Sources disagree. Most databases pick a winner. donto keeps them all.

For fifty years knowledge graphs were scarce — every triple expensive to author, schema first, facts slow. Models invert that economy: GPTKB pulled 105M typed triplesfrom one mid-tier model at ~$0.00009/claim; AutoSchemaKG built a 900M-node graph with no schema; cost falls ~10× a year. The bottleneck is no longer extraction — it's trust.

01 · Emit freely

No fixed schema, no pre-typed predicates. Models invent the predicate they need (~1M minted) and assert in every direction — the signature of abundance, not a bug.

02 · Defer to query time

Typing, alignment, identity-resolution and joining aren't write gates. They're query-time judgments — composed on demand, reversible, never destroying the raw claim.

03 · Prune by reality

Contradiction is stored paraconsistently; a claim's evidence, corroboration and lifecycle decide its standing. Reality is the verifier, not a curator.

Anyone can extract facts. The durable advantage is what happens after: how a claim earns or loses standing over time, evidence intact, contradictions preserved.

step 1 · Ingest

Register any source as an immutable, content-addressed document the substrate can always retrieve.

step 2 · Emit free

Unbounded claims, predicates invented as needed. Only write-time invariant: an evidence anchor or an explicit hypothesis flag.

step 3 · Hold incompatible

Contradictory claims stored side-by-side as legal bitemporal state. A conflict is data, not a failed write.

step 4 · Hypothesize

Typed relationship hypotheses proposed where lenses intersect — connections no schema author would pre-type.

step 5 · Attach evidence

Supports / rebuts / undercuts / qualifies — argument edges carry evidence and counter-evidence with provenance.

step 6 · Rank

Claims scored by value — information gain, novelty, downstream task-lift — not accuracy alone.

step 7 · Re-rank in time

New evidence re-scores old hypotheses. Bitemporal: standing compounds; nothing frozen at write time.

step 8 · Explain

Answers explained only from attached evidence — faithful by construction, never narrative-first.

Each invariant is enforced at the schema and API layer and guarded by 80 dedicated invariant test suites in CI — the substrate refuses to be a normal database.

Every piece of the thesis has a running counterpart — watchable in real time at scanner.donto.org.

Alignment engine

Embedding fabric (919K predicate + 338K entity vectors, bge-small/HNSW) + a continuous daemon proposing/adjudicating a 1M-row predicate closure. killedBy meets murderedBy at cosine 0.95 — no synonym table.

Always-on citer

Extraction (what) and anchoring (where) are separate stages. Every fact gets its exact evidence span — or an honest interpretation flag, never a bogus span. Doubles as a hallucination filter.

Gleaning loop

Models stop early by choice; the harness re-prompts to saturation (one article 511 → 3,227 facts). A count floor makes models pad with garbage — saturation decides done.

One engine, many lanes

donto-extract: one engine, 8 swappable model lanes as a declarative registry. Capped lanes rotate out automatically; the citer holds every lane to the same evidence standard.

donto-memory was run through LongMemEval (ICLR 2025) under audited no-leakage conditions. Where a history fits a frontier model's context, raw accuracy ties a full-context reader — the substrate earns its keep on retrieval quality, token cost, knowledge-update and abstention.

Full methodology in the LongMemEval study.

Prosopography — reconstructing people and their relationships from fragmentary, scattered and often contradictory mentions across deeds, registers, newspapers, censuses and inscriptions — is the discipline donto was built for. No single source gives you a person; you triangulate one from dozens of partial, disagreeing attestations. That is exactly the shape of contested reality, and the substrate's hardest invariants are the tools the historian actually needs.

Every mention is a claim

A witness in a deed, a name in a burial register, one line in a newspaper — each becomes an evidence-anchored claim with its exact source span, never flattened into a single “record.”

Name variants aren't noise

Mahamoodally / Mamode Ally, Lablanche / Lablache — competing surface forms held together and reconciled by similarity at query time, not a hand-kept alias table.

Is it the same person?

Identity is a hypothesis (I8). Two attestations become one person only under a lens you pick per query — strict, cluster or transitive — never a destructive merge you can't take back.

Contradictions triangulate

Born 1889 vs 1890, this parish vs that — held side by side with their sources. One marriage certificate that names both parents re-ranks the rest; you see which reconstruction the evidence actually supports.

Live in genealogy.donto.org — contested family history is prosopography at its hardest: century-old records, missing links, and identity that is itself the open question.