This page describes how we source and verify the editorial content on Discover. The goal is auditability: a reader who wants to check a specific claim should be able to follow the trail from the page back to a verifiable record.

Source types

We accept, in rough order of authority weight:

• Scholarly art-history books, including catalogues raisonnés.
• Primary museum collection records (the museum's own page for the work).
• Encyclopedic references, treated as verifiable references rather than authorities.
• Curated knowledge-graph and archive entries, used for structured facts (dates, identifiers, holdings) and as a check on prose.

Each source carries a type label in the footer so readers can judge authority at a glance. Sources are listed once per page even when they back several claims, with a "Used for: …" line summarising what they support.

City guide datasets

City guides and downloadable datasets use row-level provenance. Each exported row records sources for museum identity, location, artist identity, holding or artwork association, verification status, and last verification date.

Before a flagship city is launched, Andrew Parry and Michael Hamilton review the top 50 rows by published importance score. The remaining rows pass automated QA and a deterministic sample review. The public score discloses its categories and weights (source depth 40%, museum or collection prominence 35%, audience demand 25%) without exposing internal matching thresholds or queue logic.

Museum and artwork-location language is conservative: we use terms such as "recorded at" or "associated with" unless a source confirms current display. Museum closures, renamed institutions, ambiguous holdings, and moved collections are flagged or excluded from the public dataset until reviewed.

Citation placement

Inline ^[N] superscripts are placed deterministically, not by AI. They appear next to the specific claims a source backs, with adjacent duplicates suppressed.

Honest dateModified

Schema.org dateModified is gated on a content-change check: we only update it when the rendered page actually differs from its previous version. We do not bump dateModified on no-op republishes, which would be a freshness signal we cannot honestly back.

What we do not yet verify

We are explicit about gaps:

• Image attribution at the per-image level. Every image is Creative Commons or public domain, but our license tagging is at the source-record level. Per-image is in progress.
• Source-URL liveness. URLs in our source pool are not yet automatically re-checked for 404s; we monitor at deploy time and rely on reader reports between deploys.
• Editorial depth on long-tail entries. The richest coverage is on the most-searched pages. Lesser-known artists have shorter biographies and fewer citations.
• External review for cultural perspective. Pre-publication content guardrails reject historical content that requires human review (see AI Disclosure). Those guardrails are not a substitute for review by domain experts.