Labelers

Hypercerts runs two labelers. The Activity Labeler rates the quality of Hypercerts activity records, and the Orglabeler rates how complete and credible a Certified organization looks. Both publish their verdicts as labels that anyone can read, so your application can tell well-formed records apart from drafts, placeholders, and test data without writing its own checks.

Where it fits#

The labelers sit beside the main read path. They take in records from the relay and publish labels. Your application can read labels directly from a labeler. The Feed Service uses Orglabeler labels to filter organizations. The services overview has the full diagram.

AT Protocol background#

A label is a short, signed statement that one account makes about a record or about another account, such as "high quality" or "likely test". It has a source (src, the account that issued it), a subject (uri, the labeled record or account), and a value (val). A label does not change the data it describes. The record stays in its owner's repository, untouched, and the label is published separately.

A labeler is an account, with a service behind it, that publishes labels. It declares itself and the label values it uses in a record in its own repository, and it serves its labels through two standard methods that any client can call.

Nothing in the protocol makes a label binding. Each application chooses which labelers it listens to and what a label means for it: show a badge, sort a list, hide an item, or do nothing. AT Protocol calls this composable moderation. It allows several independent labelers to judge the same records by different standards, and it leaves the decision about whom to trust with the application. For the details, see the AT Protocol label specification.

How it works#

A shared pipeline#

Both labelers work the same way:

Ingest. A tool called Tap follows the relay, which is the service that gathers record changes from servers across the network. Tap loads older records and then streams new ones as they are published.
Score. Each labeler applies its own criteria to the records it receives.
Publish. The labeler signs a label and serves it through the standard label methods.

Labels can change. When a record is edited, or when a slower check finishes, the labeler withdraws the earlier label and publishes a new one.

Activity Labeler#

The Activity Labeler, activitylabeler.certified.one, scores org.hypercerts.claim.activity records and labels each record with a quality tier.

Label	Meaning
`pending`	The record was detected and is being evaluated
`high-quality`	A well-documented activity with comprehensive details
`standard`	An adequate activity with basic information filled in
`draft`	A minimal activity that looks like a work in progress
`likely-test`	The activity appears to contain test or placeholder data

The score rewards a meaningful title, summary, and description, and the presence of an image, work scope, contributors, locations, a date range, and rights. Placeholder text, repeated content, and obvious junk lower it. A separate classification step can downgrade content that is not meaningful to likely-test.

Orglabeler#

The Orglabeler, orglabeler.certified.one, reads app.certified.actor.profile and app.certified.actor.organization records, merges the two for each account, and labels the organization's account. Its subject is the account's DID (the permanent account ID), not a single record.

Label	Meaning
`likely-test`	Clear test evidence, such as an account on a test server or placeholder names, domains, or descriptions
`standard`	A non-test organization that scores below 70
`high-quality`	A non-test organization that scores 70 or more

The score is out of 100 points across 13 signals, covering the display name, description, organization type, website and other URLs, location, founded date, avatar, and banner. Accounts hosted on a trusted server, including the Certified production PDS (Personal Data Server), get a small bonus. Test detection runs first, so an obvious test account lands in likely-test however complete it is. Website checks run in the background, so a label can change shortly after the first score. The full scoring table is in the Orglabeler README.

Using it from your application#

Labelers expose two standard XRPC methods. XRPC is AT Protocol's convention for HTTP APIs, where each method is called at /xrpc/<method name>.

com.atproto.label.queryLabels returns the current labels for the subjects you ask about.
com.atproto.label.subscribeLabels streams labels over a WebSocket as they are published, for applications that keep their own copy.

For example, to read the Orglabeler's label for one organization, pass the organization's DID as the subject:

Terminal

curl --get https://orglabeler.hypercerts.dev/xrpc/com.atproto.label.queryLabels \
  --data-urlencode 'uriPatterns=did:plc:example'


# {
#   "labels": [
#     { "src": "did:plc:pswneepkd5lesumj7ejmkbal", "uri": "did:plc:example",
#       "val": "high-quality", "cts": "2026-08-05T22:08:58.761Z" }
#   ]
# }

For the Activity Labeler, send the same request to its host with the activity record's AT-URI (its at:// address) in uriPatterns. Hosts and labeler DIDs are listed under Running services.

When you use labels:

Check the source. Compare src with the DID of a labeler you have chosen to trust. A label from an unknown source carries no weight on its own.
Treat labels as signals, not verdicts. A standard organization is not a bad one. It has filled in fewer fields.
Handle the unlabeled case. New records and accounts are labeled a short time after they appear, so keep a fallback for subjects with no label yet.

For labels, query a labeler directly.

Status and source#

Both labelers are running in production and label new records as they are published. They have no changelog page in this documentation.

The Orglabeler's source code is in the orglabeler repository. The Activity Labeler describes its scoring at activitylabeler.hypercerts.dev/docs. Running your own labeler is outside the scope of this documentation for now.