Checkpoint: What Makes a Fact Safe to Carry Forward?

Utopia has now shown four different parts of the same problem: a source must remain inspectable, an ontology must give observations usable meaning, an identity decision must join the right history, and a correction must not erase what came before. This checkpoint connects those parts and turns them into a practical judgment about a graph observation. The central question is not whether a row exists, but whether the row is evidenced, named, attached to the right entity, and safe to reuse—especially once time enters the question.

Four independent questions behind one graph row

Imagine that a document says a company supports a public program. The sentence is easy for a person to understand, but Utopia cannot treat “understandable to a reader” as one indivisible technical state. It has to preserve several links that can fail independently.

The first link is evidence. Which source item contained the sentence? Which version of that item? Which passage supports the observation? Utopia’s pipeline keeps the document and its versions, associates extracted facts with evidence, and records when an observation did not land. Without this link, a graph edge may be plausible but cannot be inspected or challenged. A later answer would have no reliable route back to the words that prompted it.

The second link is vocabulary. Did the knowledge base have an accepted relation for “supports,” and did that relation mean what the sentence required? If not, Utopia can retain the observation with the source’s wording and an empty predicate. That row is useful evidence of a relationship the current ontology does not yet name, but it is not equivalent to a named relation in the ontology. The distinction keeps a missing vocabulary term from becoming a fabricated assertion.

The third link is identity. Which company and which program do the names refer to? Exact names, aliases, context, types, and similarity can produce a candidate, but a candidate is not automatically a settled identity. Two companies can share a name, and one company can appear under several forms. A conservative split may leave one mention separate until a person or a later decision joins it to an existing entity. The graph is less consolidated in the short term, but it avoids combining two histories merely because their labels look alike.

The fourth link is reversibility. If the ontology is refined, the entity is retyped or merged, or a source later corrects its statement, can the system show what changed? Utopia’s corrections and merges use recorded transitions rather than erasing the old state. That does not make every transition harmless, but it means that a reviewer can distinguish an earlier belief from the later correction and can identify the operation that changed ownership of facts.

These links are independent. A fact may have excellent evidence but no accepted predicate. It may have a perfectly named relation but point to the wrong entity. It may be attached to the right entity and still represent an interpretation that is awaiting review. It may be fully usable now but lack the temporal scope needed to answer a historical question. Treating these as one “confidence” score would hide the kind of repair required.