System fields

Every record carries fields you never write: per-field version stamps, a delivery counter, tombstone markers, creation and modification instants, and — on objects — a handful of built-in properties. This page lists them, says which are synced and which are peer-local, and which ones are safe to sort, page or filter on.

Record-level bookkeeping

Present on every record returned by /query and in subscribe events. Client writes addressing them are rejected.

Version ids are peer-local: two devices holding the same change name it by different _ver values. Use them for ordering and paging on the device that produced them, never as identifiers you exchange. The exchangeable identifier of a change is its changeId — see Version history.

Note. History diffs never carry _ver. Snapshot rows and added/updated records do — narrowed to your projection when you send one, in full otherwise. {"projection": {"_ver": -1}} drops it, and _addSeq/_applySeq drop by default under any projection.

Row-root stamps on objects

Every row of the per-space objects collection carries these next to id, all derived and read-only:

Both instants are the author's clock and are written in the {"$date": …} wire shape:

{ "id": "bafy…", "author": "A5k…",
  "createdAt":  { "$date": "2026-08-05T17:00:00.000Z" },
  "modifiedAt": { "$date": "2026-08-20T09:12:31.000Z" },
  "modifiedBy": "A9t…" }

"Recently modified first" is {"sort": ["-modifiedAt"]}. A filter literal must take the same shape — {"modifiedAt": {"$gte": {"$date": "2026-01-01T00:00:00Z"}}}. A bare number or string does not error; it silently matches nothing, because ordering comparisons are bracketed by type and a number never compares against an instant. See Data types.

modifiedAt and modifiedBy are one pair. Both come from a single change — the object's latest by DAG order, whatever dataset it landed on: a property write, an editor block, a chat message, a runtime-dataset record, a record delete. They carry that change's version, so they move together and never pair one change's time with another's signer. A change that arrives late regresses neither. Concurrent writers are resolved by DAG order, not by clock, so the identity that wins can be the one whose wall clock reads earlier. Deleting the object removes the row outright, stamps included.

modifiedBy is an account identity in the same encoding as author, as chat creator, as identity in GET /v1/spaces/:spaceId/members, and as id from GET /v1/account — resolve a name and icon through the members list, and through the account-global identities directory (GET /v1/identities/:identity) for a past writer who has since left the space. modifiedAt is indexed and is the conventional recency ordering; modifiedBy is not, so a filter on it scans the collection. A row without modifiedBy has either not been rebuilt yet (rebuilds run on each object's first load, plus a background sweep) or its latest change has no known signer — never "nobody modified it".

The any type's property listing (GET …/types/any/properties) carries the stamps as derived properties — modifiedBy appears there as "Modified by" — but their values sit at the row root. Filter and sort by the bare name; any.modifiedBy matches nothing.

Runtime datasets get the same trio on demand through stamp: creator / createTime / modifyTime fields (Runtime datasets); chat messages carry creator / createdAt / modifiedAt.

Why it matters. There is no server clock to trust. Every stamp is whichever device wrote the change, and the CRDT converges on a value, not the true time. Sort and display on these freely; never use them as a fence for "has everything before T arrived" — that is what sync status is for.

Built-in properties on objects

Objects carry a few properties under the universal any type. Paths use literal keys, not content-addressed property ids.

Tree placement is not a system field. The wiki usecase's parentId / pos / folder are ordinary properties of a hidden type, at <wikiTypeId>.<propId> on objects that carry it, and nothing is stamped on create — Objects.

Field scopes

Every dataset field belongs to one scope, visible as x-scope in the dataset's schema:

Negation operators ($ne, $nin, $not, $exists: false) also match rows that lack the field entirely — on the objects collection, which holds every object including type definitions, always scope by any.types first.