Layout

runs/
  <id>/
    manifest.json        the run's manifest (rnd-data-run/v1)
    files/               the run's small public files, at their paths inside the run
store.json               the stores outside this repository (rnd-data-store/v1)

Every field below is read from real manifests: a run (receipts--colab--window13l-20261018), a collection (receipts--colab--rung-prestep-muse-20261019) and a run with an identity-only private file (receipts--colab--rung-prestep-muse-20261019--try1-gate).

The run id

A run id is its source folder, with / written as --. receipts--colab--window13l-20261018 is the Publisher’s receipts/colab/window13l-20261018.

• A run is the nearest folder holding an outcome.json. Its files are the ones under that folder, so each retry (try1-gate/, try2-content-path/) is its own run.
• A collection is every other group: files outside a run folder are grouped under their window folder (for Colab), or else under their top folder. receipts--colab--rung-prestep-muse-20261019 collects the scripts and the out/ folder beside that window’s tries.

The manifest, field by field

Top level

Field	What	Example
schema	Always rnd-data-run/v1	"rnd-data-run/v1"
id	The run id, the same as its folder under runs/	"receipts--colab--window13l-20261018"
kind	run (its folder holds an outcome.json) or collection	"run"
source	The folder it came from, under its ingest’s prefix (receipts, rnd-watch)	"receipts/colab/window13l-20261018"
models	The models that generated its outputs; empty until they’re recorded	[]
licence_note	A note on the licence of what the models generated	"generating models not yet recorded for this run"
files	One entry per file in the run, public or listed private (below)
copies	Where each public file’s bytes are (below)
private_files	Files private for their content, counted by reason, never named	{"item-text": 6}
ingested	The date rnd wrote the manifest	"2026-10-19"
rnd_commit	The rnd commit that wrote it	"6c0ec58768b4114ee81804b04b8b5ec8c39e2b0f"

The four rnd-watch manifests were written by an earlier rnd (their rnd_commit is 96f1277…) and don’t carry kind.

files

Field	What
path	The file’s path inside the run, such as outcome.json or upload/llama-server.gz
sha256	Its SHA-256 in hex; null for a private file
bytes	Its size in bytes
tier	The data tier in rnd’s scale plan. Every file here is T4, raw data
licence	The file’s licence, such as CC-BY-4.0 or MIT; null only on a private entry
visibility	public, or private for a file listed without its hash
private_why	Only on a private entry: why. In a public manifest this is always identity

A public file and a listed private one, from receipts--colab--window13l-20261018:

{
 "path": "upload/llama-server.gz",
 "sha256": "b35cb0bb3d0a5c28d31d5b4cd88c1068e72d8ad0571d8d87b3fdcf3adf4a0729",
 "bytes": 127456485,
 "tier": "T4",
 "licence": "MIT (llama.cpp, commit 620d2225)",
 "visibility": "public"
},
{
 "path": "job.log",
 "sha256": null,
 "bytes": 575188,
 "tier": "T4",
 "licence": null,
 "visibility": "private",
 "private_why": "identity"
}

copies

Field	What
sha256	The hash of the copy, matching a public entry in files
store	The store that holds it: rnd-data (this repository’s git) or d (the blob store)
location	Where in that store: a path from this repository’s root, or blobs/<sha256[:2]>/<sha256>

Two files with the same bytes have the same hash, and each still has its own copy entry. A private file has no copy entry in a public manifest.

Small files and blobs

store.json sets the limit ("small_file_limit": 262144, 256 KB) and the blob layout.

• 256 KB or less: the file is in git, at runs/<id>/files/<path>, and its copy says "store": "rnd-data".
• Larger: it’s stored once, by content, at blobs/<sha256[:2]>/<sha256> in the blob store, and its copy says "store": "d". Every run that has the same file names the same blob, so a build several windows used is stored once. upload/llama-server.gz above is one, with its four split pieces.

The stores

Store	Kind	What
rnd-data	git	This repository (GitHub, public)
rnd-data-private	git	The private companion (GitHub, private): only files with a named reason
d	blobs	The studio rig’s backup drive, a blob store by SHA-256

A machine finds a blob store’s root through $RND_DATA_STORE_D or the key d in store.local.json, which git ignores, so no machine path is ever committed. The public blobs go to a Hugging Face dataset at the next phase.