Truth at query time

A candidate qualifies by clearing either the trigram bar (morphological variants: bornIn/wasBornIn) or the embedding-cosine bar (true synonyms with no shared characters: killedBy/murderedBy). The target must be more popular than the source — the rare freshly-minted variant aligns to the established term (e.g. rdfType at 6.7K uses → rdf:type at 4.1M), which also kills cycles. Production floors: semantic ≥ 0.82, combined ≥ 0.88.

Similarity can say two predicates are related; it cannot say the relation type — direction, containment, or that near-identical strings are not equivalent. The adjudicator shows the LLM each predicate's labels plus up to 6 real (subject, object) usage pairs from the substrate, and records a verdict from a closed vocabulary: exact_equivalent · inverse_equivalent · sub_property_of · close_match · not_equivalent. Verdicts below the 0.80 floor stay candidates; explicit negatives are recorded too, so look-alikes are never re-proposed. The full audit trail (similarities, generator, the model's reasoning) lives in the alignment row's provenance.

donto_predicate_alignment (24,228 rows; bitemporal, append-only) carries three independent safety flags: safe_for_query_expansion (default true), safe_for_export (false), safe_for_logical_inference (false) — an alignment good enough to widen recall is not automatically good enough to export or reason over. Only accepted rows reach the closure.

donto_predicate_closure (1.01M rows = one self row per predicate + ~9.8K real expansion edges) is rebuilt atomically — staged in a temp table, then swapped in one transaction, so readers never see a half-built closure. Each row says: a query for predicate A should also match statements stored under B, via this relation, at this confidence, swapping subject/object if it's an inverse.

donto_match_aligned() is the standard matcher plus closure expansion. A real fold, executed on the live database: ex:mrs-e-e-brackenridge has her birthplace stored under the freely-minted wasBornIn; querying the established bornIn returns both — nobody maintains a synonym table:

SELECT subject, predicate, object_iri, matched_via, alignment_confidence
FROM donto_match_aligned(p_subject := 'ex:mrs-e-e-brackenridge', p_predicate := 'bornIn');

 ex:mrs-e-e-brackenridge | bornIn    | ex:adelaide | direct           | 1.00
 ex:mrs-e-e-brackenridge | wasBornIn | ex:adelaide | exact_equivalent | 0.95

And the inverse-swap branch — asking a question in the opposite orientation to how the fact was stored:

-- stored: ex:robert-dawson killedBy ex:unruly-horse
SELECT * FROM donto_match_aligned(p_subject := 'ex:unruly-horse', p_predicate := 'killed');

 ex:unruly-horse | killedBy | ex:robert-dawson | inverse_equivalent | 0.95
 -- subject/object swapped back so the row reads in the caller's orientation

Other live folds: birthplaceOf ↔ bornIn (inverse), diedOf ↔ causeOfDeath (inverse), killed → murdered (sub-property: querying the general term also returns the specific), affiliated-with ↔ affiliatedWith. Alignment widens matching, never visibility — consumers still post-filter by their own scope.

Discord ingestion minted both ex:traves-theberge and ex:traves_theberge (hyphen vs underscore). Fingerprint embeddings put the pair at cosine 0.84; LLM adjudication judged same_referent @ 0.95; the proposal was accepted and became identity edge #124. Resolving the underscore IRI under each hypothesis:

SELECT h.name, h.threshold_same, donto_identity_resolve_iri(h.hypothesis_id, 'ex:traves_theberge');

 strict      | 0.98 | ex:traves_theberge   -- 0.95 < 0.98: NOT merged under strict
 likely      | 0.85 | ex:traves-theberge   -- clustered: resolves to the representative
 exploratory | 0.60 | ex:traves-theberge

An epistemic sweep marks genealogy predicates functional (one true value: ex:birthYear, ex:gender, ex:birthPlace…), finds subjects with multiple distinct asserted values, and creates rebuts pairs in a dedicated context — bounded scans, idempotent inserts. Some detected conflicts (gender female vs Female vs F) are really value-normalization gaps — exactly what the value-mapping and literal-canonicalization tables are built for as they come online.

SELECT predicate, total_score, windows
FROM donto_v_top_contested_predicates ORDER BY total_score DESC LIMIT 5;

 rdf:type                  | 49715.9 | 53713
 rdfs:label                | 14155.4 | 16182
 rdfType                   |  1774.2 |  1999
 locatedIn                 |  1246.2 |  1367
 interrogationInterrogator |   842.6 |   846

Contestedness is the steering wheel: it tells the discovery lenses where reality is disputed and worth another look.