Provenance and independence: roles, origins and owners

Two counts, two questions

Every search hit carries independent_sources: the number of distinct hosts carrying the passage’s story, counted across near-duplicates dropped at ingest. It is a good filter (min_independent_sources=3) and a cheap signal. It is not a measure of independence: ten sites can run one wire dispatch, copy one press release, or belong to one media group, and every one of them counts as a host.

ground computes a different number. Within each story it groups members into origins, assigns each origin an owner, and reports independent_origins — origins with distinct owners. That is the count to hold a claim to.

source_role

What kind of source a passage is, for the purpose of weighing it as evidence.

Role	Rule
official	The host is a government, regulator, court, standards body or university: authority is gov or edu, the TLD is gov, mil or edu, or the host is under .gov.uk, .gc.ca, .gov.au, .europa.eu or .ac.uk.
primary	Documentation, an academic paper, code, a published document (content_type docs, academic, code, document), or a press release — an organisation speaking for itself.
secondary	Reporting, analysis or reference material about the subject (news, article, reference, opinion, how-to, video, product).
aggregator	A host that collects other hosts’ content (news aggregators).
ugc	User-generated: forums, Q&A, social hosts, or content_type forum / qa.
unknown	None of the above applied.

The rules run in that order, so a .gov host is official whatever its content type. Each label carries a role_confidence: high for a host rule, lower for a content-type rule.

origin_type

How a passage came to exist relative to the story it belongs to.

Type	Meaning
original	The earliest member of its origin — the thing others copied, or a lone source.
wire	A news-wire dispatch (Reuters, AP, AFP, dpa, PA, …), by host or by dateline in the text.
press_release	A press release, by distributor host or by distributor marker in the text.
republish	A later copy of an origin that already exists in the set. Never an evidence item.

Origins

Within a story, members are visited earliest-first. A member joins an existing origin when it carries the same wire service, the same press-release distributor, the same canonical URL, or the same normalised title as that origin’s representative; otherwise it opens a new origin. Two passages of the same page share a title and therefore an origin, which is right: they are one document.

{
  "id": "o:3f2a9c1e4b7d",
  "story_id": "…",
  "kind": "wire",
  "owner": "Thomson Reuters",
  "hosts": ["reuters.com", "localpaper.com", "othernews.com"],
  "registrable_domains": ["localpaper.com", "othernews.com", "reuters.com"],
  "members": ["a:0", "b:0", "c:0"],
  "representative": "a:0",
  "earliest_published_at": 1757790000,
  "confidence": 0.85
}

confidence says how strong the grouping signal was: a wire dateline is stronger than a title match across two unrelated domains.

Owners

An origin’s owner is the wire service for a wire dispatch; otherwise the media group that owns the representative’s registrable domain, from a curated table of the groups whose outlets most often appear together in one story; otherwise the registrable domain itself. news.bbc.co.uk and bbc.com are one owner; thehill.com and newsnationnow.com are one owner; example.com is its own.

The table is a starter list and is stated as such in /describe. A domain it does not know is treated as independent of every other domain, so the count errs towards over-stating independence for small outlets and never towards hiding a known group.

independence_score

Per evidence item, within the packed set:

Score	Meaning
1.0	The only origin with this owner in the set.
0.5	Another origin in the set shares the owner.
0.0	A republish — which never becomes an evidence item, so you will not see it.

The set-level number is coverage.independent_origins. hosts and copies on the receipt show how much syndication was collapsed to reach it.

What this does and does not detect

It detects: wire copy by host and dateline; press releases by distributor host and marker; republishes by matching title within a story; subdomains of one registrable domain; the media groups in the ownership table.

It does not detect: a rewrite of a wire story under a new headline (that becomes a second origin); ownership outside the table; two outlets citing the same unnamed source. The labels are deterministic and heuristic, each carries a confidence, and the rules on this page are the whole method.

Claim relations

Experimental. Pass claim= and each evidence item gets a claim_relation — supports, contradicts, qualifies or unresolved — with a relation_confidence and a relation_basis of passage (judged on the full text) or snippet. The rule is lexical: the best-overlapping sentence decides; a flipped negation or a disagreeing number makes it contradicts; a hedge (however, although, partly) makes it qualifies; too little overlap is unresolved. contradictions on the response lists the ids that contradict. No model is involved, and the response says experimental: ["claim_relations"].

Next

• Grounding — where these labels appear.
• Coverage receipts — the set-level counts.
• corroborate — the host-level view of one story.