND_DATA="$PWD" rnd raw verify --run receipts--colab--rung-prestep-muse-20261019--try1-gate

--public-only re-hashes every git copy and checks the private-file rules, and skips the stores a public clone doesn’t have. Verifying has the details, and the full check the rig runs.