holdings — the placement catalog

A disposable, regenerable catalog answering: which media hold which files? — and the two questions that follow from it: is this backed up? and is it safe to wipe this?

Single file, stdlib only, Python 3.10+. SQLite for placement truth; semantics belong to OntoDAG (see Projection contract, below).

pip install holdings

holdings add-medium laptop-x1      --kind laptop --site home
holdings add-medium drive-budapest --kind drive --durability independent \
                                   --site budapest --location "safe, Budapest"
holdings scan laptop-x1 /home/you
holdings scan drive-budapest /media/you/backup-drive

holdings whereis holiday.jpg                  # which media hold it
holdings redundancy --min-copies 2 --exit-code    # the policy, as a check
holdings redundancy --on laptop-x1 --min-copies 1 # safe to reformat?
holdings due                                  # what to re-read next

Documentation

Design contract (the important part)

1. Observer, not authority. holdings only reads filesystems and backup listings. It never writes to your data, never sits in the backup or sync path. Deleting holdings and its database costs nothing but convenience.
2. Everything is regenerable by re-scanning. The catalog is a cache of facts about the world. The only original data in the whole system is your human OntoDAG categorization — which holdings never touches.
3. Content hash is identity. sha256:… is the primary key everywhere. Paths, media, snapshots, categories: all attributes of a hash.
4. Single writer, many readers. Scan on the backup-node laptop; every other device only reads. How the catalog reaches them is not part of the contract — put the SQLite file in a synced folder (Syncthing, or anything else) and every device carries the full index of everything, including drives offline in another country. Or publish it read-only and have them query it without holding it at all (see Reading a published catalog, below). Both work; neither is required.

What it is careful about

The design is mostly a list of wrong answers it refuses to give.

A sync mirror is not a backup. A file on your laptop and in Dropbox has two copies and no backup: one rm removes both. Every medium declares what its copies survive — working, independent, mirror, leased, hosted — and only the last three count. See the User Guide §2.

Seen is not verified. A rescan reuses the stored hash without opening the file, and bit rot changes neither size nor mtime, so a copy can be catalogued faithfully for years and be gone. Instances record how they are known, and --verified-within discounts evidence nobody has refreshed.

Two drives in one drawer are not two sites. --min-sites and --min-kinds make the 2 and the 1 of 3-2-1 checkable, not just the 3.

A copy that expires is not a copy that does not. A lease is counted while it holds and reported when it lapses; hosting nobody else participates in — a Radicle repo seeded only by your own node — does not count at all.

An unreadable file is not a deleted one. A failing drive and a tidied-up one must not produce the same catalog change, so unreadable paths are kept and reported rather than pruned.

Beyond mounted media

Anything that can list itself can be recorded: restic, rclone, S3, a sha256sum run over ssh, a Syncthing cluster, a git remote (GitHub, Hugging Face, Radicle), a Swarm root. Where the source can supply a SHA-256, placement is exact rather than matched by name and size. See User Guide §7.

Human categories from OntoDAG can be materialised against placement, so redundancy --category vienna answers "which Vienna photos are unbacked?" as an index lookup — no lattice shipped to the reader. Prototype; see the User Guide and ontodag#20.

The catalog can also be published read-only and queried from a phone without holding it — a lookup fetches tens of kilobytes of a catalog that may be hundreds of megabytes — or served offline from a checkout with python web/serve.py. See Publishing.

Projection contract (OntoDAG integration)

(2026-08-20: this contract's canonical statement now lives at the meet point — ontodag docs/plans/PROJECTIONS.md — which generalizes it across sources (files here, messages in ucomm) and adds retention classes. The rules below remain the agreed file-side instance and the wire format is unchanged.)

holdings project-ontodag --out placement.jsonl

Emits JSON lines: {"item": "<hash>", "supercategories": ["sys:on:<medium>", "sys:type:<ext>", "sys:backup:<n>"]} — matching OntoDAG's put(item, supercategories) model.

The agreed rules for the ingesting side:

• everything under the sys: namespace is machine-written, regenerable cache — never hand-edit, never treat as authoritative;
• ingestion is an idempotent full rebuild: drop all sys: memberships, re-ingest the stream (no incremental diffing — staleness is the only permitted failure mode, drift is not);
• the projection is rebuilt locally on each device from the synced SQLite, not synced itself; only the human layer of the DAG is persisted and synced (it is the irreplaceable original data — back it up like the KeePass vault);
• human categories are attached to the same hash-identified items, giving unified queries like get(photo, vienna, sys:on:drive-budapest).

See ontodag_ingest.py for an adaptation template.

Notes & limits

• A listing proves paths and sizes, never bytes. Entries are matched to known content when that is unambiguous, recorded as unverified: otherwise, and taken exactly when the source supplies a SHA-256. Mounting the source and scanning it is the only thing that proves content.
• Symlinks are skipped. Hidden config and caches are excluded by default (.cache, .config, .git, node_modules, Syncthing internals, …); add your own with --exclude-file.
• Concurrent writes are not supported, by design: single writer, many readers.
• diff A B is proportional to what is on A, unlike the other reports.
• holdings cannot verify a durability class or a site. Those are your assertions when you register a medium; what it can do is stop one word standing in for five different things.
• Tests: pip install -e ".[test]" && pytest — 249 tests, stdlib only, no node and no network; a guard fails if that number drifts from the suite. The browser viewer's logic has its own: node --test web/test/view.test.mjs.
• Roadmap: see ROADMAP.md.