onindex.ai/about/method/organization-v1

Method: organization fact sheet, version 1

Published at /about/method/organization-v1. 2026-10-22. Page type organization, URL /org/<slug>.

What the page claims

Verifiable facts about one organization: identity (legal name, website, kind, founding, headquarters, identifiers, industry) in the stable layer; size band, leadership, offerings, locations, recent events and organizational contact in the volatile layer. Never personal data: no personal emails, phones, home addresses, or facts about individuals beyond their public role at the organization.

Inputs

• A canonical organization from the demand backlog (seeded from Chef's lead lists and the demand study), with known aliases and any identifier already held.

Steps

1. Resolve the organization. Confirm the official website and legal name. If two organizations share a name, the page is for the one matching the identifiers; the other gets its own slug.
2. Source discovery, in order of authority. Official site (about, press, investors), the relevant government or statutory registry, Wikidata and Wikipedia where present, one or two reference directories, recent news from named outlets. Target 3–6 sources. Record every attempt and every failure.
3. Identity extraction into stable.identity. Registry beats official site on legal name and incorporation date; official site beats everything on offerings and leadership.
4. Facts extraction into volatile.facts. Size as a band, never a point number unless the source states it. Leadership with as_of. Recent events with date and source.
5. Disagreement. Official versus registry conflicts on identity facts: registry wins and the claim notes the discrepancy. Anything else unresolved becomes disputed.
6. Watch dates. Fiscal year end, announced events, pending registrations.
7. Draft the page to the schema with coverage stated in words.

Verification

Every claim is written unverified by the producer (site/produce.py): one tool-less, schema-constrained model read the organization's own page once. A claim becomes verified only through the second reader (site/second_reader.py): a second tool-less, schema-constrained model, a different one, re-reads the stored copy of the page or pages the claim cites, the same copy the first reader read, and extracts the same structured record; the claim is marked verified when that record states the claimed value under the comparison rules (website host, headquarters city, founding year, operating status, the leader at the same title, every practitioner and service named, the street address and every phone number printed, the opening hours, direct billing and emergency appointments) and no second reading of a cited page states another. A claim whose second reading is silent or different, whose cited page has no stored copy, or that is a labelled summary ("what it does", which no code rule can compare) stays unverified and the page's disclosure says so. Each verified claim's verified_by names the reader, the model and the day; the site refuses to publish a verified claim that names no verifier. The pass does not re-fetch the page and consults no other source. (The cross-page check on makers named on product pages is not built.)

Publish gates

• The official site or a registry must be accessible; a page with neither does not publish.
• Coverage ratio at least 0.5.
• No personal data (automated check on the contact block and leadership entries).
• Schema and invariants pass.

Refresh

Monthly by class; sooner on watch dates, on a change to the official site's about or press pages (ETag or content hash), or on feedback. The v1 correction channel (brief §15) will add official-statement as a source kind; until then an organization's own statement reaches us only through its website.