Skip to main content
An assertion records a claim extracted from a document and the evidence span that supports it. In “Mara joined Northstar Labs in March 2026,” an assertion can preserve the subject mention Mara, the raw predicate, and the object mention Northstar Labs before and after they are mapped to canonical records. An assertion answers what did this source say? A truth slot answers which supported values are current? Keeping both lets you inspect a disagreement without losing either source’s claim.

​
From extraction to committed evidence

The active graph-ingestion path stages model output as knowledge_candidates artifacts. Validation prepares a typed commit bundle, and commit_artifacts writes assertions and the corresponding graph or event records. Entity resolution and promotion participate in that boundary. A successful extraction work item is not, on its own, proof that canonical records were committed. Use document pipeline status to follow the artifact and work-item stages. The former separate resolve_assertions worker is outside the active path.

​
Read the parts of a claim

Current inline promotion copies the extraction score into resolution_confidence; it is not an independently calibrated identity score. A literal claim does not need an object entity. A resolved assertion also does not imply that its derived relationship is approved or that its value won a conflict.

​
Follow evidence to immutable text

Every assertion has an EvidenceSpan: Read the cited revision, slice the recorded byte range, and verify its hash when interpreting evidence. Do not apply old offsets to a document’s newest text: a later render can change the text and offsets while the original revision remains the evidence authority. Byte offsets also differ from character positions for non-ASCII text. Media-backed revisions can additionally map those spans to attachments and time ranges through their render manifest. See documents.

​
Distinguish the three timestamps

A recently ingested source can describe an old event. These fields distinguish the intended temporal meanings. Current inline promotion initializes observed_at and asserted_at to the promotion time, so those two fields alone do not reconstruct a delayed source’s original arrival time. Inspect source metadata and the evidence timestamps when that distinction matters.

​
Resolution and retention state

The record type supports the following resolution states: The active inline promotion path writes resolved assertions. These states are record semantics, not a promise that each assertion passes through six separate worker stages. Retry failures are inspected through the active work queue. assertion_key makes repeated promotion of the same node and claim idempotent. Its current construction includes the document-node ID, raw predicate, and a hash of the subject and object. A new revision with different nodes can therefore produce new assertion IDs; deduplication is not a promise that every paraphrase or re-render collapses into one record. Assertions also carry lifecycle_tier. The active tier is used by normal processing; archival can retain older resolved evidence for provenance while excluding it from normal active queries. The record’s tier is distinct from its resolution state and from a work item’s status.

​
Next steps