The claim model

Created in migration 0001_core.sql. The row is deliberately narrow — the substrate never widens the 41.7M-row table; per-claim metadata lives in keyed side tables instead:

create table donto_statement (
    statement_id  uuid primary key default gen_random_uuid(),
    subject       text not null,
    predicate     text not null,
    object_iri    text,            -- reference …
    object_lit    jsonb,           -- … XOR value
    context       text not null references donto_context(iri),
    tx_time       tstzrange not null default tstzrange(now(), null, '[)'),
    valid_time    daterange not null default daterange(null, null, '[)'),
    flags         smallint not null default 0,
    content_hash  bytea generated always as ( /* sha256, see below */ ) stored,
    constraint donto_statement_object_one_of
        check ((object_iri is not null) <> (object_lit is not null))
);

-- a graph edge with a freely-minted predicate (BEAM-10M extraction)
subject  ex:weaviate-client   predicate  performsAction   object_iri  ex:vector-encryption
context  ctx:claims/beam/cbcc52f9-…   flags 8 (asserted, E2)

-- a typed literal
subject  ex:cipher   predicate  rdfs:label   object_lit {"v":"cipher","dt":"xsd:string"}

-- a memory chunk carrying its session date as valid_time
subject  ctx:memory/episodic/000351d1-…   predicate  mem:episodic/chunk
object_lit {"v":"[Session date: 2023/05/22] User: I'm looking to improve…","dt":"xsd:string"}
valid_time [2023-05-22,)

-- a retracted row — closed tx_time, nothing deleted (I3)
subject ex:carol  predicate ex:loves  object_iri ex:wrong
tx_time ["2026-04-17 11:09:05.66+00","2026-04-17 11:09:05.66+00")

The single write surface is donto_assert(…) (SQL) / DontoClient::assert(&StatementInput) (Rust). Inside the function: require a context → auto-create it if new → permissive mode auto-registers unknown predicates / curated mode enforces the registry + minting approval → insert with content-hash conflict handling → audit. Idempotent end to end.

select donto_assert(
    p_subject    => 'ex:weaviate-client',
    p_predicate  => 'performsAction',      -- invented freely; auto-registered
    p_object_iri => 'ex:vector-encryption',
    p_object_lit => null,
    p_context    => 'ctx:claims/beam/cbcc52f9-…',
    p_polarity   => 'asserted',
    p_maturity   => 2,                     -- E2: the machine-extractor cap
    p_valid_lo   => null, p_valid_hi => null,
    p_actor      => 'agent:donto-agent'
);  -- returns uuid; no-op if the same open claim already exists

A context is where a claim lives: a named scope carrying provenance grouping, a trust mode, and a position in a hierarchy. Contexts are cheap — donto_ensure_context auto-creates them on first assert — which is why there are 68,325.

The big residents: ctx:genealogy/research-db holds 21.8M statements; ctx:claims/* (21,170 contexts) holds the extraction firehose, one context per source chunk; ctx:memory/* (20,767) is the agent-memory overlay.

Any predicate IRI may appear on a statement; the registry exists to record, describe and align predicates, not to gate writes. 99.8% of its 1,009,054 rows are status='implicit' — minted by models on first use (performsAction, dependsOn, returns-on-true, …, some seconds old at any given moment). The registry also carries alias structure (canonical_of), shape hints (domain, range_iri), algebraic properties (is_symmetric/transitive/functional), and the minting workflow status for curated contexts. Folding the 1M predicates into usable families is the alignment engine's job — at query time, never at write time.