How the record is built

Provenance in the Watts Index is not a citation stapled to a page — it is attached to each individual fact. A single artist record is assembled from many fields, and every field keeps its own record of where its value came from: the source, the source’s trust tier, a link back to the origin where one exists, and a verification status.

No canonical field is ever written directly. Every value passes through a single adjudicating write path. When a new value arrives for a field, that path:

1. 1Records the incoming value and its source as a candidate — always, even if it will not be adopted.
2. 2Compares it against the value already on record and the full set of prior candidates for that field.
3. 3Applies the source-trust hierarchy: a higher-trust source overwrites a lower one; a lower-trust source can never overwrite a higher one.
4. 4Publishes the value to the canonical record only when the trust and corroboration rules are satisfied.
5. 5Marks the field disputed when two sources of equal standing assert different values.

Because the candidate log is kept whether or not a value is adopted, the record is auditable in both directions: you can see not only what we publish, but what we considered and rejected, and why.

Not all sources deserve equal weight. Every source that feeds the index is placed in one of three trust tiers. The rule is strict and directional: a lower tier can add evidence, but it can never overwrite a fact established by a higher tier. AI may summarise, but it is never the canonical source of a fact.

The registry currently defines 83 named sources across these three tiers. It is versioned and grows as new authorities are onboarded.

One subtlety matters for trust: two sources are only independent if they do not descend from the same origin. Where a “source” is really a mirror or re-publication of another, the registry collapses them into a single upstream family, so a genuine institution and an internal re-derivation of it are never mistaken for two witnesses. Independence is judged at the level of the family, not the label.

A fact is not treated as verified because one source asserts it — even a Tier-1 one. It becomes verified only when two independent upstream families assert the same value. A single strong source is recorded and shown, but it is held as sourced, not verified, until a second, genuinely independent family concurs.

The corroboration rule is a versioned predicate — currently v1.3— so the exact standard applied to any record is itself part of the record’s history. When we tighten what counts as independent corroboration, that change is versioned rather than applied silently.

Crucially, conflicts are never resolved by averaging. If one source says a work was made in 1921 and another says 1923, the answer is not 1922. Both claims are preserved, the field is marked disputed, and it stays out of the canonical layer until the disagreement is resolved by evidence.

Requiring independent corroboration turns the catalogue into its own auditor. Because agreement only counts across independent families, an error that has propagated from a single origin into many downstream copies cannot masquerade as corroboration — the copies collapse back into one family and the fact remains merely sourced, not verified. Shared mistakes stop being self-reinforcing.

The opposite case is just as useful. When genuinely independent sources disagree, the system does not pick a winner or quietly average them — it surfaces the disagreement as a conflict. Each disputed field is a signal pointing to exactly where the world’s record of an artist or object is unsettled, which is precisely where careful research adds the most value. Errors that other datasets bury inside a single confident-looking value are, here, visible on the surface.